1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Copyright (C) 2018 Red Hat. All rights reserved.
   4 *
   5 * This file is released under the GPL.
   6 */
   7
   8#include <linux/device-mapper.h>
   9#include <linux/module.h>
  10#include <linux/init.h>
  11#include <linux/vmalloc.h>
  12#include <linux/kthread.h>
  13#include <linux/dm-io.h>
  14#include <linux/dm-kcopyd.h>
  15#include <linux/dax.h>
  16#include <linux/pfn_t.h>
  17#include <linux/libnvdimm.h>
  18
  19#define DM_MSG_PREFIX "writecache"
  20
  21#define HIGH_WATERMARK			50
  22#define LOW_WATERMARK			45
  23#define MAX_WRITEBACK_JOBS		0
  24#define ENDIO_LATENCY			16
  25#define WRITEBACK_LATENCY		64
  26#define AUTOCOMMIT_BLOCKS_SSD		65536
  27#define AUTOCOMMIT_BLOCKS_PMEM		64
  28#define AUTOCOMMIT_MSEC			1000
  29#define MAX_AGE_DIV			16
  30#define MAX_AGE_UNSPECIFIED		-1UL
  31
  32#define BITMAP_GRANULARITY	65536
  33#if BITMAP_GRANULARITY < PAGE_SIZE
  34#undef BITMAP_GRANULARITY
  35#define BITMAP_GRANULARITY	PAGE_SIZE
  36#endif
  37
  38#if IS_ENABLED(CONFIG_ARCH_HAS_PMEM_API) && IS_ENABLED(CONFIG_DAX_DRIVER)
  39#define DM_WRITECACHE_HAS_PMEM
  40#endif
  41
  42#ifdef DM_WRITECACHE_HAS_PMEM
  43#define pmem_assign(dest, src)					\
  44do {								\
  45	typeof(dest) uniq = (src);				\
  46	memcpy_flushcache(&(dest), &uniq, sizeof(dest));	\
  47} while (0)
  48#else
  49#define pmem_assign(dest, src)	((dest) = (src))
  50#endif
  51
  52#if defined(__HAVE_ARCH_MEMCPY_MCSAFE) && defined(DM_WRITECACHE_HAS_PMEM)
  53#define DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
  54#endif
  55
  56#define MEMORY_SUPERBLOCK_MAGIC		0x23489321
  57#define MEMORY_SUPERBLOCK_VERSION	1
  58
  59struct wc_memory_entry {
  60	__le64 original_sector;
  61	__le64 seq_count;
  62};
  63
  64struct wc_memory_superblock {
  65	union {
  66		struct {
  67			__le32 magic;
  68			__le32 version;
  69			__le32 block_size;
  70			__le32 pad;
  71			__le64 n_blocks;
  72			__le64 seq_count;
  73		};
  74		__le64 padding[8];
  75	};
  76	struct wc_memory_entry entries[0];
  77};
  78
  79struct wc_entry {
  80	struct rb_node rb_node;
  81	struct list_head lru;
  82	unsigned short wc_list_contiguous;
  83	bool write_in_progress
  84#if BITS_PER_LONG == 64
  85		:1
  86#endif
  87	;
  88	unsigned long index
  89#if BITS_PER_LONG == 64
  90		:47
  91#endif
  92	;
  93	unsigned long age;
  94#ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
  95	uint64_t original_sector;
  96	uint64_t seq_count;
  97#endif
  98};
  99
 100#ifdef DM_WRITECACHE_HAS_PMEM
 101#define WC_MODE_PMEM(wc)			((wc)->pmem_mode)
 102#define WC_MODE_FUA(wc)				((wc)->writeback_fua)
 103#else
 104#define WC_MODE_PMEM(wc)			false
 105#define WC_MODE_FUA(wc)				false
 106#endif
 107#define WC_MODE_SORT_FREELIST(wc)		(!WC_MODE_PMEM(wc))
 108
 109struct dm_writecache {
 110	struct mutex lock;
 111	struct list_head lru;
 112	union {
 113		struct list_head freelist;
 114		struct {
 115			struct rb_root freetree;
 116			struct wc_entry *current_free;
 117		};
 118	};
 119	struct rb_root tree;
 120
 121	size_t freelist_size;
 122	size_t writeback_size;
 123	size_t freelist_high_watermark;
 124	size_t freelist_low_watermark;
 125	unsigned long max_age;
 126
 127	unsigned uncommitted_blocks;
 128	unsigned autocommit_blocks;
 129	unsigned max_writeback_jobs;
 130
 131	int error;
 132
 133	unsigned long autocommit_jiffies;
 134	struct timer_list autocommit_timer;
 135	struct wait_queue_head freelist_wait;
 136
 137	struct timer_list max_age_timer;
 138
 139	atomic_t bio_in_progress[2];
 140	struct wait_queue_head bio_in_progress_wait[2];
 141
 142	struct dm_target *ti;
 143	struct dm_dev *dev;
 144	struct dm_dev *ssd_dev;
 145	sector_t start_sector;
 146	void *memory_map;
 147	uint64_t memory_map_size;
 148	size_t metadata_sectors;
 149	size_t n_blocks;
 150	uint64_t seq_count;
 151	void *block_start;
 152	struct wc_entry *entries;
 153	unsigned block_size;
 154	unsigned char block_size_bits;
 155
 156	bool pmem_mode:1;
