Linux Audio

Check our new training course

Loading...
v4.6
   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 "dm-bufio.h"
  10
  11#include <linux/device-mapper.h>
  12#include <linux/dm-io.h>
  13#include <linux/slab.h>
  14#include <linux/jiffies.h>
  15#include <linux/vmalloc.h>
  16#include <linux/shrinker.h>
  17#include <linux/module.h>
  18#include <linux/rbtree.h>
  19#include <linux/stacktrace.h>
  20
  21#define DM_MSG_PREFIX "bufio"
  22
  23/*
  24 * Memory management policy:
  25 *	Limit the number of buffers to DM_BUFIO_MEMORY_PERCENT of main memory
  26 *	or DM_BUFIO_VMALLOC_PERCENT of vmalloc memory (whichever is lower).
  27 *	Always allocate at least DM_BUFIO_MIN_BUFFERS buffers.
  28 *	Start background writeback when there are DM_BUFIO_WRITEBACK_PERCENT
  29 *	dirty buffers.
  30 */
  31#define DM_BUFIO_MIN_BUFFERS		8
  32
  33#define DM_BUFIO_MEMORY_PERCENT		2
  34#define DM_BUFIO_VMALLOC_PERCENT	25
  35#define DM_BUFIO_WRITEBACK_PERCENT	75
  36
  37/*
  38 * Check buffer ages in this interval (seconds)
  39 */
  40#define DM_BUFIO_WORK_TIMER_SECS	30
  41
  42/*
  43 * Free buffers when they are older than this (seconds)
  44 */
  45#define DM_BUFIO_DEFAULT_AGE_SECS	300
  46
  47/*
  48 * The nr of bytes of cached data to keep around.
  49 */
  50#define DM_BUFIO_DEFAULT_RETAIN_BYTES   (256 * 1024)
  51
  52/*
  53 * The number of bvec entries that are embedded directly in the buffer.
  54 * If the chunk size is larger, dm-io is used to do the io.
  55 */
  56#define DM_BUFIO_INLINE_VECS		16
  57
  58/*
 
 
 
 
 
 
 
 
  59 * Don't try to use kmem_cache_alloc for blocks larger than this.
  60 * For explanation, see alloc_buffer_data below.
  61 */
  62#define DM_BUFIO_BLOCK_SIZE_SLAB_LIMIT	(PAGE_SIZE >> 1)
  63#define DM_BUFIO_BLOCK_SIZE_GFP_LIMIT	(PAGE_SIZE << (MAX_ORDER - 1))
  64
  65/*
  66 * dm_buffer->list_mode
  67 */
  68#define LIST_CLEAN	0
  69#define LIST_DIRTY	1
  70#define LIST_SIZE	2
  71
  72/*
  73 * Linking of buffers:
  74 *	All buffers are linked to cache_hash with their hash_list field.
  75 *
  76 *	Clean buffers that are not being written (B_WRITING not set)
  77 *	are linked to lru[LIST_CLEAN] with their lru_list field.
  78 *
  79 *	Dirty and clean buffers that are being written are linked to
  80 *	lru[LIST_DIRTY] with their lru_list field. When the write
  81 *	finishes, the buffer cannot be relinked immediately (because we
  82 *	are in an interrupt context and relinking requires process
  83 *	context), so some clean-not-writing buffers can be held on
  84 *	dirty_lru too.  They are later added to lru in the process
  85 *	context.
  86 */
  87struct dm_bufio_client {
  88	struct mutex lock;
  89
  90	struct list_head lru[LIST_SIZE];
  91	unsigned long n_buffers[LIST_SIZE];
  92
  93	struct block_device *bdev;
  94	unsigned block_size;
  95	unsigned char sectors_per_block_bits;
  96	unsigned char pages_per_block_bits;
  97	unsigned char blocks_per_page_bits;
  98	unsigned aux_size;
  99	void (*alloc_callback)(struct dm_buffer *);
 100	void (*write_callback)(struct dm_buffer *);
 101
 102	struct dm_io_client *dm_io;
 103
 104	struct list_head reserved_buffers;
 105	unsigned need_reserved_buffers;
 106
 107	unsigned minimum_buffers;
 108
 109	struct rb_root buffer_tree;
 110	wait_queue_head_t free_buffer_wait;
 111
 112	int async_write_error;
 113
 114	struct list_head client_list;
 115	struct shrinker shrinker;
 116};
 117
 118/*
 119 * Buffer state bits.
 120 */
 121#define B_READING	0
 122#define B_WRITING	1
 123#define B_DIRTY		2
 124
 125/*
 126 * Describes how the block was allocated:
 127 * kmem_cache_alloc(), __get_free_pages() or vmalloc().
 128 * See the comment at alloc_buffer_data.
 129 */
 130enum data_mode {
 131	DATA_MODE_SLAB = 0,
 132	DATA_MODE_GET_FREE_PAGES = 1,
 133	DATA_MODE_VMALLOC = 2,
 134	DATA_MODE_LIMIT = 3
 135};
 136
 137struct dm_buffer {
 138	struct rb_node node;
 139	struct list_head lru_list;
 140	sector_t block;
 141	void *data;
 142	enum data_mode data_mode;
 143	unsigned char list_mode;		/* LIST_* */
 144	unsigned hold_count;
 145	int read_error;
 146	int write_error;
 147	unsigned long state;
 148	unsigned long last_accessed;
 149	struct dm_bufio_client *c;
 150	struct list_head write_list;
 151	struct bio bio;
 152	struct bio_vec bio_vec[DM_BUFIO_INLINE_VECS];
 153#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
 154#define MAX_STACK 10
 155	struct stack_trace stack_trace;
 156	unsigned long stack_entries[MAX_STACK];
 157#endif
 158};
 159
 160/*----------------------------------------------------------------*/
 161
 162static struct kmem_cache *dm_bufio_caches[PAGE_SHIFT - SECTOR_SHIFT];
 163static char *dm_bufio_cache_names[PAGE_SHIFT - SECTOR_SHIFT];
 164
 165static inline int dm_bufio_cache_index(struct dm_bufio_client *c)
 166{
 167	unsigned ret = c->blocks_per_page_bits - 1;
 168
 169	BUG_ON(ret >= ARRAY_SIZE(dm_bufio_caches));
 170
 171	return ret;
 172}
 173
 174#define DM_BUFIO_CACHE(c)	(dm_bufio_caches[dm_bufio_cache_index(c)])
 175#define DM_BUFIO_CACHE_NAME(c)	(dm_bufio_cache_names[dm_bufio_cache_index(c)])
 176
 177#define dm_bufio_in_request()	(!!current->bio_list)
 178
 179static void dm_bufio_lock(struct dm_bufio_client *c)
 180{
 181	mutex_lock_nested(&c->lock, dm_bufio_in_request());
 182}
 183
 184static int dm_bufio_trylock(struct dm_bufio_client *c)
 185{
 186	return mutex_trylock(&c->lock);
 187}
 188
 189static void dm_bufio_unlock(struct dm_bufio_client *c)
 190{
 191	mutex_unlock(&c->lock);
 192}
 193
 194/*
 195 * FIXME Move to sched.h?
 196 */
 197#ifdef CONFIG_PREEMPT_VOLUNTARY
 198#  define dm_bufio_cond_resched()		\
 199do {						\
 200	if (unlikely(need_resched()))		\
 201		_cond_resched();		\
 202} while (0)
 203#else
 204#  define dm_bufio_cond_resched()                do { } while (0)
 205#endif
 206
 207/*----------------------------------------------------------------*/
 208
 209/*
 210 * Default cache size: available memory divided by the ratio.
 211 */
 212static unsigned long dm_bufio_default_cache_size;
 213
 214/*
 215 * Total cache size set by the user.
 216 */
 217static unsigned long dm_bufio_cache_size;
 218
 219/*
 220 * A copy of dm_bufio_cache_size because dm_bufio_cache_size can change
 221 * at any time.  If it disagrees, the user has changed cache size.
 222 */
 223static unsigned long dm_bufio_cache_size_latch;
 224
 225static DEFINE_SPINLOCK(param_spinlock);
 226
 227/*
 228 * Buffers are freed after this timeout
 229 */
 230static unsigned dm_bufio_max_age = DM_BUFIO_DEFAULT_AGE_SECS;
 231static unsigned dm_bufio_retain_bytes = DM_BUFIO_DEFAULT_RETAIN_BYTES;
 232
 233static unsigned long dm_bufio_peak_allocated;
 234static unsigned long dm_bufio_allocated_kmem_cache;
 235static unsigned long dm_bufio_allocated_get_free_pages;
 236static unsigned long dm_bufio_allocated_vmalloc;
 237static unsigned long dm_bufio_current_allocated;
 238
 239/*----------------------------------------------------------------*/
 240
 241/*
 242 * Per-client cache: dm_bufio_cache_size / dm_bufio_client_count
 243 */
 244static unsigned long dm_bufio_cache_size_per_client;
 245
 246/*
 247 * The current number of clients.
 248 */
 249static int dm_bufio_client_count;
 250
 251/*
 252 * The list of all clients.
 253 */
 254static LIST_HEAD(dm_bufio_all_clients);
 255
 256/*
 257 * This mutex protects dm_bufio_cache_size_latch,
 258 * dm_bufio_cache_size_per_client and dm_bufio_client_count
 259 */
 260static DEFINE_MUTEX(dm_bufio_clients_lock);
 261
 262#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
 263static void buffer_record_stack(struct dm_buffer *b)
 264{
 265	b->stack_trace.nr_entries = 0;
 266	b->stack_trace.max_entries = MAX_STACK;
 267	b->stack_trace.entries = b->stack_entries;
 268	b->stack_trace.skip = 2;
 269	save_stack_trace(&b->stack_trace);
 270}
 271#endif
 272
 273/*----------------------------------------------------------------
 274 * A red/black tree acts as an index for all the buffers.
 275 *--------------------------------------------------------------*/
 276static struct dm_buffer *__find(struct dm_bufio_client *c, sector_t block)
 277{
 278	struct rb_node *n = c->buffer_tree.rb_node;
 279	struct dm_buffer *b;
 280
 281	while (n) {
 282		b = container_of(n, struct dm_buffer, node);
 283
 284		if (b->block == block)
 285			return b;
 286
 287		n = (b->block < block) ? n->rb_left : n->rb_right;
 288	}
 289
 290	return NULL;
 291}
 292
 293static void __insert(struct dm_bufio_client *c, struct dm_buffer *b)
 294{
 295	struct rb_node **new = &c->buffer_tree.rb_node, *parent = NULL;
 296	struct dm_buffer *found;
 297
 298	while (*new) {
 299		found = container_of(*new, struct dm_buffer, node);
 300
 301		if (found->block == b->block) {
 302			BUG_ON(found != b);
 303			return;
 304		}
 305
 306		parent = *new;
 307		new = (found->block < b->block) ?
 308			&((*new)->rb_left) : &((*new)->rb_right);
 309	}
 310
 311	rb_link_node(&b->node, parent, new);
 312	rb_insert_color(&b->node, &c->buffer_tree);
 313}
 314
 315static void __remove(struct dm_bufio_client *c, struct dm_buffer *b)
 316{
 317	rb_erase(&b->node, &c->buffer_tree);
 318}
 319
 320/*----------------------------------------------------------------*/
 321
 322static void adjust_total_allocated(enum data_mode data_mode, long diff)
 323{
 324	static unsigned long * const class_ptr[DATA_MODE_LIMIT] = {
 325		&dm_bufio_allocated_kmem_cache,
 326		&dm_bufio_allocated_get_free_pages,
 327		&dm_bufio_allocated_vmalloc,
 328	};
 329
 330	spin_lock(&param_spinlock);
 331
 332	*class_ptr[data_mode] += diff;
 333
 334	dm_bufio_current_allocated += diff;
 335
 336	if (dm_bufio_current_allocated > dm_bufio_peak_allocated)
 337		dm_bufio_peak_allocated = dm_bufio_current_allocated;
 338
 339	spin_unlock(&param_spinlock);
 340}
 341
 342/*
 343 * Change the number of clients and recalculate per-client limit.
 344 */
 345static void __cache_size_refresh(void)
 346{
 347	BUG_ON(!mutex_is_locked(&dm_bufio_clients_lock));
 348	BUG_ON(dm_bufio_client_count < 0);
 349
 350	dm_bufio_cache_size_latch = ACCESS_ONCE(dm_bufio_cache_size);
 351
 352	/*
 353	 * Use default if set to 0 and report the actual cache size used.
 354	 */
 355	if (!dm_bufio_cache_size_latch) {
 356		(void)cmpxchg(&dm_bufio_cache_size, 0,
 357			      dm_bufio_default_cache_size);
 358		dm_bufio_cache_size_latch = dm_bufio_default_cache_size;
 359	}
 360
 361	dm_bufio_cache_size_per_client = dm_bufio_cache_size_latch /
 362					 (dm_bufio_client_count ? : 1);
 363}
 364
 365/*
 366 * Allocating buffer data.
 367 *
 368 * Small buffers are allocated with kmem_cache, to use space optimally.
 369 *
 370 * For large buffers, we choose between get_free_pages and vmalloc.
 371 * Each has advantages and disadvantages.
 372 *
 373 * __get_free_pages can randomly fail if the memory is fragmented.
 374 * __vmalloc won't randomly fail, but vmalloc space is limited (it may be
 375 * as low as 128M) so using it for caching is not appropriate.
 376 *
 377 * If the allocation may fail we use __get_free_pages. Memory fragmentation
 378 * won't have a fatal effect here, but it just causes flushes of some other
 379 * buffers and more I/O will be performed. Don't use __get_free_pages if it
 380 * always fails (i.e. order >= MAX_ORDER).
 381 *
 382 * If the allocation shouldn't fail we use __vmalloc. This is only for the
 383 * initial reserve allocation, so there's no risk of wasting all vmalloc
 384 * space.
 385 */
 386static void *alloc_buffer_data(struct dm_bufio_client *c, gfp_t gfp_mask,
 387			       enum data_mode *data_mode)
 388{
 389	unsigned noio_flag;
 390	void *ptr;
 391
 392	if (c->block_size <= DM_BUFIO_BLOCK_SIZE_SLAB_LIMIT) {
 393		*data_mode = DATA_MODE_SLAB;
 394		return kmem_cache_alloc(DM_BUFIO_CACHE(c), gfp_mask);
 395	}
 396
 397	if (c->block_size <= DM_BUFIO_BLOCK_SIZE_GFP_LIMIT &&
 398	    gfp_mask & __GFP_NORETRY) {
 399		*data_mode = DATA_MODE_GET_FREE_PAGES;
 400		return (void *)__get_free_pages(gfp_mask,
 401						c->pages_per_block_bits);
 402	}
 403
 404	*data_mode = DATA_MODE_VMALLOC;
 405
 406	/*
 407	 * __vmalloc allocates the data pages and auxiliary structures with
 408	 * gfp_flags that were specified, but pagetables are always allocated
 409	 * with GFP_KERNEL, no matter what was specified as gfp_mask.
 410	 *
 411	 * Consequently, we must set per-process flag PF_MEMALLOC_NOIO so that
 412	 * all allocations done by this process (including pagetables) are done
 413	 * as if GFP_NOIO was specified.
 414	 */
 415
 416	if (gfp_mask & __GFP_NORETRY)
 417		noio_flag = memalloc_noio_save();
 418
 419	ptr = __vmalloc(c->block_size, gfp_mask | __GFP_HIGHMEM, PAGE_KERNEL);
 420
 421	if (gfp_mask & __GFP_NORETRY)
 422		memalloc_noio_restore(noio_flag);
 423
 424	return ptr;
 425}
 426
 427/*
 428 * Free buffer's data.
 429 */
 430static void free_buffer_data(struct dm_bufio_client *c,
 431			     void *data, enum data_mode data_mode)
 432{
 433	switch (data_mode) {
 434	case DATA_MODE_SLAB:
 435		kmem_cache_free(DM_BUFIO_CACHE(c), data);
 436		break;
 437
 438	case DATA_MODE_GET_FREE_PAGES:
 439		free_pages((unsigned long)data, c->pages_per_block_bits);
 440		break;
 441
 442	case DATA_MODE_VMALLOC:
 443		vfree(data);
 444		break;
 445
 446	default:
 447		DMCRIT("dm_bufio_free_buffer_data: bad data mode: %d",
 448		       data_mode);
 449		BUG();
 450	}
 451}
 452
 453/*
 454 * Allocate buffer and its data.
 455 */
 456static struct dm_buffer *alloc_buffer(struct dm_bufio_client *c, gfp_t gfp_mask)
 457{
 458	struct dm_buffer *b = kmalloc(sizeof(struct dm_buffer) + c->aux_size,
 459				      gfp_mask);
 460
 461	if (!b)
 462		return NULL;
 463
 464	b->c = c;
 465
 466	b->data = alloc_buffer_data(c, gfp_mask, &b->data_mode);
 467	if (!b->data) {
 468		kfree(b);
 469		return NULL;
 470	}
 471
 472	adjust_total_allocated(b->data_mode, (long)c->block_size);
 473
 474#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
 475	memset(&b->stack_trace, 0, sizeof(b->stack_trace));
 476#endif
 477	return b;
 478}
 479
 480/*
 481 * Free buffer and its data.
 482 */
 483static void free_buffer(struct dm_buffer *b)
 484{
 485	struct dm_bufio_client *c = b->c;
 486
 487	adjust_total_allocated(b->data_mode, -(long)c->block_size);
 488
 489	free_buffer_data(c, b->data, b->data_mode);
 490	kfree(b);
 491}
 492
 493/*
 494 * Link buffer to the hash list and clean or dirty queue.
 495 */
 496static void __link_buffer(struct dm_buffer *b, sector_t block, int dirty)
 497{
 498	struct dm_bufio_client *c = b->c;
 499
 500	c->n_buffers[dirty]++;
 501	b->block = block;
 502	b->list_mode = dirty;
 503	list_add(&b->lru_list, &c->lru[dirty]);
 504	__insert(b->c, b);
 505	b->last_accessed = jiffies;
 506}
 507
 508/*
 509 * Unlink buffer from the hash list and dirty or clean queue.
 510 */
 511static void __unlink_buffer(struct dm_buffer *b)
 512{
 513	struct dm_bufio_client *c = b->c;
 514
 515	BUG_ON(!c->n_buffers[b->list_mode]);
 516
 517	c->n_buffers[b->list_mode]--;
 518	__remove(b->c, b);
 519	list_del(&b->lru_list);
 520}
 521
 522/*
 523 * Place the buffer to the head of dirty or clean LRU queue.
 524 */
 525static void __relink_lru(struct dm_buffer *b, int dirty)
 526{
 527	struct dm_bufio_client *c = b->c;
 528
 529	BUG_ON(!c->n_buffers[b->list_mode]);
 530
 531	c->n_buffers[b->list_mode]--;
 532	c->n_buffers[dirty]++;
 533	b->list_mode = dirty;
 534	list_move(&b->lru_list, &c->lru[dirty]);
 535	b->last_accessed = jiffies;
 536}
 537
 538/*----------------------------------------------------------------
 539 * Submit I/O on the buffer.
 540 *
 541 * Bio interface is faster but it has some problems:
 542 *	the vector list is limited (increasing this limit increases
 543 *	memory-consumption per buffer, so it is not viable);
 544 *
 545 *	the memory must be direct-mapped, not vmalloced;
 546 *
 547 *	the I/O driver can reject requests spuriously if it thinks that
 548 *	the requests are too big for the device or if they cross a
 549 *	controller-defined memory boundary.
 550 *
 551 * If the buffer is small enough (up to DM_BUFIO_INLINE_VECS pages) and
 552 * it is not vmalloced, try using the bio interface.
 553 *
 554 * If the buffer is big, if it is vmalloced or if the underlying device
 555 * rejects the bio because it is too large, use dm-io layer to do the I/O.
 556 * The dm-io layer splits the I/O into multiple requests, avoiding the above
 557 * shortcomings.
 558 *--------------------------------------------------------------*/
 559
 560/*
 561 * dm-io completion routine. It just calls b->bio.bi_end_io, pretending
 562 * that the request was handled directly with bio interface.
 563 */
 564static void dmio_complete(unsigned long error, void *context)
 565{
 566	struct dm_buffer *b = context;
 567
 568	b->bio.bi_error = error ? -EIO : 0;
 569	b->bio.bi_end_io(&b->bio);
 570}
 571
 572static void use_dmio(struct dm_buffer *b, int rw, sector_t block,
 573		     bio_end_io_t *end_io)
 574{
 575	int r;
 576	struct dm_io_request io_req = {
 577		.bi_rw = rw,
 578		.notify.fn = dmio_complete,
 579		.notify.context = b,
 580		.client = b->c->dm_io,
 581	};
 582	struct dm_io_region region = {
 583		.bdev = b->c->bdev,
 584		.sector = block << b->c->sectors_per_block_bits,
 585		.count = b->c->block_size >> SECTOR_SHIFT,
 586	};
 587
 588	if (b->data_mode != DATA_MODE_VMALLOC) {
 589		io_req.mem.type = DM_IO_KMEM;
 590		io_req.mem.ptr.addr = b->data;
 591	} else {
 592		io_req.mem.type = DM_IO_VMA;
 593		io_req.mem.ptr.vma = b->data;
 594	}
 595
 596	b->bio.bi_end_io = end_io;
 597
 598	r = dm_io(&io_req, 1, &region, NULL);
 599	if (r) {
 600		b->bio.bi_error = r;
 601		end_io(&b->bio);
 602	}
 603}
 604
 605static void inline_endio(struct bio *bio)
 606{
 607	bio_end_io_t *end_fn = bio->bi_private;
 608	int error = bio->bi_error;
 609
 610	/*
 611	 * Reset the bio to free any attached resources
 612	 * (e.g. bio integrity profiles).
 613	 */
 614	bio_reset(bio);
 615
 616	bio->bi_error = error;
 617	end_fn(bio);
 618}
 619
 620static void use_inline_bio(struct dm_buffer *b, int rw, sector_t block,
 621			   bio_end_io_t *end_io)
 622{
 623	char *ptr;
 624	int len;
 625
 626	bio_init(&b->bio);
 627	b->bio.bi_io_vec = b->bio_vec;
 628	b->bio.bi_max_vecs = DM_BUFIO_INLINE_VECS;
 629	b->bio.bi_iter.bi_sector = block << b->c->sectors_per_block_bits;
 630	b->bio.bi_bdev = b->c->bdev;
 631	b->bio.bi_end_io = inline_endio;
 632	/*
 633	 * Use of .bi_private isn't a problem here because
 634	 * the dm_buffer's inline bio is local to bufio.
 635	 */
 636	b->bio.bi_private = end_io;
 637
 638	/*
 639	 * We assume that if len >= PAGE_SIZE ptr is page-aligned.
 640	 * If len < PAGE_SIZE the buffer doesn't cross page boundary.
 641	 */
 642	ptr = b->data;
 643	len = b->c->block_size;
 644
 645	if (len >= PAGE_SIZE)
 646		BUG_ON((unsigned long)ptr & (PAGE_SIZE - 1));
 647	else
 648		BUG_ON((unsigned long)ptr & (len - 1));
 649
 650	do {
 651		if (!bio_add_page(&b->bio, virt_to_page(ptr),
 652				  len < PAGE_SIZE ? len : PAGE_SIZE,
 653				  offset_in_page(ptr))) {
 654			BUG_ON(b->c->block_size <= PAGE_SIZE);
 655			use_dmio(b, rw, block, end_io);
 656			return;
 657		}
 658
 659		len -= PAGE_SIZE;
 660		ptr += PAGE_SIZE;
 661	} while (len > 0);
 662
 663	submit_bio(rw, &b->bio);
 664}
 665
 666static void submit_io(struct dm_buffer *b, int rw, sector_t block,
 667		      bio_end_io_t *end_io)
 668{
 669	if (rw == WRITE && b->c->write_callback)
 670		b->c->write_callback(b);
 671
 672	if (b->c->block_size <= DM_BUFIO_INLINE_VECS * PAGE_SIZE &&
 673	    b->data_mode != DATA_MODE_VMALLOC)
 674		use_inline_bio(b, rw, block, end_io);
 675	else
 676		use_dmio(b, rw, block, end_io);
 677}
 678
 679/*----------------------------------------------------------------
 680 * Writing dirty buffers
 681 *--------------------------------------------------------------*/
 682
 683/*
 684 * The endio routine for write.
 685 *
 686 * Set the error, clear B_WRITING bit and wake anyone who was waiting on
 687 * it.
 688 */
 689static void write_endio(struct bio *bio)
 690{
 691	struct dm_buffer *b = container_of(bio, struct dm_buffer, bio);
 692
 693	b->write_error = bio->bi_error;
 694	if (unlikely(bio->bi_error)) {
 695		struct dm_bufio_client *c = b->c;
 696		int error = bio->bi_error;
 697		(void)cmpxchg(&c->async_write_error, 0, error);
 698	}
 699
 700	BUG_ON(!test_bit(B_WRITING, &b->state));
 701
 702	smp_mb__before_atomic();
 703	clear_bit(B_WRITING, &b->state);
 704	smp_mb__after_atomic();
 705
 706	wake_up_bit(&b->state, B_WRITING);
 707}
 708
 709/*
 
