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