 157	bool writeback_fua:1;
 158
 159	bool overwrote_committed:1;
 160	bool memory_vmapped:1;
 161
 162	bool high_wm_percent_set:1;
 163	bool low_wm_percent_set:1;
 164	bool max_writeback_jobs_set:1;
 165	bool autocommit_blocks_set:1;
 166	bool autocommit_time_set:1;
 167	bool writeback_fua_set:1;
 168	bool flush_on_suspend:1;
 169	bool cleaner:1;
 170
 171	unsigned writeback_all;
 172	struct workqueue_struct *writeback_wq;
 173	struct work_struct writeback_work;
 174	struct work_struct flush_work;
 175
 176	struct dm_io_client *dm_io;
 177
 178	raw_spinlock_t endio_list_lock;
 179	struct list_head endio_list;
 180	struct task_struct *endio_thread;
 181
 182	struct task_struct *flush_thread;
 183	struct bio_list flush_list;
 184
 185	struct dm_kcopyd_client *dm_kcopyd;
 186	unsigned long *dirty_bitmap;
 187	unsigned dirty_bitmap_size;
 188
 189	struct bio_set bio_set;
 190	mempool_t copy_pool;
 191};
 192
 193#define WB_LIST_INLINE		16
 194
 195struct writeback_struct {
 196	struct list_head endio_entry;
 197	struct dm_writecache *wc;
 198	struct wc_entry **wc_list;
 199	unsigned wc_list_n;
 200	struct wc_entry *wc_list_inline[WB_LIST_INLINE];
 201	struct bio bio;
 202};
 203
 204struct copy_struct {
 205	struct list_head endio_entry;
 206	struct dm_writecache *wc;
 207	struct wc_entry *e;
 208	unsigned n_entries;
 209	int error;
 210};
 211
 212DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(dm_writecache_throttle,
 213					    "A percentage of time allocated for data copying");
 214
 215static void wc_lock(struct dm_writecache *wc)
 216{
 217	mutex_lock(&wc->lock);
 218}
 219
 220static void wc_unlock(struct dm_writecache *wc)
 221{
 222	mutex_unlock(&wc->lock);
 223}
 224
 225#ifdef DM_WRITECACHE_HAS_PMEM
 226static int persistent_memory_claim(struct dm_writecache *wc)
 227{
 228	int r;
 229	loff_t s;
 230	long p, da;
 231	pfn_t pfn;
 232	int id;
 233	struct page **pages;
 234	sector_t offset;
 235
 236	wc->memory_vmapped = false;
 237
 
 
 
 
 238	s = wc->memory_map_size;
 239	p = s >> PAGE_SHIFT;
 240	if (!p) {
 241		r = -EINVAL;
 242		goto err1;
 243	}
 244	if (p != s >> PAGE_SHIFT) {
 245		r = -EOVERFLOW;
 246		goto err1;
 247	}
 248
 249	offset = get_start_sect(wc->ssd_dev->bdev);
 250	if (offset & (PAGE_SIZE / 512 - 1)) {
 251		r = -EINVAL;
 252		goto err1;
 253	}
 254	offset >>= PAGE_SHIFT - 9;
 255
 256	id = dax_read_lock();
 257
 258	da = dax_direct_access(wc->ssd_dev->dax_dev, offset, p, &wc->memory_map, &pfn);
 259	if (da < 0) {
 260		wc->memory_map = NULL;
 261		r = da;
 262		goto err2;
 263	}
 264	if (!pfn_t_has_page(pfn)) {
 265		wc->memory_map = NULL;
 266		r = -EOPNOTSUPP;
 267		goto err2;
 268	}
 269	if (da != p) {
 270		long i;
 271		wc->memory_map = NULL;
 272		pages = kvmalloc_array(p, sizeof(struct page *), GFP_KERNEL);
 273		if (!pages) {
 274			r = -ENOMEM;
 275			goto err2;
 276		}
 277		i = 0;
 278		do {
 279			long daa;
 280			daa = dax_direct_access(wc->ssd_dev->dax_dev, offset + i, p - i,
 281						NULL, &pfn);
 282			if (daa <= 0) {
 283				r = daa ? daa : -EINVAL;
 284				goto err3;
 285			}
 286			if (!pfn_t_has_page(pfn)) {
 287				r = -EOPNOTSUPP;
 288				goto err3;
 289			}
 290			while (daa-- && i < p) {
 291				pages[i++] = pfn_t_to_page(pfn);
 292				pfn.val++;
 293				if (!(i & 15))
 294					cond_resched();
 295			}
 296		} while (i < p);
 297		wc->memory_map = vmap(pages, p, VM_MAP, PAGE_KERNEL);
 298		if (!wc->memory_map) {
 299			r = -ENOMEM;
 300			goto err3;
 301		}
 302		kvfree(pages);
 303		wc->memory_vmapped = true;
 304	}
 305
 306	dax_read_unlock(id);
 307
 308	wc->memory_map += (size_t)wc->start_sector << SECTOR_SHIFT;
 309	wc->memory_map_size -= (size_t)wc->start_sector << SECTOR_SHIFT;
 310
 311	return 0;
 312err3:
 313	kvfree(pages);
 314err2:
 315	dax_read_unlock(id);
 316err1:
 317	return r;
 318}
 319#else
 320static int persistent_memory_claim(struct dm_writecache *wc)
 321{
 322	BUG();
 323}
 324#endif
 325
 326static void persistent_memory_release(struct dm_writecache *wc)
 327{