 
 
 
 
 
 
 
 
 
 710 * Initiate a write on a dirty buffer, but don't wait for it.
 711 *
 712 * - If the buffer is not dirty, exit.
 713 * - If there some previous write going on, wait for it to finish (we can't
 714 *   have two writes on the same buffer simultaneously).
 715 * - Submit our write and don't wait on it. We set B_WRITING indicating
 716 *   that there is a write in progress.
 717 */
 718static void __write_dirty_buffer(struct dm_buffer *b,
 719				 struct list_head *write_list)
 720{
 721	if (!test_bit(B_DIRTY, &b->state))
 722		return;
 723
 724	clear_bit(B_DIRTY, &b->state);
 725	wait_on_bit_lock_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
 
 726
 727	if (!write_list)
 728		submit_io(b, WRITE, b->block, write_endio);
 729	else
 730		list_add_tail(&b->write_list, write_list);
 731}
 732
 733static void __flush_write_list(struct list_head *write_list)
 734{
 735	struct blk_plug plug;
 736	blk_start_plug(&plug);
 737	while (!list_empty(write_list)) {
 738		struct dm_buffer *b =
 739			list_entry(write_list->next, struct dm_buffer, write_list);
 740		list_del(&b->write_list);
 741		submit_io(b, WRITE, b->block, write_endio);
 742		dm_bufio_cond_resched();
 743	}
 744	blk_finish_plug(&plug);
 745}
 746
 747/*
 748 * Wait until any activity on the buffer finishes.  Possibly write the
 749 * buffer if it is dirty.  When this function finishes, there is no I/O
 750 * running on the buffer and the buffer is not dirty.
 751 */
 752static void __make_buffer_clean(struct dm_buffer *b)
 753{
 754	BUG_ON(b->hold_count);
 755
 756	if (!b->state)	/* fast case */
 757		return;
 758
 759	wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE);
 760	__write_dirty_buffer(b, NULL);
 761	wait_on_bit_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
 762}
 763
 764/*
 765 * Find some buffer that is not held by anybody, clean it, unlink it and
 766 * return it.
 767 */
 768static struct dm_buffer *__get_unclaimed_buffer(struct dm_bufio_client *c)
 769{
 770	struct dm_buffer *b;
 771
 772	list_for_each_entry_reverse(b, &c->lru[LIST_CLEAN], lru_list) {
 773		BUG_ON(test_bit(B_WRITING, &b->state));
 774		BUG_ON(test_bit(B_DIRTY, &b->state));
 775
 776		if (!b->hold_count) {
 777			__make_buffer_clean(b);
 778			__unlink_buffer(b);
 779			return b;
 780		}
 781		dm_bufio_cond_resched();
 782	}
 783
 784	list_for_each_entry_reverse(b, &c->lru[LIST_DIRTY], lru_list) {
 785		BUG_ON(test_bit(B_READING, &b->state));
 786
 787		if (!b->hold_count) {
 788			__make_buffer_clean(b);
 789			__unlink_buffer(b);
 790			return b;
 791		}
 792		dm_bufio_cond_resched();
 793	}
 794
 795	return NULL;
 796}
 797
 798/*
 799 * Wait until some other threads free some buffer or release hold count on
 800 * some buffer.
 801 *
 802 * This function is entered with c->lock held, drops it and regains it
 803 * before exiting.
 804 */
 805static void __wait_for_free_buffer(struct dm_bufio_client *c)
 806{
 807	DECLARE_WAITQUEUE(wait, current);
 808
 809	add_wait_queue(&c->free_buffer_wait, &wait);
 810	set_task_state(current, TASK_UNINTERRUPTIBLE);
 811	dm_bufio_unlock(c);
 812
 813	io_schedule();
 814
 
 815	remove_wait_queue(&c->free_buffer_wait, &wait);
 816
 817	dm_bufio_lock(c);
 818}
 819
 820enum new_flag {
 821	NF_FRESH = 0,
 822	NF_READ = 1,
 823	NF_GET = 2,
 824	NF_PREFETCH = 3
 825};
 826
 827/*
 828 * Allocate a new buffer. If the allocation is not possible, wait until
 829 * some other thread frees a buffer.
 830 *
 831 * May drop the lock and regain it.
 832 */
 833static struct dm_buffer *__alloc_buffer_wait_no_callback(struct dm_bufio_client *c, enum new_flag nf)
 834{
 835	struct dm_buffer *b;
 836
 837	/*
 838	 * dm-bufio is resistant to allocation failures (it just keeps
 839	 * one buffer reserved in cases all the allocations fail).
 840	 * So set flags to not try too hard:
 841	 *	GFP_NOIO: don't recurse into the I/O layer
 842	 *	__GFP_NORETRY: don't retry and rather return failure
 843	 *	__GFP_NOMEMALLOC: don't use emergency reserves
 844	 *	__GFP_NOWARN: don't print a warning in case of failure
 845	 *
 846	 * For debugging, if we set the cache size to 1, no new buffers will
 847	 * be allocated.
 848	 */
 849	while (1) {
 850		if (dm_bufio_cache_size_latch != 1) {
 851			b = alloc_buffer(c, GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
 852			if (b)
 853				return b;
 854		}
 855
 856		if (nf == NF_PREFETCH)
 857			return NULL;
 858
 859		if (!list_empty(&c->reserved_buffers)) {
 860			b = list_entry(c->reserved_buffers.next,
 861				       struct dm_buffer, lru_list);
 862			list_del(&b->lru_list);
 863			c->need_reserved_buffers++;
 864
 865			return b;
 866		}
 867
 868		b = __get_unclaimed_buffer(c);
 869		if (b)
 870			return b;
 871
 872		__wait_for_free_buffer(c);
 873	}
 874}
 875
 876static struct dm_buffer *__alloc_buffer_wait(struct dm_bufio_client *c, enum new_flag nf)
 877{
 878	struct dm_buffer *b = __alloc_buffer_wait_no_callback(c, nf);
 879
 880	if (!b)
 881		return NULL;
 882
 883	if (c->alloc_callback)
 884		c->alloc_callback(b);
 885
 886	return b;
 887}
 888
 889/*
 890 * Free a buffer and wake other threads waiting for free buffers.
 891 */
 892static void __free_buffer_wake(struct dm_buffer *b)
 893{
 894	struct dm_bufio_client *c = b->c;
 895
 896	if (!c->need_reserved_buffers)
 897		free_buffer(b);
 898	else {
 899		list_add(&b->lru_list, &c->reserved_buffers);
 900		c->need_reserved_buffers--;
 901	}
 902
 903	wake_up(&c->free_buffer_wait);
 904}
 905
 906static void __write_dirty_buffers_async(struct dm_bufio_client *c, int no_wait,
 907					struct list_head *write_list)
 908{
 909	struct dm_buffer *b, *tmp;
 910
 911	list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
 912		BUG_ON(test_bit(B_READING, &b->state));
 913
 914		if (!test_bit(B_DIRTY, &b->state) &&
 915		    !test_bit(B_WRITING, &b->state)) {
 916			__relink_lru(b, LIST_CLEAN);
 917			continue;
 918		}
 919
 920		if (no_wait && test_bit(B_WRITING, &b->state))
 921			return;
 922
 923		__write_dirty_buffer(b, write_list);
 924		dm_bufio_cond_resched();
 925	}
 926}
 927
 928/*
 929 * Get writeback threshold and buffer limit for a given client.
 930 */
 931static void __get_memory_limit(struct dm_bufio_client *c,
 932			       unsigned long *threshold_buffers,
 933			       unsigned long *limit_buffers)
 934{
 935	unsigned long buffers;
 936
 937	if (ACCESS_ONCE(dm_bufio_cache_size) != dm_bufio_cache_size_latch) {
 938		mutex_lock(&dm_bufio_clients_lock);
 939		__cache_size_refresh();
 940		mutex_unlock(&dm_bufio_clients_lock);
 941	}
 942
 943	buffers = dm_bufio_cache_size_per_client >>
 944		  (c->sectors_per_block_bits + SECTOR_SHIFT);
 945
 946	if (buffers < c->minimum_buffers)
 947		buffers = c->minimum_buffers;
 948
 949	*limit_buffers = buffers;
 950	*threshold_buffers = buffers * DM_BUFIO_WRITEBACK_PERCENT / 100;
 951}
 952
 953/*
 954 * Check if we're over watermark.
 955 * If we are over threshold_buffers, start freeing buffers.
 956 * If we're over "limit_buffers", block until we get under the limit.
 957 */
 958static void __check_watermark(struct dm_bufio_client *c,
 959			      struct list_head *write_list)
 960{
 961	unsigned long threshold_buffers, limit_buffers;
 962
 963	__get_memory_limit(c, &threshold_buffers, &limit_buffers);
 964
 965	while (c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY] >
 966	       limit_buffers) {
 967
 968		struct dm_buffer *b = __get_unclaimed_buffer(c);
 969
 970		if (!b)
 971			return;
 972
 973		__free_buffer_wake(b);
 974		dm_bufio_cond_resched();
 975	}
 976
 977	if (c->n_buffers[LIST_DIRTY] > threshold_buffers)
 978		__write_dirty_buffers_async(c, 1, write_list);
 979}
 980
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 981/*----------------------------------------------------------------
 982 * Getting a buffer
 983 *--------------------------------------------------------------*/
 984
 985static struct dm_buffer *__bufio_new(struct dm_bufio_client *c, sector_t block,
 986				     enum new_flag nf, int *need_submit,
 987				     struct list_head *write_list)
 988{
 989	struct dm_buffer *b, *new_b = NULL;
 990
 991	*need_submit = 0;
 992
 993	b = __find(c, block);
 994	if (b)
 995		goto found_buffer;
 996
 997	if (nf == NF_GET)
 998		return NULL;
 999
1000	new_b = __alloc_buffer_wait(c, nf);
1001	if (!new_b)
1002		return NULL;
1003
1004	/*
1005	 * We've had a period where the mutex was unlocked, so need to
1006	 * recheck the hash table.
1007	 */
1008	b = __find(c, block);
1009	if (b) {
1010		__free_buffer_wake(new_b);
1011		goto found_buffer;
1012	}
1013
1014	__check_watermark(c, write_list);
1015
1016	b = new_b;
1017	b->hold_count = 1;
1018	b->read_error = 0;
1019	b->write_error = 0;
1020	__link_buffer(b, block, LIST_CLEAN);
1021
1022	if (nf == NF_FRESH) {
1023		b->state = 0;
1024		return b;
1025	}
1026
1027	b->state = 1 << B_READING;
1028	*need_submit = 1;
1029
1030	return b;
1031
1032found_buffer:
1033	if (nf == NF_PREFETCH)
1034		return NULL;
1035	/*
1036	 * Note: it is essential that we don't wait for the buffer to be
1037	 * read if dm_bufio_get function is used. Both dm_bufio_get and
1038	 * dm_bufio_prefetch can be used in the driver request routine.
1039	 * If the user called both dm_bufio_prefetch and dm_bufio_get on
1040	 * the same buffer, it would deadlock if we waited.
1041	 */
1042	if (nf == NF_GET && unlikely(test_bit(B_READING, &b->state)))
1043		return NULL;
1044
1045	b->hold_count++;
1046	__relink_lru(b, test_bit(B_DIRTY, &b->state) ||
1047		     test_bit(B_WRITING, &b->state));
1048	return b;
1049}
1050
1051/*
1052 * The endio routine for reading: set the error, clear the bit and wake up
1053 * anyone waiting on the buffer.
1054 */
1055static void read_endio(struct bio *bio)
1056{
1057	struct dm_buffer *b = container_of(bio, struct dm_buffer, bio);
1058
1059	b->read_error = bio->bi_error;
1060
1061	BUG_ON(!test_bit(B_READING, &b->state));
1062
1063	smp_mb__before_atomic();
1064	clear_bit(B_READING, &b->state);
1065	smp_mb__after_atomic();
1066
1067	wake_up_bit(&b->state, B_READING);
1068}
1069
1070/*
1071 * A common routine for dm_bufio_new and dm_bufio_read.  Operation of these
1072 * functions is similar except that dm_bufio_new doesn't read the
1073 * buffer from the disk (assuming that the caller overwrites all the data
1074 * and uses dm_bufio_mark_buffer_dirty to write new data back).
1075 */
1076static void *new_read(struct dm_bufio_client *c, sector_t block,
1077		      enum new_flag nf, struct dm_buffer **bp)
1078{
1079	int need_submit;
1080	struct dm_buffer *b;
1081
1082	LIST_HEAD(write_list);
1083
1084	dm_bufio_lock(c);
1085	b = __bufio_new(c, block, nf, &need_submit, &write_list);
1086#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1087	if (b && b->hold_count == 1)
1088		buffer_record_stack(b);
1089#endif
1090	dm_bufio_unlock(c);
1091
1092	__flush_write_list(&write_list);
1093
1094	if (!b)
1095		return NULL;
1096
1097	if (need_submit)
1098		submit_io(b, READ, b->block, read_endio);
1099
1100	wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE);
1101
1102	if (b->read_error) {
1103		int error = b->read_error;
1104
1105		dm_bufio_release(b);
1106
1107		return ERR_PTR(error);
1108	}
1109
1110	*bp = b;
1111
1112	return b->data;
1113}
1114
1115void *dm_bufio_get(struct dm_bufio_client *c, sector_t block,
1116		   struct dm_buffer **bp)
1117{
1118	return new_read(c, block, NF_GET, bp);
1119}
1120EXPORT_SYMBOL_GPL(dm_bufio_get);
1121
1122void *dm_bufio_read(struct dm_bufio_client *c, sector_t block,
1123		    struct dm_buffer **bp)
1124{
1125	BUG_ON(dm_bufio_in_request());
1126
1127	return new_read(c, block, NF_READ, bp);
1128}
1129EXPORT_SYMBOL_GPL(dm_bufio_read);
1130
1131void *dm_bufio_new(struct dm_bufio_client *c, sector_t block,
1132		   struct dm_buffer **bp)
1133{
1134	BUG_ON(dm_bufio_in_request());
1135
1136	return new_read(c, block, NF_FRESH, bp);
1137}
1138EXPORT_SYMBOL_GPL(dm_bufio_new);
1139
1140void dm_bufio_prefetch(struct dm_bufio_client *c,
1141		       sector_t block, unsigned n_blocks)
1142{
1143	struct blk_plug plug;
1144
1145	LIST_HEAD(write_list);
1146
1147	BUG_ON(dm_bufio_in_request());
1148
1149	blk_start_plug(&plug);
1150	dm_bufio_lock(c);
1151
1152	for (; n_blocks--; block++) {
1153		int need_submit;
1154		struct dm_buffer *b;
1155		b = __bufio_new(c, block, NF_PREFETCH, &need_submit,
1156				&write_list);
1157		if (unlikely(!list_empty(&write_list))) {
1158			dm_bufio_unlock(c);
1159			blk_finish_plug(&plug);
1160			__flush_write_list(&write_list);
1161			blk_start_plug(&plug);
1162			dm_bufio_lock(c);
1163		}
1164		if (unlikely(b != NULL)) {
1165			dm_bufio_unlock(c);
1166
1167			if (need_submit)
1168				submit_io(b, READ, b->block, read_endio);
1169			dm_bufio_release(b);
1170
1171			dm_bufio_cond_resched();
1172
1173			if (!n_blocks)
1174				goto flush_plug;
1175			dm_bufio_lock(c);
1176		}
1177	}
1178
1179	dm_bufio_unlock(c);
1180
1181flush_plug:
1182	blk_finish_plug(&plug);
1183}
1184EXPORT_SYMBOL_GPL(dm_bufio_prefetch);
1185
1186void dm_bufio_release(struct dm_buffer *b)
1187{
1188	struct dm_bufio_client *c = b->c;
1189
1190	dm_bufio_lock(c);
1191
1192	BUG_ON(!b->hold_count);
1193
1194	b->hold_count--;
1195	if (!b->hold_count) {
1196		wake_up(&c->free_buffer_wait);
1197
1198		/*
1199		 * If there were errors on the buffer, and the buffer is not
1200		 * to be written, free the buffer. There is no point in caching
1201		 * invalid buffer.
1202		 */
1203		if ((b->read_error || b->write_error) &&
1204		    !test_bit(B_READING, &b->state) &&
1205		    !test_bit(B_WRITING, &b->state) &&
1206		    !test_bit(B_DIRTY, &b->state)) {
1207			__unlink_buffer(b);
1208			__free_buffer_wake(b);
1209		}
1210	}
1211
1212	dm_bufio_unlock(c);
1213}
1214EXPORT_SYMBOL_GPL(dm_bufio_release);
1215
1216void dm_bufio_mark_buffer_dirty(struct dm_buffer *b)
1217{
1218	struct dm_bufio_client *c = b->c;
1219
1220	dm_bufio_lock(c);
1221
1222	BUG_ON(test_bit(B_READING, &b->state));
1223
1224	if (!test_and_set_bit(B_DIRTY, &b->state))
1225		__relink_lru(b, LIST_DIRTY);
1226
1227	dm_bufio_unlock(c);
1228}
1229EXPORT_SYMBOL_GPL(dm_bufio_mark_buffer_dirty);
1230
1231void dm_bufio_write_dirty_buffers_async(struct dm_bufio_client *c)
1232{
1233	LIST_HEAD(write_list);
1234
1235	BUG_ON(dm_bufio_in_request());
1236
1237	dm_bufio_lock(c);
1238	__write_dirty_buffers_async(c, 0, &write_list);
1239	dm_bufio_unlock(c);
1240	__flush_write_list(&write_list);
1241}
1242EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers_async);
1243
1244/*
1245 * For performance, it is essential that the buffers are written asynchronously
1246 * and simultaneously (so that the block layer can merge the writes) and then
1247 * waited upon.
1248 *
1249 * Finally, we flush hardware disk cache.
1250 */
1251int dm_bufio_write_dirty_buffers(struct dm_bufio_client *c)
1252{
1253	int a, f;
1254	unsigned long buffers_processed = 0;
1255	struct dm_buffer *b, *tmp;
1256
1257	LIST_HEAD(write_list);
1258
1259	dm_bufio_lock(c);
1260	__write_dirty_buffers_async(c, 0, &write_list);
1261	dm_bufio_unlock(c);
1262	__flush_write_list(&write_list);
1263	dm_bufio_lock(c);
1264
1265again:
1266	list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
1267		int dropped_lock = 0;
1268
1269		if (buffers_processed < c->n_buffers[LIST_DIRTY])
1270			buffers_processed++;
1271
1272		BUG_ON(test_bit(B_READING, &b->state));
1273
1274		if (test_bit(B_WRITING, &b->state)) {
1275			if (buffers_processed < c->n_buffers[LIST_DIRTY]) {
1276				dropped_lock = 1;
1277				b->hold_count++;
1278				dm_bufio_unlock(c);
1279				wait_on_bit_io(&b->state, B_WRITING,
1280					       TASK_UNINTERRUPTIBLE);
 
1281				dm_bufio_lock(c);
1282				b->hold_count--;
1283			} else
1284				wait_on_bit_io(&b->state, B_WRITING,
1285					       TASK_UNINTERRUPTIBLE);
 
1286		}
1287
1288		if (!test_bit(B_DIRTY, &b->state) &&
1289		    !test_bit(B_WRITING, &b->state))
1290			__relink_lru(b, LIST_CLEAN);
1291
1292		dm_bufio_cond_resched();
1293
1294		/*
1295		 * If we dropped the lock, the list is no longer consistent,
1296		 * so we must restart the search.
1297		 *
1298		 * In the most common case, the buffer just processed is
1299		 * relinked to the clean list, so we won't loop scanning the
1300		 * same buffer again and again.
1301		 *
1302		 * This may livelock if there is another thread simultaneously
1303		 * dirtying buffers, so we count the number of buffers walked
1304		 * and if it exceeds the total number of buffers, it means that
1305		 * someone is doing some writes simultaneously with us.  In
1306		 * this case, stop, dropping the lock.
1307		 */
1308		if (dropped_lock)
1309			goto again;
1310	}
1311	wake_up(&c->free_buffer_wait);
1312	dm_bufio_unlock(c);
1313
1314	a = xchg(&c->async_write_error, 0);
1315	f = dm_bufio_issue_flush(c);
1316	if (a)
1317		return a;
1318
1319	return f;
1320}
1321EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers);
1322
1323/*
1324 * Use dm-io to send and empty barrier flush the device.
1325 */
1326int dm_bufio_issue_flush(struct dm_bufio_client *c)
1327{
1328	struct dm_io_request io_req = {
1329		.bi_rw = WRITE_FLUSH,
1330		.mem.type = DM_IO_KMEM,
1331		.mem.ptr.addr = NULL,
1332		.client = c->dm_io,
1333	};
1334	struct dm_io_region io_reg = {
1335		.bdev = c->bdev,
1336		.sector = 0,
1337		.count = 0,
1338	};
1339
1340	BUG_ON(dm_bufio_in_request());
1341
1342	return dm_io(&io_req, 1, &io_reg, NULL);
1343}
1344EXPORT_SYMBOL_GPL(dm_bufio_issue_flush);
1345
1346/*
1347 * We first delete any other buffer that may be at that new location.
1348 *
1349 * Then, we write the buffer to the original location if it was dirty.
1350 *
1351 * Then, if we are the only one who is holding the buffer, relink the buffer
1352 * in the hash queue for the new location.
1353 *
1354 * If there was someone else holding the buffer, we write it to the new
1355 * location but not relink it, because that other user needs to have the buffer
1356 * at the same place.
1357 */
1358void dm_bufio_release_move(struct dm_buffer *b, sector_t new_block)
1359{
1360	struct dm_bufio_client *c = b->c;
1361	struct dm_buffer *new;
1362
1363	BUG_ON(dm_bufio_in_request());
1364
1365	dm_bufio_lock(c);
1366
1367retry:
1368	new = __find(c, new_block);
1369	if (new) {
1370		if (new->hold_count) {
1371			__wait_for_free_buffer(c);
1372			goto retry;
1373		}
1374
1375		/*
1376		 * FIXME: Is there any point waiting for a write that's going
1377		 * to be overwritten in a bit?
1378		 */
1379		__make_buffer_clean(new);
1380		__unlink_buffer(new);
1381		__free_buffer_wake(new);
1382	}
1383
1384	BUG_ON(!b->hold_count);
1385	BUG_ON(test_bit(B_READING, &b->state));
1386
1387	__write_dirty_buffer(b, NULL);
1388	if (b->hold_count == 1) {
1389		wait_on_bit_io(&b->state, B_WRITING,
1390			       TASK_UNINTERRUPTIBLE);
1391		set_bit(B_DIRTY, &b->state);
1392		__unlink_buffer(b);
1393		__link_buffer(b, new_block, LIST_DIRTY);
1394	} else {
1395		sector_t old_block;
1396		wait_on_bit_lock_io(&b->state, B_WRITING,
1397				    TASK_UNINTERRUPTIBLE);
1398		/*
1399		 * Relink buffer to "new_block" so that write_callback
1400		 * sees "new_block" as a block number.
1401		 * After the write, link the buffer back to old_block.
1402		 * All this must be done in bufio lock, so that block number
1403		 * change isn't visible to other threads.
1404		 */
1405		old_block = b->block;
1406		__unlink_buffer(b);
1407		__link_buffer(b, new_block, b->list_mode);
1408		submit_io(b, WRITE, new_block, write_endio);
1409		wait_on_bit_io(&b->state, B_WRITING,
1410			       TASK_UNINTERRUPTIBLE);
1411		__unlink_buffer(b);
1412		__link_buffer(b, old_block, b->list_mode);
1413	}
1414
1415	dm_bufio_unlock(c);
1416	dm_bufio_release(b);
1417}
1418EXPORT_SYMBOL_GPL(dm_bufio_release_move);
1419
1420/*
1421 * Free the given buffer.
1422 *
1423 * This is just a hint, if the buffer is in use or dirty, this function
1424 * does nothing.
1425 */
1426void dm_bufio_forget(struct dm_bufio_client *c, sector_t block)
1427{
1428	struct dm_buffer *b;
1429
1430	dm_bufio_lock(c);
1431
1432	b = __find(c, block);
1433	if (b && likely(!b->hold_count) && likely(!b->state)) {
1434		__unlink_buffer(b);
1435		__free_buffer_wake(b);
1436	}
1437
1438	dm_bufio_unlock(c);
1439}
1440EXPORT_SYMBOL(dm_bufio_forget);
1441
1442void dm_bufio_set_minimum_buffers(struct dm_bufio_client *c, unsigned n)
1443{
1444	c->minimum_buffers = n;
1445}
1446EXPORT_SYMBOL(dm_bufio_set_minimum_buffers);
1447
1448unsigned dm_bufio_get_block_size(struct dm_bufio_client *c)
1449{
1450	return c->block_size;
1451}
1452EXPORT_SYMBOL_GPL(dm_bufio_get_block_size);
1453
1454sector_t dm_bufio_get_device_size(struct dm_bufio_client *c)
1455{
1456	return i_size_read(c->bdev->bd_inode) >>
1457			   (SECTOR_SHIFT + c->sectors_per_block_bits);
1458}
1459EXPORT_SYMBOL_GPL(dm_bufio_get_device_size);
1460
1461sector_t dm_bufio_get_block_number(struct dm_buffer *b)
1462{
1463	return b->block;
1464}
1465EXPORT_SYMBOL_GPL(dm_bufio_get_block_number);
1466
1467void *dm_bufio_get_block_data(struct dm_buffer *b)
1468{
1469	return b->data;
1470}
1471EXPORT_SYMBOL_GPL(dm_bufio_get_block_data);
1472
1473void *dm_bufio_get_aux_data(struct dm_buffer *b)
1474{
1475	return b + 1;
1476}
1477EXPORT_SYMBOL_GPL(dm_bufio_get_aux_data);
1478
1479struct dm_bufio_client *dm_bufio_get_client(struct dm_buffer *b)
1480{
1481	return b->c;
1482}
1483EXPORT_SYMBOL_GPL(dm_bufio_get_client);
1484
1485static void drop_buffers(struct dm_bufio_client *c)
1486{
1487	struct dm_buffer *b;
1488	int i;
1489	bool warned = false;
1490
1491	BUG_ON(dm_bufio_in_request());
1492
1493	/*
1494	 * An optimization so that the buffers are not written one-by-one.
1495	 */
1496	dm_bufio_write_dirty_buffers_async(c);
1497
1498	dm_bufio_lock(c);
1499
1500	while ((b = __get_unclaimed_buffer(c)))
1501		__free_buffer_wake(b);
1502
1503	for (i = 0; i < LIST_SIZE; i++)
1504		list_for_each_entry(b, &c->lru[i], lru_list) {
1505			WARN_ON(!warned);
1506			warned = true;
1507			DMERR("leaked buffer %llx, hold count %u, list %d",
1508			      (unsigned long long)b->block, b->hold_count, i);
1509#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1510			print_stack_trace(&b->stack_trace, 1);
1511			b->hold_count = 0; /* mark unclaimed to avoid BUG_ON below */
1512#endif
1513		}
1514
1515#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1516	while ((b = __get_unclaimed_buffer(c)))
1517		__free_buffer_wake(b);
1518#endif
1519
1520	for (i = 0; i < LIST_SIZE; i++)
1521		BUG_ON(!list_empty(&c->lru[i]));
1522
1523	dm_bufio_unlock(c);
1524}
1525
1526/*
1527 * We may not be able to evict this buffer if IO pending or the client
1528 * is still using it.  Caller is expected to know buffer is too old.
1529 *
1530 * And if GFP_NOFS is used, we must not do any I/O because we hold
1531 * dm_bufio_clients_lock and we would risk deadlock if the I/O gets
1532 * rerouted to different bufio client.
1533 */
1534static bool __try_evict_buffer(struct dm_buffer *b, gfp_t gfp)
 