 328	if (wc->memory_vmapped)
 329		vunmap(wc->memory_map - ((size_t)wc->start_sector << SECTOR_SHIFT));
 330}
 331
 332static struct page *persistent_memory_page(void *addr)
 333{
 334	if (is_vmalloc_addr(addr))
 335		return vmalloc_to_page(addr);
 336	else
 337		return virt_to_page(addr);
 338}
 339
 340static unsigned persistent_memory_page_offset(void *addr)
 341{
 342	return (unsigned long)addr & (PAGE_SIZE - 1);
 343}
 344
 345static void persistent_memory_flush_cache(void *ptr, size_t size)
 346{
 347	if (is_vmalloc_addr(ptr))
 348		flush_kernel_vmap_range(ptr, size);
 349}
 350
 351static void persistent_memory_invalidate_cache(void *ptr, size_t size)
 352{
 353	if (is_vmalloc_addr(ptr))
 354		invalidate_kernel_vmap_range(ptr, size);
 355}
 356
 357static struct wc_memory_superblock *sb(struct dm_writecache *wc)
 358{
 359	return wc->memory_map;
 360}
 361
 362static struct wc_memory_entry *memory_entry(struct dm_writecache *wc, struct wc_entry *e)
 363{
 364	return &sb(wc)->entries[e->index];
 365}
 366
 367static void *memory_data(struct dm_writecache *wc, struct wc_entry *e)
 368{
 369	return (char *)wc->block_start + (e->index << wc->block_size_bits);
 370}
 371
 372static sector_t cache_sector(struct dm_writecache *wc, struct wc_entry *e)
 373{
 374	return wc->start_sector + wc->metadata_sectors +
 375		((sector_t)e->index << (wc->block_size_bits - SECTOR_SHIFT));
 376}
 377
 378static uint64_t read_original_sector(struct dm_writecache *wc, struct wc_entry *e)
 379{
 380#ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 381	return e->original_sector;
 382#else
 383	return le64_to_cpu(memory_entry(wc, e)->original_sector);
 384#endif
 385}
 386
 387static uint64_t read_seq_count(struct dm_writecache *wc, struct wc_entry *e)
 388{
 389#ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 390	return e->seq_count;
 391#else
 392	return le64_to_cpu(memory_entry(wc, e)->seq_count);
 393#endif
 394}
 395
 396static void clear_seq_count(struct dm_writecache *wc, struct wc_entry *e)
 397{
 398#ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 399	e->seq_count = -1;
 400#endif
 401	pmem_assign(memory_entry(wc, e)->seq_count, cpu_to_le64(-1));
 402}
 403
 404static void write_original_sector_seq_count(struct dm_writecache *wc, struct wc_entry *e,
 405					    uint64_t original_sector, uint64_t seq_count)
 406{
 407	struct wc_memory_entry me;
 408#ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 409	e->original_sector = original_sector;
 410	e->seq_count = seq_count;
 411#endif
 412	me.original_sector = cpu_to_le64(original_sector);
 413	me.seq_count = cpu_to_le64(seq_count);
 414	pmem_assign(*memory_entry(wc, e), me);
 415}
 416
 417#define writecache_error(wc, err, msg, arg...)				\
 418do {									\
 419	if (!cmpxchg(&(wc)->error, 0, err))				\
 420		DMERR(msg, ##arg);					\
 421	wake_up(&(wc)->freelist_wait);					\
 422} while (0)
 423
 424#define writecache_has_error(wc)	(unlikely(READ_ONCE((wc)->error)))
 425
 426static void writecache_flush_all_metadata(struct dm_writecache *wc)
 427{
 428	if (!WC_MODE_PMEM(wc))
 429		memset(wc->dirty_bitmap, -1, wc->dirty_bitmap_size);
 430}
 431
 432static void writecache_flush_region(struct dm_writecache *wc, void *ptr, size_t size)
 433{
 434	if (!WC_MODE_PMEM(wc))
 435		__set_bit(((char *)ptr - (char *)wc->memory_map) / BITMAP_GRANULARITY,
 436			  wc->dirty_bitmap);
 437}
 438
 439static void writecache_disk_flush(struct dm_writecache *wc, struct dm_dev *dev);
 440
 441struct io_notify {
 442	struct dm_writecache *wc;
 443	struct completion c;
 444	atomic_t count;
 445};
 446
 447static void writecache_notify_io(unsigned long error, void *context)
 448{
 449	struct io_notify *endio = context;
 450
 451	if (unlikely(error != 0))
 452		writecache_error(endio->wc, -EIO, "error writing metadata");
 453	BUG_ON(atomic_read(&endio->count) <= 0);
 454	if (atomic_dec_and_test(&endio->count))
 455		complete(&endio->c);
 456}
 457
 458static void writecache_wait_for_ios(struct dm_writecache *wc, int direction)