1535{
1536	if (!(gfp & __GFP_FS)) {
 
 
 
1537		if (test_bit(B_READING, &b->state) ||
1538		    test_bit(B_WRITING, &b->state) ||
1539		    test_bit(B_DIRTY, &b->state))
1540			return false;
1541	}
1542
1543	if (b->hold_count)
1544		return false;
1545
1546	__make_buffer_clean(b);
1547	__unlink_buffer(b);
1548	__free_buffer_wake(b);
1549
1550	return true;
1551}
1552
1553static unsigned get_retain_buffers(struct dm_bufio_client *c)
1554{
1555        unsigned retain_bytes = ACCESS_ONCE(dm_bufio_retain_bytes);
1556        return retain_bytes / c->block_size;
1557}
1558
1559static unsigned long __scan(struct dm_bufio_client *c, unsigned long nr_to_scan,
1560			    gfp_t gfp_mask)
1561{
1562	int l;
1563	struct dm_buffer *b, *tmp;
1564	unsigned long freed = 0;
1565	unsigned long count = nr_to_scan;
1566	unsigned retain_target = get_retain_buffers(c);
1567
1568	for (l = 0; l < LIST_SIZE; l++) {
1569		list_for_each_entry_safe_reverse(b, tmp, &c->lru[l], lru_list) {
1570			if (__try_evict_buffer(b, gfp_mask))
1571				freed++;
1572			if (!--nr_to_scan || ((count - freed) <= retain_target))
1573				return freed;
1574			dm_bufio_cond_resched();
1575		}
 
1576	}
1577	return freed;
1578}
1579
1580static unsigned long
1581dm_bufio_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
1582{
1583	struct dm_bufio_client *c;
1584	unsigned long freed;
1585
1586	c = container_of(shrink, struct dm_bufio_client, shrinker);
1587	if (sc->gfp_mask & __GFP_FS)
1588		dm_bufio_lock(c);
1589	else if (!dm_bufio_trylock(c))
1590		return SHRINK_STOP;
1591
1592	freed  = __scan(c, sc->nr_to_scan, sc->gfp_mask);
1593	dm_bufio_unlock(c);
1594	return freed;
1595}
1596
1597static unsigned long
1598dm_bufio_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
1599{
1600	struct dm_bufio_client *c;
1601	unsigned long count;
1602
1603	c = container_of(shrink, struct dm_bufio_client, shrinker);
1604	if (sc->gfp_mask & __GFP_FS)
1605		dm_bufio_lock(c);
1606	else if (!dm_bufio_trylock(c))
1607		return 0;
1608
1609	count = c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY];
1610	dm_bufio_unlock(c);
1611	return count;
1612}
1613
1614/*
1615 * Create the buffering interface
1616 */
1617struct dm_bufio_client *dm_bufio_client_create(struct block_device *bdev, unsigned block_size,
1618					       unsigned reserved_buffers, unsigned aux_size,
1619					       void (*alloc_callback)(struct dm_buffer *),
1620					       void (*write_callback)(struct dm_buffer *))
1621{
1622	int r;
1623	struct dm_bufio_client *c;
1624	unsigned i;
1625
1626	BUG_ON(block_size < 1 << SECTOR_SHIFT ||
1627	       (block_size & (block_size - 1)));
1628
1629	c = kzalloc(sizeof(*c), GFP_KERNEL);
1630	if (!c) {
1631		r = -ENOMEM;
1632		goto bad_client;
1633	}
1634	c->buffer_tree = RB_ROOT;
 
 
 
 
1635
1636	c->bdev = bdev;
1637	c->block_size = block_size;
1638	c->sectors_per_block_bits = __ffs(block_size) - SECTOR_SHIFT;
1639	c->pages_per_block_bits = (__ffs(block_size) >= PAGE_SHIFT) ?
1640				  __ffs(block_size) - PAGE_SHIFT : 0;
1641	c->blocks_per_page_bits = (__ffs(block_size) < PAGE_SHIFT ?
1642				  PAGE_SHIFT - __ffs(block_size) : 0);
1643
1644	c->aux_size = aux_size;
1645	c->alloc_callback = alloc_callback;
1646	c->write_callback = write_callback;
1647
1648	for (i = 0; i < LIST_SIZE; i++) {
1649		INIT_LIST_HEAD(&c->lru[i]);
1650		c->n_buffers[i] = 0;
1651	}
1652
 
 
 
1653	mutex_init(&c->lock);
1654	INIT_LIST_HEAD(&c->reserved_buffers);
1655	c->need_reserved_buffers = reserved_buffers;
1656
1657	c->minimum_buffers = DM_BUFIO_MIN_BUFFERS;
1658
1659	init_waitqueue_head(&c->free_buffer_wait);
1660	c->async_write_error = 0;
1661
1662	c->dm_io = dm_io_client_create();
1663	if (IS_ERR(c->dm_io)) {
1664		r = PTR_ERR(c->dm_io);
1665		goto bad_dm_io;
1666	}
1667
1668	mutex_lock(&dm_bufio_clients_lock);
1669	if (c->blocks_per_page_bits) {
1670		if (!DM_BUFIO_CACHE_NAME(c)) {
1671			DM_BUFIO_CACHE_NAME(c) = kasprintf(GFP_KERNEL, "dm_bufio_cache-%u", c->block_size);
1672			if (!DM_BUFIO_CACHE_NAME(c)) {
1673				r = -ENOMEM;
1674				mutex_unlock(&dm_bufio_clients_lock);
1675				goto bad_cache;
1676			}
1677		}
1678
1679		if (!DM_BUFIO_CACHE(c)) {
1680			DM_BUFIO_CACHE(c) = kmem_cache_create(DM_BUFIO_CACHE_NAME(c),
1681							      c->block_size,
1682							      c->block_size, 0, NULL);
1683			if (!DM_BUFIO_CACHE(c)) {
1684				r = -ENOMEM;
1685				mutex_unlock(&dm_bufio_clients_lock);
1686				goto bad_cache;
1687			}
1688		}
1689	}
1690	mutex_unlock(&dm_bufio_clients_lock);
1691
1692	while (c->need_reserved_buffers) {
1693		struct dm_buffer *b = alloc_buffer(c, GFP_KERNEL);
1694
1695		if (!b) {
1696			r = -ENOMEM;
1697			goto bad_buffer;
1698		}
1699		__free_buffer_wake(b);
1700	}
1701
1702	mutex_lock(&dm_bufio_clients_lock);
1703	dm_bufio_client_count++;
1704	list_add(&c->client_list, &dm_bufio_all_clients);
1705	__cache_size_refresh();
1706	mutex_unlock(&dm_bufio_clients_lock);
1707
1708	c->shrinker.count_objects = dm_bufio_shrink_count;
1709	c->shrinker.scan_objects = dm_bufio_shrink_scan;
1710	c->shrinker.seeks = 1;
1711	c->shrinker.batch = 0;
1712	register_shrinker(&c->shrinker);
1713
1714	return c;
1715
1716bad_buffer:
1717bad_cache:
1718	while (!list_empty(&c->reserved_buffers)) {
1719		struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1720						 struct dm_buffer, lru_list);
1721		list_del(&b->lru_list);
1722		free_buffer(b);
1723	}
1724	dm_io_client_destroy(c->dm_io);
1725bad_dm_io:
 
 
1726	kfree(c);
1727bad_client:
1728	return ERR_PTR(r);
1729}
1730EXPORT_SYMBOL_GPL(dm_bufio_client_create);
1731
1732/*
1733 * Free the buffering interface.
1734 * It is required that there are no references on any buffers.
1735 */
1736void dm_bufio_client_destroy(struct dm_bufio_client *c)
1737{
1738	unsigned i;
1739
1740	drop_buffers(c);
1741
1742	unregister_shrinker(&c->shrinker);
1743
1744	mutex_lock(&dm_bufio_clients_lock);
1745
1746	list_del(&c->client_list);
1747	dm_bufio_client_count--;
1748	__cache_size_refresh();
1749
1750	mutex_unlock(&dm_bufio_clients_lock);
1751
1752	BUG_ON(!RB_EMPTY_ROOT(&c->buffer_tree));
 
 
1753	BUG_ON(c->need_reserved_buffers);
1754
1755	while (!list_empty(&c->reserved_buffers)) {
1756		struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1757						 struct dm_buffer, lru_list);
1758		list_del(&b->lru_list);
1759		free_buffer(b);
1760	}
1761
1762	for (i = 0; i < LIST_SIZE; i++)
1763		if (c->n_buffers[i])
1764			DMERR("leaked buffer count %d: %ld", i, c->n_buffers[i]);
1765
1766	for (i = 0; i < LIST_SIZE; i++)
1767		BUG_ON(c->n_buffers[i]);
1768
1769	dm_io_client_destroy(c->dm_io);
 
1770	kfree(c);
1771}
1772EXPORT_SYMBOL_GPL(dm_bufio_client_destroy);
1773
1774static unsigned get_max_age_hz(void)
1775{
1776	unsigned max_age = ACCESS_ONCE(dm_bufio_max_age);
1777
1778	if (max_age > UINT_MAX / HZ)
1779		max_age = UINT_MAX / HZ;
1780
1781	return max_age * HZ;
1782}
1783
1784static bool older_than(struct dm_buffer *b, unsigned long age_hz)
1785{
1786	return time_after_eq(jiffies, b->last_accessed + age_hz);
1787}
1788
1789static void __evict_old_buffers(struct dm_bufio_client *c, unsigned long age_hz)
1790{
1791	struct dm_buffer *b, *tmp;
1792	unsigned retain_target = get_retain_buffers(c);
1793	unsigned count;
1794
1795	dm_bufio_lock(c);
1796
1797	count = c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY];
1798	list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_CLEAN], lru_list) {
1799		if (count <= retain_target)
1800			break;
1801
1802		if (!older_than(b, age_hz))
1803			break;
1804
1805		if (__try_evict_buffer(b, 0))
1806			count--;
1807
1808		dm_bufio_cond_resched();
1809	}
1810
1811	dm_bufio_unlock(c);
1812}
1813
1814static void cleanup_old_buffers(void)
1815{
1816	unsigned long max_age_hz = get_max_age_hz();
1817	struct dm_bufio_client *c;
1818
 
 
 
1819	mutex_lock(&dm_bufio_clients_lock);
 
 
 
1820
1821	list_for_each_entry(c, &dm_bufio_all_clients, client_list)
1822		__evict_old_buffers(c, max_age_hz);
 
 
 
 
 
 
1823
 
 
 
1824	mutex_unlock(&dm_bufio_clients_lock);
1825}
1826
1827static struct workqueue_struct *dm_bufio_wq;
1828static struct delayed_work dm_bufio_work;
1829
1830static void work_fn(struct work_struct *w)
1831{
1832	cleanup_old_buffers();
1833
1834	queue_delayed_work(dm_bufio_wq, &dm_bufio_work,
1835			   DM_BUFIO_WORK_TIMER_SECS * HZ);
1836}
1837
1838/*----------------------------------------------------------------
1839 * Module setup
1840 *--------------------------------------------------------------*/
1841
1842/*
1843 * This is called only once for the whole dm_bufio module.
1844 * It initializes memory limit.
1845 */
1846static int __init dm_bufio_init(void)
1847{
1848	__u64 mem;
1849
1850	dm_bufio_allocated_kmem_cache = 0;
1851	dm_bufio_allocated_get_free_pages = 0;
1852	dm_bufio_allocated_vmalloc = 0;
1853	dm_bufio_current_allocated = 0;
1854
1855	memset(&dm_bufio_caches, 0, sizeof dm_bufio_caches);
1856	memset(&dm_bufio_cache_names, 0, sizeof dm_bufio_cache_names);
1857
1858	mem = (__u64)((totalram_pages - totalhigh_pages) *
1859		      DM_BUFIO_MEMORY_PERCENT / 100) << PAGE_SHIFT;
1860
1861	if (mem > ULONG_MAX)
1862		mem = ULONG_MAX;
1863
1864#ifdef CONFIG_MMU
1865	/*
1866	 * Get the size of vmalloc space the same way as VMALLOC_TOTAL
1867	 * in fs/proc/internal.h
1868	 */
1869	if (mem > (VMALLOC_END - VMALLOC_START) * DM_BUFIO_VMALLOC_PERCENT / 100)
1870		mem = (VMALLOC_END - VMALLOC_START) * DM_BUFIO_VMALLOC_PERCENT / 100;
1871#endif
1872
1873	dm_bufio_default_cache_size = mem;
1874
1875	mutex_lock(&dm_bufio_clients_lock);
1876	__cache_size_refresh();
1877	mutex_unlock(&dm_bufio_clients_lock);
1878
1879	dm_bufio_wq = create_singlethread_workqueue("dm_bufio_cache");
1880	if (!dm_bufio_wq)
1881		return -ENOMEM;
1882
1883	INIT_DELAYED_WORK(&dm_bufio_work, work_fn);
1884	queue_delayed_work(dm_bufio_wq, &dm_bufio_work,
1885			   DM_BUFIO_WORK_TIMER_SECS * HZ);
1886
1887	return 0;
1888}
1889
1890/*
1891 * This is called once when unloading the dm_bufio module.
1892 */
1893static void __exit dm_bufio_exit(void)
1894{
1895	int bug = 0;
1896	int i;
1897
1898	cancel_delayed_work_sync(&dm_bufio_work);
1899	destroy_workqueue(dm_bufio_wq);
1900
1901	for (i = 0; i < ARRAY_SIZE(dm_bufio_caches); i++)
1902		kmem_cache_destroy(dm_bufio_caches[i]);
 
 
 
 
1903
1904	for (i = 0; i < ARRAY_SIZE(dm_bufio_cache_names); i++)
1905		kfree(dm_bufio_cache_names[i]);
1906
1907	if (dm_bufio_client_count) {
1908		DMCRIT("%s: dm_bufio_client_count leaked: %d",
1909			__func__, dm_bufio_client_count);
1910		bug = 1;
1911	}
1912
1913	if (dm_bufio_current_allocated) {
1914		DMCRIT("%s: dm_bufio_current_allocated leaked: %lu",
1915			__func__, dm_bufio_current_allocated);
1916		bug = 1;
1917	}
1918
1919	if (dm_bufio_allocated_get_free_pages) {
1920		DMCRIT("%s: dm_bufio_allocated_get_free_pages leaked: %lu",
1921		       __func__, dm_bufio_allocated_get_free_pages);
1922		bug = 1;
1923	}
1924
1925	if (dm_bufio_allocated_vmalloc) {
1926		DMCRIT("%s: dm_bufio_vmalloc leaked: %lu",
1927		       __func__, dm_bufio_allocated_vmalloc);
1928		bug = 1;
1929	}
1930
1931	BUG_ON(bug);
 