 459{
 460	wait_event(wc->bio_in_progress_wait[direction],
 461		   !atomic_read(&wc->bio_in_progress[direction]));
 462}
 463
 464static void ssd_commit_flushed(struct dm_writecache *wc, bool wait_for_ios)
 465{
 466	struct dm_io_region region;
 467	struct dm_io_request req;
 468	struct io_notify endio = {
 469		wc,
 470		COMPLETION_INITIALIZER_ONSTACK(endio.c),
 471		ATOMIC_INIT(1),
 472	};
 473	unsigned bitmap_bits = wc->dirty_bitmap_size * 8;
 474	unsigned i = 0;
 475
 476	while (1) {
 477		unsigned j;
 478		i = find_next_bit(wc->dirty_bitmap, bitmap_bits, i);
 479		if (unlikely(i == bitmap_bits))
 480			break;
 481		j = find_next_zero_bit(wc->dirty_bitmap, bitmap_bits, i);
 482
 483		region.bdev = wc->ssd_dev->bdev;
 484		region.sector = (sector_t)i * (BITMAP_GRANULARITY >> SECTOR_SHIFT);
 485		region.count = (sector_t)(j - i) * (BITMAP_GRANULARITY >> SECTOR_SHIFT);
 486
 487		if (unlikely(region.sector >= wc->metadata_sectors))
 488			break;
 489		if (unlikely(region.sector + region.count > wc->metadata_sectors))
 490			region.count = wc->metadata_sectors - region.sector;
 491
 492		region.sector += wc->start_sector;
 493		atomic_inc(&endio.count);
 494		req.bi_op = REQ_OP_WRITE;
 495		req.bi_op_flags = REQ_SYNC;
 496		req.mem.type = DM_IO_VMA;
 497		req.mem.ptr.vma = (char *)wc->memory_map + (size_t)i * BITMAP_GRANULARITY;
 498		req.client = wc->dm_io;
 499		req.notify.fn = writecache_notify_io;
 500		req.notify.context = &endio;
 501
 502		/* writing via async dm-io (implied by notify.fn above) won't return an error */
 503	        (void) dm_io(&req, 1, &region, NULL);
 504		i = j;
 505	}
 506
 507	writecache_notify_io(0, &endio);
 508	wait_for_completion_io(&endio.c);
 509
 510	if (wait_for_ios)
 511		writecache_wait_for_ios(wc, WRITE);
 512
 513	writecache_disk_flush(wc, wc->ssd_dev);
 514
 515	memset(wc->dirty_bitmap, 0, wc->dirty_bitmap_size);
 516}
 517
 518static void ssd_commit_superblock(struct dm_writecache *wc)
 519{
 520	int r;
 521	struct dm_io_region region;
 522	struct dm_io_request req;
 523
 524	region.bdev = wc->ssd_dev->bdev;
 525	region.sector = 0;
 526	region.count = PAGE_SIZE;
 527
 528	if (unlikely(region.sector + region.count > wc->metadata_sectors))
 529		region.count = wc->metadata_sectors - region.sector;
 530
 531	region.sector += wc->start_sector;
 532
 533	req.bi_op = REQ_OP_WRITE;
 534	req.bi_op_flags = REQ_SYNC | REQ_FUA;
 535	req.mem.type = DM_IO_VMA;
 536	req.mem.ptr.vma = (char *)wc->memory_map;
 537	req.client = wc->dm_io;
 538	req.notify.fn = NULL;
 539	req.notify.context = NULL;
 540
 541	r = dm_io(&req, 1, &region, NULL);
 542	if (unlikely(r))
 543		writecache_error(wc, r, "error writing superblock");
 544}
 545
 546static void writecache_commit_flushed(struct dm_writecache *wc, bool wait_for_ios)
 547{
 548	if (WC_MODE_PMEM(wc))
 549		pmem_wmb();
 550	else
 551		ssd_commit_flushed(wc, wait_for_ios);
 552}
 553
 554static void writecache_disk_flush(struct dm_writecache *wc, struct dm_dev *dev)
 555{
 556	int r;
 557	struct dm_io_region region;
 558	struct dm_io_request req;
 559
 560	region.bdev = dev->bdev;
 561	region.sector = 0;
 562	region.count = 0;
 563	req.bi_op = REQ_OP_WRITE;
 564	req.bi_op_flags = REQ_PREFLUSH;
 565	req.mem.type = DM_IO_KMEM;
 566	req.mem.ptr.addr = NULL;
 567	req.client = wc->dm_io;
 568	req.notify.fn = NULL;
 569
 570	r = dm_io(&req, 1, &region, NULL);
 571	if (unlikely(r))
 572		writecache_error(wc, r, "error flushing metadata: %d", r);
 573}
 574
 
 
 
 
 
 
 575#define WFE_RETURN_FOLLOWING	1
 576#define WFE_LOWEST_SEQ		2
 577
 578static struct wc_entry *writecache_find_entry(struct dm_writecache *wc,
 579					      uint64_t block, int flags)
 580{
 581	struct wc_entry *e;
 582	struct rb_node *node = wc->tree.rb_node;
 583
 584	if (unlikely(!node))
 585		return NULL;
 586
 587	while (1) {
 588		e = container_of(node, struct wc_entry, rb_node);
 589		if (read_original_sector(wc, e) == block)
 590			break;
 591
 592		node = (read_original_sector(wc, e) >= block ?