1932}
1933
1934module_init(dm_bufio_init)
1935module_exit(dm_bufio_exit)
1936
1937module_param_named(max_cache_size_bytes, dm_bufio_cache_size, ulong, S_IRUGO | S_IWUSR);
1938MODULE_PARM_DESC(max_cache_size_bytes, "Size of metadata cache");
1939
1940module_param_named(max_age_seconds, dm_bufio_max_age, uint, S_IRUGO | S_IWUSR);
1941MODULE_PARM_DESC(max_age_seconds, "Max age of a buffer in seconds");
1942
1943module_param_named(retain_bytes, dm_bufio_retain_bytes, uint, S_IRUGO | S_IWUSR);
1944MODULE_PARM_DESC(retain_bytes, "Try to keep at least this many bytes cached in memory");
1945
1946module_param_named(peak_allocated_bytes, dm_bufio_peak_allocated, ulong, S_IRUGO | S_IWUSR);
1947MODULE_PARM_DESC(peak_allocated_bytes, "Tracks the maximum allocated memory");
1948
1949module_param_named(allocated_kmem_cache_bytes, dm_bufio_allocated_kmem_cache, ulong, S_IRUGO);
1950MODULE_PARM_DESC(allocated_kmem_cache_bytes, "Memory allocated with kmem_cache_alloc");
1951
1952module_param_named(allocated_get_free_pages_bytes, dm_bufio_allocated_get_free_pages, ulong, S_IRUGO);
1953MODULE_PARM_DESC(allocated_get_free_pages_bytes, "Memory allocated with get_free_pages");
1954
1955module_param_named(allocated_vmalloc_bytes, dm_bufio_allocated_vmalloc, ulong, S_IRUGO);
1956MODULE_PARM_DESC(allocated_vmalloc_bytes, "Memory allocated with vmalloc");
1957
1958module_param_named(current_allocated_bytes, dm_bufio_current_allocated, ulong, S_IRUGO);
1959MODULE_PARM_DESC(current_allocated_bytes, "Memory currently used by the cache");
1960
1961MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
1962MODULE_DESCRIPTION(DM_NAME " buffered I/O library");
1963MODULE_LICENSE("GPL");
v3.15
   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 "dm-bufio.h"
  10
  11#include <linux/device-mapper.h>
  12#include <linux/dm-io.h>
  13#include <linux/slab.h>
 
  14#include <linux/vmalloc.h>
  15#include <linux/shrinker.h>
  16#include <linux/module.h>
 
 
  17
  18#define DM_MSG_PREFIX "bufio"
  19
  20/*
  21 * Memory management policy:
  22 *	Limit the number of buffers to DM_BUFIO_MEMORY_PERCENT of main memory
  23 *	or DM_BUFIO_VMALLOC_PERCENT of vmalloc memory (whichever is lower).
  24 *	Always allocate at least DM_BUFIO_MIN_BUFFERS buffers.
  25 *	Start background writeback when there are DM_BUFIO_WRITEBACK_PERCENT
  26 *	dirty buffers.
  27 */
  28#define DM_BUFIO_MIN_BUFFERS		8
  29
  30#define DM_BUFIO_MEMORY_PERCENT		2
  31#define DM_BUFIO_VMALLOC_PERCENT	25
  32#define DM_BUFIO_WRITEBACK_PERCENT	75
  33
  34/*
  35 * Check buffer ages in this interval (seconds)
  36 */
  37#define DM_BUFIO_WORK_TIMER_SECS	10
  38
  39/*
  40 * Free buffers when they are older than this (seconds)
  41 */
  42#define DM_BUFIO_DEFAULT_AGE_SECS	60
 
 
 
 
 
  43
  44/*
  45 * The number of bvec entries that are embedded directly in the buffer.
  46 * If the chunk size is larger, dm-io is used to do the io.
  47 */
  48#define DM_BUFIO_INLINE_VECS		16
  49
  50/*
  51 * Buffer hash
  52 */
  53#define DM_BUFIO_HASH_BITS	20
  54#define DM_BUFIO_HASH(block) \
  55	((((block) >> DM_BUFIO_HASH_BITS) ^ (block)) & \
  56	 ((1 << DM_BUFIO_HASH_BITS) - 1))
  57
  58/*
  59 * Don't try to use kmem_cache_alloc for blocks larger than this.
  60 * For explanation, see alloc_buffer_data below.
  61 */
  62#define DM_BUFIO_BLOCK_SIZE_SLAB_LIMIT	(PAGE_SIZE >> 1)
  63#define DM_BUFIO_BLOCK_SIZE_GFP_LIMIT	(PAGE_SIZE << (MAX_ORDER - 1))
  64
  65/*
  66 * dm_buffer->list_mode
  67 */
  68#define LIST_CLEAN	0
  69#define LIST_DIRTY	1
  70#define LIST_SIZE	2
  71
  72/*
  73 * Linking of buffers:
  74 *	All buffers are linked to cache_hash with their hash_list field.
  75 *
  76 *	Clean buffers that are not being written (B_WRITING not set)
  77 *	are linked to lru[LIST_CLEAN] with their lru_list field.
  78 *
  79 *	Dirty and clean buffers that are being written are linked to
  80 *	lru[LIST_DIRTY] with their lru_list field. When the write
  81 *	finishes, the buffer cannot be relinked immediately (because we
  82 *	are in an interrupt context and relinking requires process
  83 *	context), so some clean-not-writing buffers can be held on
  84 *	dirty_lru too.  They are later added to lru in the process
  85 *	context.
  86 */
  87struct dm_bufio_client {
  88	struct mutex lock;
  89
  90	struct list_head lru[LIST_SIZE];
  91	unsigned long n_buffers[LIST_SIZE];
  92
  93	struct block_device *bdev;
  94	unsigned block_size;
  95	unsigned char sectors_per_block_bits;
  96	unsigned char pages_per_block_bits;
  97	unsigned char blocks_per_page_bits;
  98	unsigned aux_size;
  99	void (*alloc_callback)(struct dm_buffer *);
 100	void (*write_callback)(struct dm_buffer *);
 101
 102	struct dm_io_client *dm_io;
 103
 104	struct list_head reserved_buffers;
 105	unsigned need_reserved_buffers;
 106
 107	unsigned minimum_buffers;
 108
 109	struct hlist_head *cache_hash;
 110	wait_queue_head_t free_buffer_wait;
 111
 112	int async_write_error;
 113
 114	struct list_head client_list;
 115	struct shrinker shrinker;
 116};
 117
 118/*
 119 * Buffer state bits.
 120 */
 121#define B_READING	0
 122#define B_WRITING	1
 123#define B_DIRTY		2
 124
 125/*
 126 * Describes how the block was allocated:
 127 * kmem_cache_alloc(), __get_free_pages() or vmalloc().
 128 * See the comment at alloc_buffer_data.
 129 */
 130enum data_mode {
 131	DATA_MODE_SLAB = 0,
 132	DATA_MODE_GET_FREE_PAGES = 1,
 133	DATA_MODE_VMALLOC = 2,
 134	DATA_MODE_LIMIT = 3
 135};
 136
 137struct dm_buffer {
 138	struct hlist_node hash_list;
 139	struct list_head lru_list;
 140	sector_t block;
 141	void *data;
 142	enum data_mode data_mode;
 143	unsigned char list_mode;		/* LIST_* */
 144	unsigned hold_count;
 145	int read_error;
 146	int write_error;
 147	unsigned long state;
 148	unsigned long last_accessed;
 149	struct dm_bufio_client *c;
 150	struct list_head write_list;
 151	struct bio bio;
 152	struct bio_vec bio_vec[DM_BUFIO_INLINE_VECS];
 
 
 
 
 
 153};
 154
 155/*----------------------------------------------------------------*/
 156
 157static struct kmem_cache *dm_bufio_caches[PAGE_SHIFT - SECTOR_SHIFT];
 158static char *dm_bufio_cache_names[PAGE_SHIFT - SECTOR_SHIFT];
 159
 160static inline int dm_bufio_cache_index(struct dm_bufio_client *c)
 161{
 162	unsigned ret = c->blocks_per_page_bits - 1;
 163
 164	BUG_ON(ret >= ARRAY_SIZE(dm_bufio_caches));
 165
 166	return ret;
 167}
 168
 169#define DM_BUFIO_CACHE(c)	(dm_bufio_caches[dm_bufio_cache_index(c)])
 170#define DM_BUFIO_CACHE_NAME(c)	(dm_bufio_cache_names[dm_bufio_cache_index(c)])
 171
 172#define dm_bufio_in_request()	(!!current->bio_list)
 173
 174static void dm_bufio_lock(struct dm_bufio_client *c)
 175{
 176	mutex_lock_nested(&c->lock, dm_bufio_in_request());
 177}
 178
 179static int dm_bufio_trylock(struct dm_bufio_client *c)
 180{
 181	return mutex_trylock(&c->lock);
 182}
 183
 184static void dm_bufio_unlock(struct dm_bufio_client *c)
 185{
 186	mutex_unlock(&c->lock);
 187}
 188
 189/*
 190 * FIXME Move to sched.h?
 191 */
 192#ifdef CONFIG_PREEMPT_VOLUNTARY
 193#  define dm_bufio_cond_resched()		\
 194do {						\
 195	if (unlikely(need_resched()))		\
 196		_cond_resched();		\
 197} while (0)
 198#else
 199#  define dm_bufio_cond_resched()                do { } while (0)
 200#endif
 201
 202/*----------------------------------------------------------------*/
 203
 204/*
 205 * Default cache size: available memory divided by the ratio.
 206 */
 207static unsigned long dm_bufio_default_cache_size;
 208
 209/*
 210 * Total cache size set by the user.
 211 */
 212static unsigned long dm_bufio_cache_size;
 213
 214/*
 215 * A copy of dm_bufio_cache_size because dm_bufio_cache_size can change
 216 * at any time.  If it disagrees, the user has changed cache size.
 217 */
 218static unsigned long dm_bufio_cache_size_latch;
 219
 220static DEFINE_SPINLOCK(param_spinlock);
 221
 222/*
 223 * Buffers are freed after this timeout
 224 */
 225static unsigned dm_bufio_max_age = DM_BUFIO_DEFAULT_AGE_SECS;
 
 226
 227static unsigned long dm_bufio_peak_allocated;
 228static unsigned long dm_bufio_allocated_kmem_cache;
 229static unsigned long dm_bufio_allocated_get_free_pages;
 230static unsigned long dm_bufio_allocated_vmalloc;
 231static unsigned long dm_bufio_current_allocated;
 232
 233/*----------------------------------------------------------------*/
 234
 235/*
 236 * Per-client cache: dm_bufio_cache_size / dm_bufio_client_count
 237 */
 238static unsigned long dm_bufio_cache_size_per_client;
 239
 240/*
 241 * The current number of clients.
 242 */
 243static int dm_bufio_client_count;
 244
 245/*
 246 * The list of all clients.
 247 */
 248static LIST_HEAD(dm_bufio_all_clients);
 249
 250/*
 251 * This mutex protects dm_bufio_cache_size_latch,
 252 * dm_bufio_cache_size_per_client and dm_bufio_client_count
 253 */
 254static DEFINE_MUTEX(dm_bufio_clients_lock);
 255
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 256/*----------------------------------------------------------------*/
 257
 258static void adjust_total_allocated(enum data_mode data_mode, long diff)
 259{
 260	static unsigned long * const class_ptr[DATA_MODE_LIMIT] = {
 261		&dm_bufio_allocated_kmem_cache,
 262		&dm_bufio_allocated_get_free_pages,
 263		&dm_bufio_allocated_vmalloc,
 264	};
 265
 266	spin_lock(&param_spinlock);
 267
 268	*class_ptr[data_mode] += diff;
 269
 270	dm_bufio_current_allocated += diff;
 271
 272	if (dm_bufio_current_allocated > dm_bufio_peak_allocated)
 273		dm_bufio_peak_allocated = dm_bufio_current_allocated;
 274
 275	spin_unlock(&param_spinlock);
 276}
 277
 278/*
 279 * Change the number of clients and recalculate per-client limit.
 280 */
 281static void __cache_size_refresh(void)
 282{
 283	BUG_ON(!mutex_is_locked(&dm_bufio_clients_lock));
 284	BUG_ON(dm_bufio_client_count < 0);
 285
 286	dm_bufio_cache_size_latch = ACCESS_ONCE(dm_bufio_cache_size);
 287
 288	/*
 289	 * Use default if set to 0 and report the actual cache size used.
 290	 */
 291	if (!dm_bufio_cache_size_latch) {
 292		(void)cmpxchg(&dm_bufio_cache_size, 0,
 293			      dm_bufio_default_cache_size);
 294		dm_bufio_cache_size_latch = dm_bufio_default_cache_size;
 295	}
 296
 297	dm_bufio_cache_size_per_client = dm_bufio_cache_size_latch /
 298					 (dm_bufio_client_count ? : 1);
 299}
 300
 301/*
 302 * Allocating buffer data.
 303 *
 304 * Small buffers are allocated with kmem_cache, to use space optimally.
 305 *
 306 * For large buffers, we choose between get_free_pages and vmalloc.
 307 * Each has advantages and disadvantages.
 308 *
 309 * __get_free_pages can randomly fail if the memory is fragmented.
 310 * __vmalloc won't randomly fail, but vmalloc space is limited (it may be
 311 * as low as 128M) so using it for caching is not appropriate.
 312 *
 313 * If the allocation may fail we use __get_free_pages. Memory fragmentation
 314 * won't have a fatal effect here, but it just causes flushes of some other
 315 * buffers and more I/O will be performed. Don't use __get_free_pages if it
 316 * always fails (i.e. order >= MAX_ORDER).
 317 *
 318 * If the allocation shouldn't fail we use __vmalloc. This is only for the
 319 * initial reserve allocation, so there's no risk of wasting all vmalloc
 320 * space.
 321 */
 322static void *alloc_buffer_data(struct dm_bufio_client *c, gfp_t gfp_mask,
 323			       enum data_mode *data_mode)
 324{
 325	unsigned noio_flag;
 326	void *ptr;
 327
 328	if (c->block_size <= DM_BUFIO_BLOCK_SIZE_SLAB_LIMIT) {
 329		*data_mode = DATA_MODE_SLAB;
 330		return kmem_cache_alloc(DM_BUFIO_CACHE(c), gfp_mask);
 331	}
 332
 333	if (c->block_size <= DM_BUFIO_BLOCK_SIZE_GFP_LIMIT &&
 334	    gfp_mask & __GFP_NORETRY) {
 335		*data_mode = DATA_MODE_GET_FREE_PAGES;
 336		return (void *)__get_free_pages(gfp_mask,
 337						c->pages_per_block_bits);
 338	}
 339
 340	*data_mode = DATA_MODE_VMALLOC;
 341
 342	/*
 343	 * __vmalloc allocates the data pages and auxiliary structures with
 344	 * gfp_flags that were specified, but pagetables are always allocated
 345	 * with GFP_KERNEL, no matter what was specified as gfp_mask.
 346	 *
 347	 * Consequently, we must set per-process flag PF_MEMALLOC_NOIO so that
 348	 * all allocations done by this process (including pagetables) are done
 349	 * as if GFP_NOIO was specified.
 350	 */
 351
 352	if (gfp_mask & __GFP_NORETRY)
 353		noio_flag = memalloc_noio_save();
 354
 355	ptr = __vmalloc(c->block_size, gfp_mask | __GFP_HIGHMEM, PAGE_KERNEL);
 356
 357	if (gfp_mask & __GFP_NORETRY)
 358		memalloc_noio_restore(noio_flag);
 359
 360	return ptr;
 361}
 362
 363/*
 364 * Free buffer's data.
 365 */
 366static void free_buffer_data(struct dm_bufio_client *c,
 367			     void *data, enum data_mode data_mode)
 368{
 369	switch (data_mode) {
 370	case DATA_MODE_SLAB:
 371		kmem_cache_free(DM_BUFIO_CACHE(c), data);
 372		break;
 373
 374	case DATA_MODE_GET_FREE_PAGES:
 375		free_pages((unsigned long)data, c->pages_per_block_bits);
 376		break;
 377
 378	case DATA_MODE_VMALLOC:
 379		vfree(data);
 380		break;
 381
 382	default:
 383		DMCRIT("dm_bufio_free_buffer_data: bad data mode: %d",
 384		       data_mode);
 385		BUG();
 386	}
 387}
 388
 389/*
 390 * Allocate buffer and its data.
 391 */
 392static struct dm_buffer *alloc_buffer(struct dm_bufio_client *c, gfp_t gfp_mask)
 393{
 394	struct dm_buffer *b = kmalloc(sizeof(struct dm_buffer) + c->aux_size,
 395				      gfp_mask);
 396
 397	if (!b)
 398		return NULL;
 399
 400	b->c = c;
 401
 402	b->data = alloc_buffer_data(c, gfp_mask, &b->data_mode);
 403	if (!b->data) {
 404		kfree(b);
 405		return NULL;
 406	}
 407
 408	adjust_total_allocated(b->data_mode, (long)c->block_size);
 409
 
 
 
 410	return b;
 411}
 412
 413/*
 414 * Free buffer and its data.
 415 */
 416static void free_buffer(struct dm_buffer *b)
 417{
 418	struct dm_bufio_client *c = b->c;
 419
 420	adjust_total_allocated(b->data_mode, -(long)c->block_size);
 421
 422	free_buffer_data(c, b->data, b->data_mode);
 423	kfree(b);
 424}
 425
 426/*
 427 * Link buffer to the hash list and clean or dirty queue.
 428 */
 429static void __link_buffer(struct dm_buffer *b, sector_t block, int dirty)
 430{
 431	struct dm_bufio_client *c = b->c;
 432
 433	c->n_buffers[dirty]++;
 434	b->block = block;
 435	b->list_mode = dirty;
 436	list_add(&b->lru_list, &c->lru[dirty]);
 437	hlist_add_head(&b->hash_list, &c->cache_hash[DM_BUFIO_HASH(block)]);
 438	b->last_accessed = jiffies;
 439}
 440
 441/*
 442 * Unlink buffer from the hash list and dirty or clean queue.
 443 */
 444static void __unlink_buffer(struct dm_buffer *b)
 445{
 446	struct dm_bufio_client *c = b->c;
 447
 448	BUG_ON(!c->n_buffers[b->list_mode]);
 449
 450	c->n_buffers[b->list_mode]--;
 451	hlist_del(&b->hash_list);
 452	list_del(&b->lru_list);
 453}
 454
 455/*
 456 * Place the buffer to the head of dirty or clean LRU queue.
 457 */
 458static void __relink_lru(struct dm_buffer *b, int dirty)
 459{
 460	struct dm_bufio_client *c = b->c;
 461
 462	BUG_ON(!c->n_buffers[b->list_mode]);
 463
 464	c->n_buffers[b->list_mode]--;
 465	c->n_buffers[dirty]++;
 466	b->list_mode = dirty;
 467	list_move(&b->lru_list, &c->lru[dirty]);
 
 468}
 469
 470/*----------------------------------------------------------------
 471 * Submit I/O on the buffer.
 472 *
 473 * Bio interface is faster but it has some problems:
 474 *	the vector list is limited (increasing this limit increases
 475 *	memory-consumption per buffer, so it is not viable);
 476 *
 477 *	the memory must be direct-mapped, not vmalloced;
 478 *
 479 *	the I/O driver can reject requests spuriously if it thinks that
 480 *	the requests are too big for the device or if they cross a
 481 *	controller-defined memory boundary.
 482 *
 483 * If the buffer is small enough (up to DM_BUFIO_INLINE_VECS pages) and
 484 * it is not vmalloced, try using the bio interface.
 485 *
 486 * If the buffer is big, if it is vmalloced or if the underlying device
 487 * rejects the bio because it is too large, use dm-io layer to do the I/O.
 488 * The dm-io layer splits the I/O into multiple requests, avoiding the above
 489 * shortcomings.
 490 *--------------------------------------------------------------*/
 491
 492/*
 493 * dm-io completion routine. It just calls b->bio.bi_end_io, pretending
 494 * that the request was handled directly with bio interface.
 495 */
 496static void dmio_complete(unsigned long error, void *context)
 497{
 498	struct dm_buffer *b = context;
 499
 500	b->bio.bi_end_io(&b->bio, error ? -EIO : 0);
 
 501}
 502
 503static void use_dmio(struct dm_buffer *b, int rw, sector_t block,
 504		     bio_end_io_t *end_io)
 505{
 506	int r;
 507	struct dm_io_request io_req = {
 508		.bi_rw = rw,
 509		.notify.fn = dmio_complete,
 510		.notify.context = b,
 511		.client = b->c->dm_io,
 512	};
 513	struct dm_io_region region = {
 514		.bdev = b->c->bdev,
 515		.sector = block << b->c->sectors_per_block_bits,
 516		.count = b->c->block_size >> SECTOR_SHIFT,
 517	};
 518
 519	if (b->data_mode != DATA_MODE_VMALLOC) {
 520		io_req.mem.type = DM_IO_KMEM;
 521		io_req.mem.ptr.addr = b->data;
 522	} else {
 523		io_req.mem.type = DM_IO_VMA;
 524		io_req.mem.ptr.vma = b->data;
 525	}
 526
 527	b->bio.bi_end_io = end_io;
 528
 529	r = dm_io(&io_req, 1, &region, NULL);
 530	if (r)
 531		end_io(&b->bio, r);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 532}
 533
 534static void use_inline_bio(struct dm_buffer *b, int rw, sector_t block,
 535			   bio_end_io_t *end_io)
 536{
 537	char *ptr;
 538	int len;
 539
 540	bio_init(&b->bio);
 541	b->bio.bi_io_vec = b->bio_vec;
 542	b->bio.bi_max_vecs = DM_BUFIO_INLINE_VECS;
 543	b->bio.bi_iter.bi_sector = block << b->c->sectors_per_block_bits;
 544	b->bio.bi_bdev = b->c->bdev;
 545	b->bio.bi_end_io = end_io;
 
 
 
 
 
 546
 547	/*
 548	 * We assume that if len >= PAGE_SIZE ptr is page-aligned.
 549	 * If len < PAGE_SIZE the buffer doesn't cross page boundary.
 550	 */
 551	ptr = b->data;
 552	len = b->c->block_size;
 553
 554	if (len >= PAGE_SIZE)
 555		BUG_ON((unsigned long)ptr & (PAGE_SIZE - 1));
 556	else
 557		BUG_ON((unsigned long)ptr & (len - 1));
 558
 559	do {
 560		if (!bio_add_page(&b->bio, virt_to_page(ptr),
 561				  len < PAGE_SIZE ? len : PAGE_SIZE,
 562				  virt_to_phys(ptr) & (PAGE_SIZE - 1))) {
 563			BUG_ON(b->c->block_size <= PAGE_SIZE);
 564			use_dmio(b, rw, block, end_io);
 565			return;
 566		}
 567
 568		len -= PAGE_SIZE;
 569		ptr += PAGE_SIZE;
 570	} while (len > 0);
 571
 572	submit_bio(rw, &b->bio);
 573}
 574
 575static void submit_io(struct dm_buffer *b, int rw, sector_t block,
 576		      bio_end_io_t *end_io)
 577{
 578	if (rw == WRITE && b->c->write_callback)
 579		b->c->write_callback(b);
 580
 581	if (b->c->block_size <= DM_BUFIO_INLINE_VECS * PAGE_SIZE &&
 582	    b->data_mode != DATA_MODE_VMALLOC)
 583		use_inline_bio(b, rw, block, end_io);
 584	else
 585		use_dmio(b, rw, block, end_io);
 586}
 587
 588/*----------------------------------------------------------------
 589 * Writing dirty buffers
 590 *--------------------------------------------------------------*/
 591
 592/*
 593 * The endio routine for write.
 594 *
 595 * Set the error, clear B_WRITING bit and wake anyone who was waiting on
 596 * it.
 597 */
 598static void write_endio(struct bio *bio, int error)
 599{
 600	struct dm_buffer *b = container_of(bio, struct dm_buffer, bio);
 601
 602	b->write_error = error;
 603	if (unlikely(error)) {
 604		struct dm_bufio_client *c = b->c;
 
 605		(void)cmpxchg(&c->async_write_error, 0, error);
 606	}
 607
 608	BUG_ON(!test_bit(B_WRITING, &b->state));
 609
 610	smp_mb__before_clear_bit();
 611	clear_bit(B_WRITING, &b->state);
 612	smp_mb__after_clear_bit();
 613
 614	wake_up_bit(&b->state, B_WRITING);
 615}
 616
 617/*
 618 * This function is called when wait_on_bit is actually waiting.
 619 */
 620static int do_io_schedule(void *word)
 621{
 622	io_schedule();
 623
 624	return 0;
 625}
 626
 627/*
 628 * Initiate a write on a dirty buffer, but don't wait for it.
 629 *
 630 * - If the buffer is not dirty, exit.
 631 * - If there some previous write going on, wait for it to finish (we can't
 632 *   have two writes on the same buffer simultaneously).
 633 * - Submit our write and don't wait on it. We set B_WRITING indicating
 634 *   that there is a write in progress.
 635 */
 636static void __write_dirty_buffer(struct dm_buffer *b,
 637				 struct list_head *write_list)
 638{
 639	if (!test_bit(B_DIRTY, &b->state))
 640		return;
 641
 642	clear_bit(B_DIRTY, &b->state);
 643	wait_on_bit_lock(&b->state, B_WRITING,
 644			 do_io_schedule, TASK_UNINTERRUPTIBLE);
 645
 646	if (!write_list)
 647		submit_io(b, WRITE, b->block, write_endio);
 648	else
 649		list_add_tail(&b->write_list, write_list);
 650}
 651
 652static void __flush_write_list(struct list_head *write_list)
 653{
 654	struct blk_plug plug;
 655	blk_start_plug(&plug);
 656	while (!list_empty(write_list)) {
 657		struct dm_buffer *b =
 658			list_entry(write_list->next, struct dm_buffer, write_list);
 659		list_del(&b->write_list);
 660		submit_io(b, WRITE, b->block, write_endio);
 661		dm_bufio_cond_resched();
 662	}
 663	blk_finish_plug(&plug);
 664}
 665
 666/*
 667 * Wait until any activity on the buffer finishes.  Possibly write the
 668 * buffer if it is dirty.  When this function finishes, there is no I/O
 669 * running on the buffer and the buffer is not dirty.
 670 */
 671static void __make_buffer_clean(struct dm_buffer *b)
 672{
 673	BUG_ON(b->hold_count);
 674
 675	if (!b->state)	/* fast case */
 676		return;
 677
 678	wait_on_bit(&b->state, B_READING, do_io_schedule, TASK_UNINTERRUPTIBLE);
 679	__write_dirty_buffer(b, NULL);
 680	wait_on_bit(&b->state, B_WRITING, do_io_schedule, TASK_UNINTERRUPTIBLE);
 681}
 682
 683/*
 684 * Find some buffer that is not held by anybody, clean it, unlink it and
 685 * return it.
 686 */
 687static struct dm_buffer *__get_unclaimed_buffer(struct dm_bufio_client *c)
 688{
 689	struct dm_buffer *b;
 690
 691	list_for_each_entry_reverse(b, &c->lru[LIST_CLEAN], lru_list) {
 692		BUG_ON(test_bit(B_WRITING, &b->state));
 693		BUG_ON(test_bit(B_DIRTY, &b->state));
 694
 695		if (!b->hold_count) {
 696			__make_buffer_clean(b);
 697			__unlink_buffer(b);
 698			return b;
 699		}
 700		dm_bufio_cond_resched();
 701	}
 702
 703	list_for_each_entry_reverse(b, &c->lru[LIST_DIRTY], lru_list) {
 704		BUG_ON(test_bit(B_READING, &b->state));
 705
 706		if (!b->hold_count) {
 707			__make_buffer_clean(b);
 708			__unlink_buffer(b);
 709			return b;
 710		}
 711		dm_bufio_cond_resched();
 712	}
 713
 714	return NULL;
 715}
 716
 717/*
 718 * Wait until some other threads free some buffer or release hold count on
 719 * some buffer.
 720 *
 721 * This function is entered with c->lock held, drops it and regains it
 722 * before exiting.
 723 */
 724static void __wait_for_free_buffer(struct dm_bufio_client *c)
 725{
 726	DECLARE_WAITQUEUE(wait, current);
 727
 728	add_wait_queue(&c->free_buffer_wait, &wait);
 729	set_task_state(current, TASK_UNINTERRUPTIBLE);
 730	dm_bufio_unlock(c);
 731
 732	io_schedule();
 733
 734	set_task_state(current, TASK_RUNNING);
 735	remove_wait_queue(&c->free_buffer_wait, &wait);
 736
 737	dm_bufio_lock(c);
 738}
 739
 740enum new_flag {
 741	NF_FRESH = 0,
 742	NF_READ = 1,
 743	NF_GET = 2,
 744	NF_PREFETCH = 3
 745};
 746
 747/*
 748 * Allocate a new buffer. If the allocation is not possible, wait until
 749 * some other thread frees a buffer.
 750 *
 751 * May drop the lock and regain it.
 752 */
 753static struct dm_buffer *__alloc_buffer_wait_no_callback(struct dm_bufio_client *c, enum new_flag nf)
 754{
 755	struct dm_buffer *b;
 756
 757	/*
 758	 * dm-bufio is resistant to allocation failures (it just keeps
 759	 * one buffer reserved in cases all the allocations fail).
 760	 * So set flags to not try too hard:
 761	 *	GFP_NOIO: don't recurse into the I/O layer
 762	 *	__GFP_NORETRY: don't retry and rather return failure
 763	 *	__GFP_NOMEMALLOC: don't use emergency reserves
 764	 *	__GFP_NOWARN: don't print a warning in case of failure
 765	 *
 766	 * For debugging, if we set the cache size to 1, no new buffers will
 767	 * be allocated.
 768	 */
 769	while (1) {
 770		if (dm_bufio_cache_size_latch != 1) {
 771			b = alloc_buffer(c, GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
 772			if (b)
 773				return b;
 774		}
 775
 776		if (nf == NF_PREFETCH)
 777			return NULL;
 778
 779		if (!list_empty(&c->reserved_buffers)) {
 780			b = list_entry(c->reserved_buffers.next,
 781				       struct dm_buffer, lru_list);
 782			list_del(&b->lru_list);
 783			c->need_reserved_buffers++;
 784
 785			return b;
 786		}
 787
 788		b = __get_unclaimed_buffer(c);
 789		if (b)
 790			return b;
 791
 792		__wait_for_free_buffer(c);
 793	}
 794}
 795
 796static struct dm_buffer *__alloc_buffer_wait(struct dm_bufio_client *c, enum new_flag nf)
 797{
 798	struct dm_buffer *b = __alloc_buffer_wait_no_callback(c, nf);
 799
 800	if (!b)
 801		return NULL;
 802
 803	if (c->alloc_callback)
 804		c->alloc_callback(b);
 805
 806	return b;
 807}
 808
 809/*
 810 * Free a buffer and wake other threads waiting for free buffers.
 811 */
 812static void __free_buffer_wake(struct dm_buffer *b)
 813{
 814	struct dm_bufio_client *c = b->c;
 815
 816	if (!c->need_reserved_buffers)
 817		free_buffer(b);
 818	else {
 819		list_add(&b->lru_list, &c->reserved_buffers);
 820		c->need_reserved_buffers--;
 821	}
 822
 823	wake_up(&c->free_buffer_wait);
 824}
 825
 826static void __write_dirty_buffers_async(struct dm_bufio_client *c, int no_wait,
 827					struct list_head *write_list)
 828{
 829	struct dm_buffer *b, *tmp;
 830
 831	list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
 832		BUG_ON(test_bit(B_READING, &b->state));
 833
 834		if (!test_bit(B_DIRTY, &b->state) &&
 835		    !test_bit(B_WRITING, &b->state)) {
 836			__relink_lru(b, LIST_CLEAN);
 837			continue;
 838		}
 839
 840		if (no_wait && test_bit(B_WRITING, &b->state))
 841			return;
 842
 843		__write_dirty_buffer(b, write_list);
 844		dm_bufio_cond_resched();
 845	}
 846}
 847
 848/*
 849 * Get writeback threshold and buffer limit for a given client.
 850 */
 851static void __get_memory_limit(struct dm_bufio_client *c,
 852			       unsigned long *threshold_buffers,
 853			       unsigned long *limit_buffers)
 854{
 855	unsigned long buffers;
 856
 857	if (ACCESS_ONCE(dm_bufio_cache_size) != dm_bufio_cache_size_latch) {
 858		mutex_lock(&dm_bufio_clients_lock);
 859		__cache_size_refresh();
 860		mutex_unlock(&dm_bufio_clients_lock);
 861	}
 862
 863	buffers = dm_bufio_cache_size_per_client >>
 864		  (c->sectors_per_block_bits + SECTOR_SHIFT);
 865
 866	if (buffers < c->minimum_buffers)
 867		buffers = c->minimum_buffers;
 868
 869	*limit_buffers = buffers;
 870	*threshold_buffers = buffers * DM_BUFIO_WRITEBACK_PERCENT / 100;
 871}
 872
 873/*
 874 * Check if we're over watermark.
 875 * If we are over threshold_buffers, start freeing buffers.
 876 * If we're over "limit_buffers", block until we get under the limit.
 877 */
 878static void __check_watermark(struct dm_bufio_client *c,
 879			      struct list_head *write_list)
 880{
 881	unsigned long threshold_buffers, limit_buffers;
 882
 883	__get_memory_limit(c, &threshold_buffers, &limit_buffers);
 884
 885	while (c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY] >
 886	       limit_buffers) {
 887
 888		struct dm_buffer *b = __get_unclaimed_buffer(c);
 889
 890		if (!b)
 891			return;
 892
 893		__free_buffer_wake(b);
 894		dm_bufio_cond_resched();
 895	}
 896
 897	if (c->n_buffers[LIST_DIRTY] > threshold_buffers)
 898		__write_dirty_buffers_async(c, 1, write_list);
 899}
 900
 901/*
 902 * Find a buffer in the hash.
 903 */
 904static struct dm_buffer *__find(struct dm_bufio_client *c, sector_t block)
 905{
 906	struct dm_buffer *b;
 907
 908	hlist_for_each_entry(b, &c->cache_hash[DM_BUFIO_HASH(block)],
 909			     hash_list) {
 910		dm_bufio_cond_resched();
 911		if (b->block == block)
 912			return b;
 913	}
 914
 915	return NULL;
 916}
 917
 918/*----------------------------------------------------------------
 919 * Getting a buffer
 920 *--------------------------------------------------------------*/
 921
 922static struct dm_buffer *__bufio_new(struct dm_bufio_client *c, sector_t block,
 923				     enum new_flag nf, int *need_submit,
 924				     struct list_head *write_list)
 925{
 926	struct dm_buffer *b, *new_b = NULL;
 927
 928	*need_submit = 0;
 929
 930	b = __find(c, block);
 931	if (b)
 932		goto found_buffer;
 933
 934	if (nf == NF_GET)
 935		return NULL;
 936
 937	new_b = __alloc_buffer_wait(c, nf);
 938	if (!new_b)
 939		return NULL;
 940
 941	/*
 942	 * We've had a period where the mutex was unlocked, so need to
 943	 * recheck the hash table.
 944	 */
 945	b = __find(c, block);
 946	if (b) {
 947		__free_buffer_wake(new_b);
 948		goto found_buffer;
 949	}
 950
 951	__check_watermark(c, write_list);
 952
 953	b = new_b;
 954	b->hold_count = 1;
 955	b->read_error = 0;
 956	b->write_error = 0;
 957	__link_buffer(b, block, LIST_CLEAN);
 958
 959	if (nf == NF_FRESH) {
 960		b->state = 0;
 961		return b;
 962	}
 963
 964	b->state = 1 << B_READING;
 965	*need_submit = 1;
 966
 967	return b;
 968
 969found_buffer:
 970	if (nf == NF_PREFETCH)
 971		return NULL;
 972	/*
 973	 * Note: it is essential that we don't wait for the buffer to be
 974	 * read if dm_bufio_get function is used. Both dm_bufio_get and
 975	 * dm_bufio_prefetch can be used in the driver request routine.
 976	 * If the user called both dm_bufio_prefetch and dm_bufio_get on
 977	 * the same buffer, it would deadlock if we waited.
 978	 */
 979	if (nf == NF_GET && unlikely(test_bit(B_READING, &b->state)))
 980		return NULL;
 981
 982	b->hold_count++;
 983	__relink_lru(b, test_bit(B_DIRTY, &b->state) ||
 984		     test_bit(B_WRITING, &b->state));
 985	return b;
 986}
 987
 988/*
 989 * The endio routine for reading: set the error, clear the bit and wake up
 990 * anyone waiting on the buffer.
 991 */
 992static void read_endio(struct bio *bio, int error)
 993{
 994	struct dm_buffer *b = container_of(bio, struct dm_buffer, bio);
 995
 996	b->read_error = error;
 997
 998	BUG_ON(!test_bit(B_READING, &b->state));
 999
1000	smp_mb__before_clear_bit();
1001	clear_bit(B_READING, &b->state);
1002	smp_mb__after_clear_bit();
1003
1004	wake_up_bit(&b->state, B_READING);
1005}
1006
1007/*
1008 * A common routine for dm_bufio_new and dm_bufio_read.  Operation of these
1009 * functions is similar except that dm_bufio_new doesn't read the
1010 * buffer from the disk (assuming that the caller overwrites all the data
1011 * and uses dm_bufio_mark_buffer_dirty to write new data back).
1012 */
1013static void *new_read(struct dm_bufio_client *c, sector_t block,
1014		      enum new_flag nf, struct dm_buffer **bp)
1015{
1016	int need_submit;
1017	struct dm_buffer *b;
1018
1019	LIST_HEAD(write_list);
1020
1021	dm_bufio_lock(c);
1022	b = __bufio_new(c, block, nf, &need_submit, &write_list);
 