 593			e->rb_node.rb_left : e->rb_node.rb_right);
 594		if (unlikely(!node)) {
 595			if (!(flags & WFE_RETURN_FOLLOWING))
 596				return NULL;
 597			if (read_original_sector(wc, e) >= block) {
 598				return e;
 599			} else {
 600				node = rb_next(&e->rb_node);
 601				if (unlikely(!node))
 602					return NULL;
 603				e = container_of(node, struct wc_entry, rb_node);
 604				return e;
 605			}
 606		}
 607	}
 608
 609	while (1) {
 610		struct wc_entry *e2;
 611		if (flags & WFE_LOWEST_SEQ)
 612			node = rb_prev(&e->rb_node);
 613		else
 614			node = rb_next(&e->rb_node);
 615		if (unlikely(!node))
 616			return e;
 617		e2 = container_of(node, struct wc_entry, rb_node);
 618		if (read_original_sector(wc, e2) != block)
 619			return e;
 620		e = e2;
 621	}
 622}
 623
 624static void writecache_insert_entry(struct dm_writecache *wc, struct wc_entry *ins)
 625{
 626	struct wc_entry *e;
 627	struct rb_node **node = &wc->tree.rb_node, *parent = NULL;
 628
 629	while (*node) {
 630		e = container_of(*node, struct wc_entry, rb_node);
 631		parent = &e->rb_node;
 632		if (read_original_sector(wc, e) > read_original_sector(wc, ins))
 633			node = &parent->rb_left;
 634		else
 635			node = &parent->rb_right;
 636	}
 637	rb_link_node(&ins->rb_node, parent, node);
 638	rb_insert_color(&ins->rb_node, &wc->tree);
 639	list_add(&ins->lru, &wc->lru);
 640	ins->age = jiffies;
 641}
 642
 643static void writecache_unlink(struct dm_writecache *wc, struct wc_entry *e)
 644{
 645	list_del(&e->lru);
 646	rb_erase(&e->rb_node, &wc->tree);
 647}
 648
 649static void writecache_add_to_freelist(struct dm_writecache *wc, struct wc_entry *e)
 650{
 651	if (WC_MODE_SORT_FREELIST(wc)) {
 652		struct rb_node **node = &wc->freetree.rb_node, *parent = NULL;
 653		if (unlikely(!*node))
 654			wc->current_free = e;
 655		while (*node) {
 656			parent = *node;
 657			if (&e->rb_node < *node)
 658				node = &parent->rb_left;
 659			else
 660				node = &parent->rb_right;
 661		}
 662		rb_link_node(&e->rb_node, parent, node);
 663		rb_insert_color(&e->rb_node, &wc->freetree);
 664	} else {
 665		list_add_tail(&e->lru, &wc->freelist);
 666	}
 667	wc->freelist_size++;
 668}
 669
 670static inline void writecache_verify_watermark(struct dm_writecache *wc)
 671{
 672	if (unlikely(wc->freelist_size + wc->writeback_size <= wc->freelist_high_watermark))
 673		queue_work(wc->writeback_wq, &wc->writeback_work);
 674}
 675
 676static void writecache_max_age_timer(struct timer_list *t)
 677{
 678	struct dm_writecache *wc = from_timer(wc, t, max_age_timer);
 679
 680	if (!dm_suspended(wc->ti) && !writecache_has_error(wc)) {
 681		queue_work(wc->writeback_wq, &wc->writeback_work);
 682		mod_timer(&wc->max_age_timer, jiffies + wc->max_age / MAX_AGE_DIV);
 683	}
 684}
 685
 686static struct wc_entry *writecache_pop_from_freelist(struct dm_writecache *wc, sector_t expected_sector)
 687{
 688	struct wc_entry *e;
 689
 690	if (WC_MODE_SORT_FREELIST(wc)) {
 691		struct rb_node *next;
 692		if (unlikely(!wc->current_free))
 693			return NULL;
 694		e = wc->current_free;
 695		if (expected_sector != (sector_t)-1 && unlikely(cache_sector(wc, e) != expected_sector))
 696			return NULL;
 697		next = rb_next(&e->rb_node);
 698		rb_erase(&e->rb_node, &wc->freetree);
 699		if (unlikely(!next))
 700			next = rb_first(&wc->freetree);
 701		wc->current_free = next ? container_of(next, struct wc_entry, rb_node) : NULL;
 702	} else {
 703		if (unlikely(list_empty(&wc->freelist)))
 704			return NULL;
 705		e = container_of(wc->freelist.next, struct wc_entry, lru);
 706		if (expected_sector != (sector_t)-1 && unlikely(cache_sector(wc, e) != expected_sector))
 707			return NULL;
 708		list_del(&e->lru);
 709	}
 710	wc->freelist_size--;
 711
 712	writecache_verify_watermark(wc);
 713
 714	return e;
 715}
 716
 717static void writecache_free_entry(struct dm_writecache *wc, struct wc_entry *e)
 718{
 719	writecache_unlink(wc, e);
 720	writecache_add_to_freelist(wc, e);
 721	clear_seq_count(wc, e);
 722	writecache_flush_region(wc, memory_entry(wc, e), sizeof(struct wc_memory_entry));
 723	if (unlikely(waitqueue_active(&wc->freelist_wait)))
 724		wake_up(&wc->freelist_wait);
 725}
 726
 727static void writecache_wait_on_freelist(struct dm_writecache *wc)
 728{
 729	DEFINE_WAIT(wait);
 730
 731	prepare_to_wait(&wc->freelist_wait, &wait, TASK_UNINTERRUPTIBLE);
 732	wc_unlock(wc);
 733	io_schedule();
 734	finish_wait(&wc->freelist_wait, &wait);
 735	wc_lock(wc);
 736}
 737
 738static void writecache_poison_lists(struct dm_writecache *wc)
 739{
 740	/*
 741	 * Catch incorrect access to these values while the device is suspended.