 
 
 
1023	dm_bufio_unlock(c);
1024
1025	__flush_write_list(&write_list);
1026
1027	if (!b)
1028		return b;
1029
1030	if (need_submit)
1031		submit_io(b, READ, b->block, read_endio);
1032
1033	wait_on_bit(&b->state, B_READING, do_io_schedule, TASK_UNINTERRUPTIBLE);
1034
1035	if (b->read_error) {
1036		int error = b->read_error;
1037
1038		dm_bufio_release(b);
1039
1040		return ERR_PTR(error);
1041	}
1042
1043	*bp = b;
1044
1045	return b->data;
1046}
1047
1048void *dm_bufio_get(struct dm_bufio_client *c, sector_t block,
1049		   struct dm_buffer **bp)
1050{
1051	return new_read(c, block, NF_GET, bp);
1052}
1053EXPORT_SYMBOL_GPL(dm_bufio_get);
1054
1055void *dm_bufio_read(struct dm_bufio_client *c, sector_t block,
1056		    struct dm_buffer **bp)
1057{
1058	BUG_ON(dm_bufio_in_request());
1059
1060	return new_read(c, block, NF_READ, bp);
1061}
1062EXPORT_SYMBOL_GPL(dm_bufio_read);
1063
1064void *dm_bufio_new(struct dm_bufio_client *c, sector_t block,
1065		   struct dm_buffer **bp)
1066{
1067	BUG_ON(dm_bufio_in_request());
1068
1069	return new_read(c, block, NF_FRESH, bp);
1070}
1071EXPORT_SYMBOL_GPL(dm_bufio_new);
1072
1073void dm_bufio_prefetch(struct dm_bufio_client *c,
1074		       sector_t block, unsigned n_blocks)
1075{
1076	struct blk_plug plug;
1077
1078	LIST_HEAD(write_list);
1079
1080	BUG_ON(dm_bufio_in_request());
1081
1082	blk_start_plug(&plug);
1083	dm_bufio_lock(c);
1084
1085	for (; n_blocks--; block++) {
1086		int need_submit;
1087		struct dm_buffer *b;
1088		b = __bufio_new(c, block, NF_PREFETCH, &need_submit,
1089				&write_list);
1090		if (unlikely(!list_empty(&write_list))) {
1091			dm_bufio_unlock(c);
1092			blk_finish_plug(&plug);
1093			__flush_write_list(&write_list);
1094			blk_start_plug(&plug);
1095			dm_bufio_lock(c);
1096		}
1097		if (unlikely(b != NULL)) {
1098			dm_bufio_unlock(c);
1099
1100			if (need_submit)
1101				submit_io(b, READ, b->block, read_endio);
1102			dm_bufio_release(b);
1103
1104			dm_bufio_cond_resched();
1105
1106			if (!n_blocks)
1107				goto flush_plug;
1108			dm_bufio_lock(c);
1109		}
1110	}
1111
1112	dm_bufio_unlock(c);
1113
1114flush_plug:
1115	blk_finish_plug(&plug);
1116}
1117EXPORT_SYMBOL_GPL(dm_bufio_prefetch);
1118
1119void dm_bufio_release(struct dm_buffer *b)
1120{
1121	struct dm_bufio_client *c = b->c;
1122
1123	dm_bufio_lock(c);
1124
1125	BUG_ON(!b->hold_count);
1126
1127	b->hold_count--;
1128	if (!b->hold_count) {
1129		wake_up(&c->free_buffer_wait);
1130
1131		/*
1132		 * If there were errors on the buffer, and the buffer is not
1133		 * to be written, free the buffer. There is no point in caching
1134		 * invalid buffer.
1135		 */
1136		if ((b->read_error || b->write_error) &&
1137		    !test_bit(B_READING, &b->state) &&
1138		    !test_bit(B_WRITING, &b->state) &&
1139		    !test_bit(B_DIRTY, &b->state)) {
1140			__unlink_buffer(b);
1141			__free_buffer_wake(b);
1142		}
1143	}
1144
1145	dm_bufio_unlock(c);
1146}
1147EXPORT_SYMBOL_GPL(dm_bufio_release);
1148
1149void dm_bufio_mark_buffer_dirty(struct dm_buffer *b)
1150{
1151	struct dm_bufio_client *c = b->c;
1152
1153	dm_bufio_lock(c);
1154
1155	BUG_ON(test_bit(B_READING, &b->state));
1156
1157	if (!test_and_set_bit(B_DIRTY, &b->state))
1158		__relink_lru(b, LIST_DIRTY);
1159
1160	dm_bufio_unlock(c);
1161}
1162EXPORT_SYMBOL_GPL(dm_bufio_mark_buffer_dirty);
1163
1164void dm_bufio_write_dirty_buffers_async(struct dm_bufio_client *c)
1165{
1166	LIST_HEAD(write_list);
1167
1168	BUG_ON(dm_bufio_in_request());
1169
1170	dm_bufio_lock(c);
1171	__write_dirty_buffers_async(c, 0, &write_list);
1172	dm_bufio_unlock(c);
1173	__flush_write_list(&write_list);
1174}
1175EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers_async);
1176
1177/*
1178 * For performance, it is essential that the buffers are written asynchronously
1179 * and simultaneously (so that the block layer can merge the writes) and then
1180 * waited upon.
1181 *
1182 * Finally, we flush hardware disk cache.
1183 */
1184int dm_bufio_write_dirty_buffers(struct dm_bufio_client *c)
1185{
1186	int a, f;
1187	unsigned long buffers_processed = 0;
1188	struct dm_buffer *b, *tmp;
1189
1190	LIST_HEAD(write_list);
1191
1192	dm_bufio_lock(c);
1193	__write_dirty_buffers_async(c, 0, &write_list);
1194	dm_bufio_unlock(c);
1195	__flush_write_list(&write_list);
1196	dm_bufio_lock(c);
1197
1198again:
1199	list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
1200		int dropped_lock = 0;
1201
1202		if (buffers_processed < c->n_buffers[LIST_DIRTY])
1203			buffers_processed++;
1204
1205		BUG_ON(test_bit(B_READING, &b->state));
1206
1207		if (test_bit(B_WRITING, &b->state)) {
1208			if (buffers_processed < c->n_buffers[LIST_DIRTY]) {
1209				dropped_lock = 1;
1210				b->hold_count++;
1211				dm_bufio_unlock(c);
1212				wait_on_bit(&b->state, B_WRITING,
1213					    do_io_schedule,
1214					    TASK_UNINTERRUPTIBLE);
1215				dm_bufio_lock(c);
1216				b->hold_count--;
1217			} else
1218				wait_on_bit(&b->state, B_WRITING,
1219					    do_io_schedule,
1220					    TASK_UNINTERRUPTIBLE);
1221		}
1222
1223		if (!test_bit(B_DIRTY, &b->state) &&
1224		    !test_bit(B_WRITING, &b->state))
1225			__relink_lru(b, LIST_CLEAN);
1226
1227		dm_bufio_cond_resched();
1228
1229		/*
1230		 * If we dropped the lock, the list is no longer consistent,
1231		 * so we must restart the search.
1232		 *
1233		 * In the most common case, the buffer just processed is
1234		 * relinked to the clean list, so we won't loop scanning the
1235		 * same buffer again and again.
1236		 *
1237		 * This may livelock if there is another thread simultaneously
1238		 * dirtying buffers, so we count the number of buffers walked
1239		 * and if it exceeds the total number of buffers, it means that
1240		 * someone is doing some writes simultaneously with us.  In
1241		 * this case, stop, dropping the lock.
1242		 */
1243		if (dropped_lock)
1244			goto again;
1245	}
1246	wake_up(&c->free_buffer_wait);
1247	dm_bufio_unlock(c);
1248
1249	a = xchg(&c->async_write_error, 0);
1250	f = dm_bufio_issue_flush(c);
1251	if (a)
1252		return a;
1253
1254	return f;
1255}
1256EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers);
1257
1258/*
1259 * Use dm-io to send and empty barrier flush the device.
1260 */
1261int dm_bufio_issue_flush(struct dm_bufio_client *c)
1262{
1263	struct dm_io_request io_req = {
1264		.bi_rw = WRITE_FLUSH,
1265		.mem.type = DM_IO_KMEM,
1266		.mem.ptr.addr = NULL,
1267		.client = c->dm_io,
1268	};
1269	struct dm_io_region io_reg = {
1270		.bdev = c->bdev,
1271		.sector = 0,
1272		.count = 0,
1273	};
1274
1275	BUG_ON(dm_bufio_in_request());
1276
1277	return dm_io(&io_req, 1, &io_reg, NULL);
1278}
1279EXPORT_SYMBOL_GPL(dm_bufio_issue_flush);
1280
1281/*
1282 * We first delete any other buffer that may be at that new location.
1283 *
1284 * Then, we write the buffer to the original location if it was dirty.
1285 *
1286 * Then, if we are the only one who is holding the buffer, relink the buffer
1287 * in the hash queue for the new location.
1288 *
1289 * If there was someone else holding the buffer, we write it to the new
1290 * location but not relink it, because that other user needs to have the buffer
1291 * at the same place.
1292 */
1293void dm_bufio_release_move(struct dm_buffer *b, sector_t new_block)
1294{
1295	struct dm_bufio_client *c = b->c;
1296	struct dm_buffer *new;
1297
1298	BUG_ON(dm_bufio_in_request());
1299
1300	dm_bufio_lock(c);
1301
1302retry:
1303	new = __find(c, new_block);
1304	if (new) {
1305		if (new->hold_count) {
1306			__wait_for_free_buffer(c);
1307			goto retry;
1308		}
1309
1310		/*
1311		 * FIXME: Is there any point waiting for a write that's going
1312		 * to be overwritten in a bit?
1313		 */
1314		__make_buffer_clean(new);
1315		__unlink_buffer(new);
1316		__free_buffer_wake(new);
1317	}
1318
1319	BUG_ON(!b->hold_count);
1320	BUG_ON(test_bit(B_READING, &b->state));
1321
1322	__write_dirty_buffer(b, NULL);
1323	if (b->hold_count == 1) {
1324		wait_on_bit(&b->state, B_WRITING,
1325			    do_io_schedule, TASK_UNINTERRUPTIBLE);
1326		set_bit(B_DIRTY, &b->state);
1327		__unlink_buffer(b);
1328		__link_buffer(b, new_block, LIST_DIRTY);
1329	} else {
1330		sector_t old_block;
1331		wait_on_bit_lock(&b->state, B_WRITING,
1332				 do_io_schedule, TASK_UNINTERRUPTIBLE);
1333		/*
1334		 * Relink buffer to "new_block" so that write_callback
1335		 * sees "new_block" as a block number.
1336		 * After the write, link the buffer back to old_block.
1337		 * All this must be done in bufio lock, so that block number
1338		 * change isn't visible to other threads.
1339		 */
1340		old_block = b->block;
1341		__unlink_buffer(b);
1342		__link_buffer(b, new_block, b->list_mode);
1343		submit_io(b, WRITE, new_block, write_endio);
1344		wait_on_bit(&b->state, B_WRITING,
1345			    do_io_schedule, TASK_UNINTERRUPTIBLE);
1346		__unlink_buffer(b);
1347		__link_buffer(b, old_block, b->list_mode);
1348	}
1349
1350	dm_bufio_unlock(c);
1351	dm_bufio_release(b);
1352}
1353EXPORT_SYMBOL_GPL(dm_bufio_release_move);
1354
1355/*
1356 * Free the given buffer.
1357 *
1358 * This is just a hint, if the buffer is in use or dirty, this function
1359 * does nothing.
1360 */
1361void dm_bufio_forget(struct dm_bufio_client *c, sector_t block)
1362{
1363	struct dm_buffer *b;
1364
1365	dm_bufio_lock(c);
1366
1367	b = __find(c, block);
1368	if (b && likely(!b->hold_count) && likely(!b->state)) {
1369		__unlink_buffer(b);
1370		__free_buffer_wake(b);
1371	}
1372
1373	dm_bufio_unlock(c);
1374}
1375EXPORT_SYMBOL(dm_bufio_forget);
1376
1377void dm_bufio_set_minimum_buffers(struct dm_bufio_client *c, unsigned n)
1378{
1379	c->minimum_buffers = n;
1380}
1381EXPORT_SYMBOL(dm_bufio_set_minimum_buffers);
1382
1383unsigned dm_bufio_get_block_size(struct dm_bufio_client *c)
1384{
1385	return c->block_size;
1386}
1387EXPORT_SYMBOL_GPL(dm_bufio_get_block_size);
1388
1389sector_t dm_bufio_get_device_size(struct dm_bufio_client *c)
1390{
1391	return i_size_read(c->bdev->bd_inode) >>
1392			   (SECTOR_SHIFT + c->sectors_per_block_bits);
1393}
1394EXPORT_SYMBOL_GPL(dm_bufio_get_device_size);
1395
1396sector_t dm_bufio_get_block_number(struct dm_buffer *b)
1397{
1398	return b->block;
1399}
1400EXPORT_SYMBOL_GPL(dm_bufio_get_block_number);
1401
1402void *dm_bufio_get_block_data(struct dm_buffer *b)
1403{
1404	return b->data;
1405}
1406EXPORT_SYMBOL_GPL(dm_bufio_get_block_data);
1407
1408void *dm_bufio_get_aux_data(struct dm_buffer *b)
1409{
1410	return b + 1;
1411}
1412EXPORT_SYMBOL_GPL(dm_bufio_get_aux_data);
1413
1414struct dm_bufio_client *dm_bufio_get_client(struct dm_buffer *b)
1415{
1416	return b->c;
1417}
1418EXPORT_SYMBOL_GPL(dm_bufio_get_client);
1419
1420static void drop_buffers(struct dm_bufio_client *c)
1421{
1422	struct dm_buffer *b;
1423	int i;
 
1424
1425	BUG_ON(dm_bufio_in_request());
1426
1427	/*
1428	 * An optimization so that the buffers are not written one-by-one.
1429	 */
1430	dm_bufio_write_dirty_buffers_async(c);
1431
1432	dm_bufio_lock(c);
1433
1434	while ((b = __get_unclaimed_buffer(c)))
1435		__free_buffer_wake(b);
1436
1437	for (i = 0; i < LIST_SIZE; i++)
1438		list_for_each_entry(b, &c->lru[i], lru_list)
 
 
1439			DMERR("leaked buffer %llx, hold count %u, list %d",
1440			      (unsigned long long)b->block, b->hold_count, i);
 
 
 
 
 
 
 
 
 
 
1441
1442	for (i = 0; i < LIST_SIZE; i++)
1443		BUG_ON(!list_empty(&c->lru[i]));
1444
1445	dm_bufio_unlock(c);
1446}
1447
1448/*
1449 * Test if the buffer is unused and too old, and commit it.
1450 * At if noio is set, we must not do any I/O because we hold
1451 * dm_bufio_clients_lock and we would risk deadlock if the I/O gets rerouted to
1452 * different bufio client.
 
 
1453 */
1454static int __cleanup_old_buffer(struct dm_buffer *b, gfp_t gfp,
1455				unsigned long max_jiffies)
1456{
1457	if (jiffies - b->last_accessed < max_jiffies)
1458		return 0;
1459
1460	if (!(gfp & __GFP_IO)) {
1461		if (test_bit(B_READING, &b->state) ||
1462		    test_bit(B_WRITING, &b->state) ||
1463		    test_bit(B_DIRTY, &b->state))
1464			return 0;
1465	}
1466
1467	if (b->hold_count)
1468		return 0;
1469
1470	__make_buffer_clean(b);
1471	__unlink_buffer(b);
1472	__free_buffer_wake(b);
1473
1474	return 1;
 
 
 
 
 
 
1475}
1476
1477static long __scan(struct dm_bufio_client *c, unsigned long nr_to_scan,
1478		   gfp_t gfp_mask)
1479{
1480	int l;
1481	struct dm_buffer *b, *tmp;
1482	long freed = 0;
 
 
1483
1484	for (l = 0; l < LIST_SIZE; l++) {
1485		list_for_each_entry_safe_reverse(b, tmp, &c->lru[l], lru_list) {
1486			freed += __cleanup_old_buffer(b, gfp_mask, 0);
1487			if (!--nr_to_scan)
1488				break;
 
 
1489		}
1490		dm_bufio_cond_resched();
1491	}
1492	return freed;
1493}
1494
1495static unsigned long
1496dm_bufio_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
1497{
1498	struct dm_bufio_client *c;
1499	unsigned long freed;
1500
1501	c = container_of(shrink, struct dm_bufio_client, shrinker);
1502	if (sc->gfp_mask & __GFP_IO)
1503		dm_bufio_lock(c);
1504	else if (!dm_bufio_trylock(c))
1505		return SHRINK_STOP;
1506
1507	freed  = __scan(c, sc->nr_to_scan, sc->gfp_mask);
1508	dm_bufio_unlock(c);
1509	return freed;
1510}
1511
1512static unsigned long
1513dm_bufio_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
1514{
1515	struct dm_bufio_client *c;
1516	unsigned long count;
1517
1518	c = container_of(shrink, struct dm_bufio_client, shrinker);
1519	if (sc->gfp_mask & __GFP_IO)
1520		dm_bufio_lock(c);
1521	else if (!dm_bufio_trylock(c))
1522		return 0;
1523
1524	count = c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY];
1525	dm_bufio_unlock(c);
1526	return count;
1527}
1528
1529/*
1530 * Create the buffering interface
1531 */
1532struct dm_bufio_client *dm_bufio_client_create(struct block_device *bdev, unsigned block_size,
1533					       unsigned reserved_buffers, unsigned aux_size,
1534					       void (*alloc_callback)(struct dm_buffer *),
1535					       void (*write_callback)(struct dm_buffer *))
1536{
1537	int r;
1538	struct dm_bufio_client *c;
1539	unsigned i;
1540
1541	BUG_ON(block_size < 1 << SECTOR_SHIFT ||
1542	       (block_size & (block_size - 1)));
1543
1544	c = kmalloc(sizeof(*c), GFP_KERNEL);
1545	if (!c) {
1546		r = -ENOMEM;
1547		goto bad_client;
1548	}
1549	c->cache_hash = vmalloc(sizeof(struct hlist_head) << DM_BUFIO_HASH_BITS);
1550	if (!c->cache_hash) {
1551		r = -ENOMEM;
1552		goto bad_hash;
1553	}
1554
1555	c->bdev = bdev;
1556	c->block_size = block_size;
1557	c->sectors_per_block_bits = ffs(block_size) - 1 - SECTOR_SHIFT;
1558	c->pages_per_block_bits = (ffs(block_size) - 1 >= PAGE_SHIFT) ?
1559				  ffs(block_size) - 1 - PAGE_SHIFT : 0;
1560	c->blocks_per_page_bits = (ffs(block_size) - 1 < PAGE_SHIFT ?
1561				  PAGE_SHIFT - (ffs(block_size) - 1) : 0);
1562
1563	c->aux_size = aux_size;
1564	c->alloc_callback = alloc_callback;
1565	c->write_callback = write_callback;
1566
1567	for (i = 0; i < LIST_SIZE; i++) {
1568		INIT_LIST_HEAD(&c->lru[i]);
1569		c->n_buffers[i] = 0;
1570	}
1571
1572	for (i = 0; i < 1 << DM_BUFIO_HASH_BITS; i++)
1573		INIT_HLIST_HEAD(&c->cache_hash[i]);
1574
1575	mutex_init(&c->lock);
1576	INIT_LIST_HEAD(&c->reserved_buffers);
1577	c->need_reserved_buffers = reserved_buffers;
1578
1579	c->minimum_buffers = DM_BUFIO_MIN_BUFFERS;
1580
1581	init_waitqueue_head(&c->free_buffer_wait);
1582	c->async_write_error = 0;
1583
1584	c->dm_io = dm_io_client_create();
1585	if (IS_ERR(c->dm_io)) {
1586		r = PTR_ERR(c->dm_io);
1587		goto bad_dm_io;
1588	}
1589
1590	mutex_lock(&dm_bufio_clients_lock);
1591	if (c->blocks_per_page_bits) {
1592		if (!DM_BUFIO_CACHE_NAME(c)) {
1593			DM_BUFIO_CACHE_NAME(c) = kasprintf(GFP_KERNEL, "dm_bufio_cache-%u", c->block_size);
1594			if (!DM_BUFIO_CACHE_NAME(c)) {
1595				r = -ENOMEM;
1596				mutex_unlock(&dm_bufio_clients_lock);
1597				goto bad_cache;
1598			}
1599		}
1600
1601		if (!DM_BUFIO_CACHE(c)) {
1602			DM_BUFIO_CACHE(c) = kmem_cache_create(DM_BUFIO_CACHE_NAME(c),
1603							      c->block_size,
1604							      c->block_size, 0, NULL);
1605			if (!DM_BUFIO_CACHE(c)) {
1606				r = -ENOMEM;
1607				mutex_unlock(&dm_bufio_clients_lock);
1608				goto bad_cache;
1609			}
1610		}
1611	}
1612	mutex_unlock(&dm_bufio_clients_lock);
1613
1614	while (c->need_reserved_buffers) {
1615		struct dm_buffer *b = alloc_buffer(c, GFP_KERNEL);
1616
1617		if (!b) {
1618			r = -ENOMEM;
1619			goto bad_buffer;
1620		}
1621		__free_buffer_wake(b);
1622	}
1623
1624	mutex_lock(&dm_bufio_clients_lock);
1625	dm_bufio_client_count++;
1626	list_add(&c->client_list, &dm_bufio_all_clients);
1627	__cache_size_refresh();
1628	mutex_unlock(&dm_bufio_clients_lock);
1629
1630	c->shrinker.count_objects = dm_bufio_shrink_count;
1631	c->shrinker.scan_objects = dm_bufio_shrink_scan;
1632	c->shrinker.seeks = 1;
1633	c->shrinker.batch = 0;
1634	register_shrinker(&c->shrinker);
1635
1636	return c;
1637
1638bad_buffer:
1639bad_cache:
1640	while (!list_empty(&c->reserved_buffers)) {
1641		struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1642						 struct dm_buffer, lru_list);
1643		list_del(&b->lru_list);
1644		free_buffer(b);
1645	}
1646	dm_io_client_destroy(c->dm_io);
1647bad_dm_io:
1648	vfree(c->cache_hash);
1649bad_hash:
1650	kfree(c);
1651bad_client:
1652	return ERR_PTR(r);
1653}
1654EXPORT_SYMBOL_GPL(dm_bufio_client_create);
1655
1656/*
1657 * Free the buffering interface.
1658 * It is required that there are no references on any buffers.
1659 */
1660void dm_bufio_client_destroy(struct dm_bufio_client *c)
1661{
1662	unsigned i;
1663
1664	drop_buffers(c);
1665
1666	unregister_shrinker(&c->shrinker);
1667
1668	mutex_lock(&dm_bufio_clients_lock);
1669
1670	list_del(&c->client_list);
1671	dm_bufio_client_count--;
1672	__cache_size_refresh();
1673
1674	mutex_unlock(&dm_bufio_clients_lock);
1675
1676	for (i = 0; i < 1 << DM_BUFIO_HASH_BITS; i++)
1677		BUG_ON(!hlist_empty(&c->cache_hash[i]));
1678
1679	BUG_ON(c->need_reserved_buffers);
1680
1681	while (!list_empty(&c->reserved_buffers)) {
1682		struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1683						 struct dm_buffer, lru_list);
1684		list_del(&b->lru_list);
1685		free_buffer(b);
1686	}
1687
1688	for (i = 0; i < LIST_SIZE; i++)
1689		if (c->n_buffers[i])
1690			DMERR("leaked buffer count %d: %ld", i, c->n_buffers[i]);
1691
1692	for (i = 0; i < LIST_SIZE; i++)
1693		BUG_ON(c->n_buffers[i]);
1694
1695	dm_io_client_destroy(c->dm_io);
1696	vfree(c->cache_hash);
1697	kfree(c);
1698}
1699EXPORT_SYMBOL_GPL(dm_bufio_client_destroy);
1700
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1701static void cleanup_old_buffers(void)
1702{
1703	unsigned long max_age = ACCESS_ONCE(dm_bufio_max_age);
1704	struct dm_bufio_client *c;
1705
1706	if (max_age > ULONG_MAX / HZ)
1707		max_age = ULONG_MAX / HZ;
1708
1709	mutex_lock(&dm_bufio_clients_lock);
1710	list_for_each_entry(c, &dm_bufio_all_clients, client_list) {
1711		if (!dm_bufio_trylock(c))
1712			continue;
1713
1714		while (!list_empty(&c->lru[LIST_CLEAN])) {
1715			struct dm_buffer *b;
1716			b = list_entry(c->lru[LIST_CLEAN].prev,
1717				       struct dm_buffer, lru_list);
1718			if (!__cleanup_old_buffer(b, 0, max_age * HZ))
1719				break;
1720			dm_bufio_cond_resched();
1721		}
1722
1723		dm_bufio_unlock(c);
1724		dm_bufio_cond_resched();
1725	}
1726	mutex_unlock(&dm_bufio_clients_lock);
1727}
1728
1729static struct workqueue_struct *dm_bufio_wq;
1730static struct delayed_work dm_bufio_work;
1731
1732static void work_fn(struct work_struct *w)
1733{
1734	cleanup_old_buffers();
1735
1736	queue_delayed_work(dm_bufio_wq, &dm_bufio_work,
1737			   DM_BUFIO_WORK_TIMER_SECS * HZ);
1738}
1739
1740/*----------------------------------------------------------------
1741 * Module setup
1742 *--------------------------------------------------------------*/
1743
1744/*
1745 * This is called only once for the whole dm_bufio module.
1746 * It initializes memory limit.
1747 */
1748static int __init dm_bufio_init(void)
1749{
1750	__u64 mem;
1751
1752	dm_bufio_allocated_kmem_cache = 0;
1753	dm_bufio_allocated_get_free_pages = 0;
1754	dm_bufio_allocated_vmalloc = 0;
1755	dm_bufio_current_allocated = 0;
1756
1757	memset(&dm_bufio_caches, 0, sizeof dm_bufio_caches);
1758	memset(&dm_bufio_cache_names, 0, sizeof dm_bufio_cache_names);
1759
1760	mem = (__u64)((totalram_pages - totalhigh_pages) *
1761		      DM_BUFIO_MEMORY_PERCENT / 100) << PAGE_SHIFT;
1762
1763	if (mem > ULONG_MAX)
1764		mem = ULONG_MAX;
1765
1766#ifdef CONFIG_MMU
1767	/*
1768	 * Get the size of vmalloc space the same way as VMALLOC_TOTAL
1769	 * in fs/proc/internal.h
1770	 */
1771	if (mem > (VMALLOC_END - VMALLOC_START) * DM_BUFIO_VMALLOC_PERCENT / 100)
1772		mem = (VMALLOC_END - VMALLOC_START) * DM_BUFIO_VMALLOC_PERCENT / 100;
1773#endif
1774
1775	dm_bufio_default_cache_size = mem;
1776
1777	mutex_lock(&dm_bufio_clients_lock);
1778	__cache_size_refresh();
1779	mutex_unlock(&dm_bufio_clients_lock);
1780
1781	dm_bufio_wq = create_singlethread_workqueue("dm_bufio_cache");
1782	if (!dm_bufio_wq)
1783		return -ENOMEM;
1784
1785	INIT_DELAYED_WORK(&dm_bufio_work, work_fn);
1786	queue_delayed_work(dm_bufio_wq, &dm_bufio_work,
1787			   DM_BUFIO_WORK_TIMER_SECS * HZ);
1788
1789	return 0;
1790}
1791
1792/*
1793 * This is called once when unloading the dm_bufio module.
1794 */
1795static void __exit dm_bufio_exit(void)
1796{
1797	int bug = 0;
1798	int i;
1799
1800	cancel_delayed_work_sync(&dm_bufio_work);
1801	destroy_workqueue(dm_bufio_wq);
1802
1803	for (i = 0; i < ARRAY_SIZE(dm_bufio_caches); i++) {
1804		struct kmem_cache *kc = dm_bufio_caches[i];
1805
1806		if (kc)
1807			kmem_cache_destroy(kc);
1808	}
1809
1810	for (i = 0; i < ARRAY_SIZE(dm_bufio_cache_names); i++)
1811		kfree(dm_bufio_cache_names[i]);
1812
1813	if (dm_bufio_client_count) {
1814		DMCRIT("%s: dm_bufio_client_count leaked: %d",
1815			__func__, dm_bufio_client_count);
1816		bug = 1;
1817	}
1818
1819	if (dm_bufio_current_allocated) {
1820		DMCRIT("%s: dm_bufio_current_allocated leaked: %lu",
1821			__func__, dm_bufio_current_allocated);
1822		bug = 1;
1823	}
1824
1825	if (dm_bufio_allocated_get_free_pages) {
1826		DMCRIT("%s: dm_bufio_allocated_get_free_pages leaked: %lu",
1827		       __func__, dm_bufio_allocated_get_free_pages);
1828		bug = 1;
1829	}
1830
1831	if (dm_bufio_allocated_vmalloc) {
1832		DMCRIT("%s: dm_bufio_vmalloc leaked: %lu",
1833		       __func__, dm_bufio_allocated_vmalloc);
1834		bug = 1;
1835	}
1836
1837	if (bug)
1838		BUG();
1839}
1840
1841module_init(dm_bufio_init)
1842module_exit(dm_bufio_exit)
1843
1844module_param_named(max_cache_size_bytes, dm_bufio_cache_size, ulong, S_IRUGO | S_IWUSR);
1845MODULE_PARM_DESC(max_cache_size_bytes, "Size of metadata cache");
1846
1847module_param_named(max_age_seconds, dm_bufio_max_age, uint, S_IRUGO | S_IWUSR);
1848MODULE_PARM_DESC(max_age_seconds, "Max age of a buffer in seconds");
 
 
 
1849
1850module_param_named(peak_allocated_bytes, dm_bufio_peak_allocated, ulong, S_IRUGO | S_IWUSR);
1851MODULE_PARM_DESC(peak_allocated_bytes, "Tracks the maximum allocated memory");
1852
1853module_param_named(allocated_kmem_cache_bytes, dm_bufio_allocated_kmem_cache, ulong, S_IRUGO);
1854MODULE_PARM_DESC(allocated_kmem_cache_bytes, "Memory allocated with kmem_cache_alloc");
1855
1856module_param_named(allocated_get_free_pages_bytes, dm_bufio_allocated_get_free_pages, ulong, S_IRUGO);
1857MODULE_PARM_DESC(allocated_get_free_pages_bytes, "Memory allocated with get_free_pages");
1858
1859module_param_named(allocated_vmalloc_bytes, dm_bufio_allocated_vmalloc, ulong, S_IRUGO);
1860MODULE_PARM_DESC(allocated_vmalloc_bytes, "Memory allocated with vmalloc");
1861
1862module_param_named(current_allocated_bytes, dm_bufio_current_allocated, ulong, S_IRUGO);
1863MODULE_PARM_DESC(current_allocated_bytes, "Memory currently used by the cache");
1864
1865MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
1866MODULE_DESCRIPTION(DM_NAME " buffered I/O library");
1867MODULE_LICENSE("GPL");