 742	 */
 743	memset(&wc->tree, -1, sizeof wc->tree);
 744	wc->lru.next = LIST_POISON1;
 745	wc->lru.prev = LIST_POISON2;
 746	wc->freelist.next = LIST_POISON1;
 747	wc->freelist.prev = LIST_POISON2;
 748}
 749
 750static void writecache_flush_entry(struct dm_writecache *wc, struct wc_entry *e)
 751{
 752	writecache_flush_region(wc, memory_entry(wc, e), sizeof(struct wc_memory_entry));
 753	if (WC_MODE_PMEM(wc))
 754		writecache_flush_region(wc, memory_data(wc, e), wc->block_size);
 755}
 756
 757static bool writecache_entry_is_committed(struct dm_writecache *wc, struct wc_entry *e)
 758{
 759	return read_seq_count(wc, e) < wc->seq_count;
 760}
 761
 762static void writecache_flush(struct dm_writecache *wc)
 763{
 764	struct wc_entry *e, *e2;
 765	bool need_flush_after_free;
 766
 767	wc->uncommitted_blocks = 0;
 768	del_timer(&wc->autocommit_timer);
 769
 770	if (list_empty(&wc->lru))
 771		return;
 772
 773	e = container_of(wc->lru.next, struct wc_entry, lru);
 774	if (writecache_entry_is_committed(wc, e)) {
 775		if (wc->overwrote_committed) {
 776			writecache_wait_for_ios(wc, WRITE);
 777			writecache_disk_flush(wc, wc->ssd_dev);
 778			wc->overwrote_committed = false;
 779		}
 780		return;
 781	}
 782	while (1) {
 783		writecache_flush_entry(wc, e);
 784		if (unlikely(e->lru.next == &wc->lru))
 785			break;
 786		e2 = container_of(e->lru.next, struct wc_entry, lru);
 787		if (writecache_entry_is_committed(wc, e2))
 788			break;
 789		e = e2;
 790		cond_resched();
 791	}
 792	writecache_commit_flushed(wc, true);
 
 
 
 793
 794	wc->seq_count++;
 795	pmem_assign(sb(wc)->seq_count, cpu_to_le64(wc->seq_count));
 796	if (WC_MODE_PMEM(wc))
 797		writecache_commit_flushed(wc, false);
 798	else
 799		ssd_commit_superblock(wc);
 800
 801	wc->overwrote_committed = false;
 802
 803	need_flush_after_free = false;
 804	while (1) {
 805		/* Free another committed entry with lower seq-count */
 806		struct rb_node *rb_node = rb_prev(&e->rb_node);
 807
 808		if (rb_node) {
 809			e2 = container_of(rb_node, struct wc_entry, rb_node);
 810			if (read_original_sector(wc, e2) == read_original_sector(wc, e) &&
 811			    likely(!e2->write_in_progress)) {
 812				writecache_free_entry(wc, e2);
 813				need_flush_after_free = true;
 814			}
 815		}
 816		if (unlikely(e->lru.prev == &wc->lru))
 817			break;
 818		e = container_of(e->lru.prev, struct wc_entry, lru);
 819		cond_resched();
 820	}
 821
 822	if (need_flush_after_free)
 823		writecache_commit_flushed(wc, false);
 824}
 825
 826static void writecache_flush_work(struct work_struct *work)
 827{
 828	struct dm_writecache *wc = container_of(work, struct dm_writecache, flush_work);
 829
 830	wc_lock(wc);
 831	writecache_flush(wc);
 832	wc_unlock(wc);
 833}
 834
 835static void writecache_autocommit_timer(struct timer_list *t)
 836{
 837	struct dm_writecache *wc = from_timer(wc, t, autocommit_timer);
 838	if (!writecache_has_error(wc))
 839		queue_work(wc->writeback_wq, &wc->flush_work);
 840}
 841
 842static void writecache_schedule_autocommit(struct dm_writecache *wc)
 843{
 844	if (!timer_pending(&wc->autocommit_timer))
 845		mod_timer(&wc->autocommit_timer, jiffies + wc->autocommit_jiffies);
 846}
 847
 848static void writecache_discard(struct dm_writecache *wc, sector_t start, sector_t end)
 849{
 850	struct wc_entry *e;
 851	bool discarded_something = false;
 852
 853	e = writecache_find_entry(wc, start, WFE_RETURN_FOLLOWING | WFE_LOWEST_SEQ);
 854	if (unlikely(!e))
 855		return;
 856
 857	while (read_original_sector(wc, e) < end) {
 858		struct rb_node *node = rb_next(&e->rb_node);
 859
 860		if (likely(!e->write_in_progress)) {
 861			if (!discarded_something) {
 862				if (!WC_MODE_PMEM(wc)) {
 863					writecache_wait_for_ios(wc, READ);
 864					writecache_wait_for_ios(wc, WRITE);
 865				}
 866				discarded_something = true;
 867			}
 868			if (!writecache_entry_is_committed(wc, e))
 869				wc->uncommitted_blocks--;
 870			writecache_free_entry(wc, e);
 871		}
 872
 873		if (unlikely(!node))
 874			break;
 875
 876		e = container_of(node, struct wc_entry, rb_node);
 877	}
 878
 879	if (discarded_something)
 880		writecache_commit_flushed(wc, false);
 881}
 882
 883static bool writecache_wait_for_writeback(struct dm_writecache *wc)
 884{
 885	if (wc->writeback_size) {
 886		writecache_wait_on_freelist(wc);
 887		return true;
 888	}
 889	return false;
 890}
 891
 892static void writecache_suspend(struct dm_target *ti)
 893{
 894	struct dm_writecache *wc = ti->private;
 895	bool flush_on_suspend;
 896
 897	del_timer_sync(&wc->autocommit_timer);
 898	del_timer_sync(&wc->max_age_timer);
 899
 900	wc_lock(wc);
 901	writecache_flush(wc);
 902	flush_on_suspend = wc->flush_on_suspend;
 903	if (flush_on_suspend) {
 904		wc->flush_on_suspend = false;
 905		wc->writeback_all++;
 906		queue_work(wc->writeback_wq, &wc->writeback_work);
 907	}
 908	wc_unlock(wc);
 909
 910	drain_workqueue(wc->writeback_wq);
 911
 912	wc_lock(wc);
 913	if (flush_on_suspend)
 914		wc->writeback_all--;
 915	while (writecache_wait_for_writeback(wc));
 916
 917	if (WC_MODE_PMEM(wc))
 918		persistent_memory_flush_cache(wc->memory_map, wc->memory_map_size);
 919
 920	writecache_poison_lists(wc);
 921
 922	wc_unlock(wc);
 923}
 924
 925static int writecache_alloc_entries(struct dm_writecache *wc)
 926{
 927	size_t b;
 928
 929	if (wc->entries)
 930		return 0;
 931	wc->entries = vmalloc(array_size(sizeof(struct wc_entry), wc->n_blocks));
 932	if (!wc->entries)
 933		return -ENOMEM;
 934	for (b = 0; b < wc->n_blocks; b++) {
 935		struct wc_entry *e = &wc->entries[b];
 936		e->index = b;
 937		e->write_in_progress = false;
 938		cond_resched();
 939	}
 940
 941	return 0;
 942}
 943
 944static int writecache_read_metadata(struct dm_writecache *wc, sector_t n_sectors)
 945{
 946	struct dm_io_region region;
 947	struct dm_io_request req;
 948
 949	region.bdev = wc->ssd_dev->bdev;
 950	region.sector = wc->start_sector;
 951	region.count = n_sectors;
 952	req.bi_op = REQ_OP_READ;
 953	req.bi_op_flags = REQ_SYNC;
 954	req.mem.type = DM_IO_VMA;
 955	req.mem.ptr.vma = (char *)wc->memory_map;
 956	req.client = wc->dm_io;
 957	req.notify.fn = NULL;
 958
 959	return dm_io(&req, 1, &region, NULL);
 960}
 961
 962static void writecache_resume(struct dm_target *ti)
 963{
 964	struct dm_writecache *wc = ti->private;
 965	size_t b;
 966	bool need_flush = false;
 967	__le64 sb_seq_count;
 968	int r;
 969
 970	wc_lock(wc);
 971
 972	if (WC_MODE_PMEM(wc)) {
 973		persistent_memory_invalidate_cache(wc->memory_map, wc->memory_map_size);
 974	} else {
 975		r = writecache_read_metadata(wc, wc->metadata_sectors);
 976		if (r) {
 977			size_t sb_entries_offset;
 978			writecache_error(wc, r, "unable to read metadata: %d", r);
 979			sb_entries_offset = offsetof(struct wc_memory_superblock, entries);
 980			memset((char *)wc->memory_map + sb_entries_offset, -1,
 981			       (wc->metadata_sectors << SECTOR_SHIFT) - sb_entries_offset);
 982		}
 983	}
 984
 985	wc->tree = RB_ROOT;
 986	INIT_LIST_HEAD(&wc->lru);
 987	if (WC_MODE_SORT_FREELIST(wc)) {
 988		wc->freetree = RB_ROOT;
 989		wc->current_free = NULL;
 990	} else {
 991		INIT_LIST_HEAD(&wc->freelist);
 992	}
 993	wc->freelist_size = 0;
 994
 995	r = memcpy_mcsafe(&sb_seq_count, &sb(wc)->seq_count, sizeof(uint64_t));
 996	if (r) {
 997		writecache_error(wc, r, "hardware memory error when reading superblock: %d", r);
 998		sb_seq_count = cpu_to_le64(0);
 999	}
1000	wc->seq_count = le64_to_cpu(sb_seq_count);
1001
1002#ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
1003	for (b = 0; b < wc->n_blocks; b++) {
1004		struct wc_entry *e = &wc->entries[b];
1005		struct wc_memory_entry wme;
1006		if (writecache_has_error(wc)) {
1007			e->original_sector = -1;
1008			e->seq_count = -1;
1009			continue;
1010		}
1011		r = memcpy_mcsafe(&wme, memory_entry(wc, e), sizeof(struct wc_memory_entry));
1012		if (r) {
1013			writecache_error(wc, r, "hardware memory error when reading metadata entry %lu: %d",
1014					 (unsigned long)b, r);
1015			e->original_sector = -1;
1016			e->seq_count = -1;
1017		} else {
1018			e->original_sector = le64_to_cpu(wme.original_sector);
1019			e->seq_count = le64_to_cpu(wme.seq_count);
1020		}
1021		cond_resched();
1022	}
1023#endif
1024	for (b = 0; b < wc->n_blocks; b++) {
1025		struct wc_entry *e = &wc->entries[b];
1026		if (!writecache_entry_is_committed(wc, e)) {
1027			if (read_seq_count(wc, e) != -1) {
1028erase_this:
1029				clear_seq_count(wc, e);
1030				need_flush = true;
1031			}
1032			writecache_add_to_freelist(wc, e);
1033		} else {
1034			struct wc_entry *old;
1035
1036			old = writecache_find_entry(wc, read_original_sector(wc, e), 0);
1037			if (!old) {
1038				writecache_insert_entry(wc, e);
1039			} else {
1040				if (read_seq_count(wc, old) == read_seq_count(wc, e)) {
1041					writecache_error(wc, -EINVAL,
1042						 "two identical entries, position %llu, sector %llu, sequence %llu",
1043						 (unsigned long long)b, (unsigned long long)read_original_sector(wc, e),
1044						 (unsigned long long)read_seq_count(wc, e));
1045				}
1046				if (read_seq_count(wc, old) > read_seq_count(wc, e)) {
1047					goto erase_this;
1048				} else {
1049					writecache_free_entry(wc, old);
1050					writecache_insert_entry(wc, e);
1051					need_flush = true;
1052				}
1053			}
1054		}
1055		cond_resched();
1056	}
1057
1058	if (need_flush) {
1059		writecache_flush_all_metadata(wc);
1060		writecache_commit_flushed(wc, false);
1061	}
1062
1063	writecache_verify_watermark(wc);
1064
1065	if (wc->max_age != MAX_AGE_UNSPECIFIED)
1066		mod_timer(&wc->max_age_timer, jiffies + wc->max_age / MAX_AGE_DIV);
1067
1068	wc_unlock(wc);
1069}
1070
1071static int process_flush_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1072{
1073	if (argc != 1)
1074		return -EINVAL;
1075
1076	wc_lock(wc);
1077	if (dm_suspended(wc->ti)) {
1078		wc_unlock(wc);
1079		return -EBUSY;
1080	}
1081	if (writecache_has_error(wc)) {
1082		wc_unlock(wc);
1083		return -EIO;
1084	}
1085
1086	writecache_flush(wc);
1087	wc->writeback_all++;
1088	queue_work(wc->writeback_wq, &wc->writeback_work);
1089	wc_unlock(wc);
1090
1091	flush_workqueue(wc->writeback_wq);
1092
1093	wc_lock(wc);
1094	wc->writeback_all--;
1095	if (writecache_has_error(wc)) {
1096		wc_unlock(wc);
1097		return -EIO;
1098	}
1099	wc_unlock(wc);
1100
1101	return 0;
1102}
1103
1104static int process_flush_on_suspend_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1105{
1106	if (argc != 1)
1107		return -EINVAL;
1108
1109	wc_lock(wc);
1110	wc->flush_on_suspend = true;
1111	wc_unlock(wc);
1112
1113	return 0;
1114}
1115
1116static void activate_cleaner(struct dm_writecache *wc)
1117{
1118	wc->flush_on_suspend = true;
1119	wc->cleaner = true;
1120	wc->freelist_high_watermark = wc->n_blocks;
1121	wc->freelist_low_watermark = wc->n_blocks;
1122}
1123
1124static int process_cleaner_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1125{
1126	if (argc != 1)
1127		return -EINVAL;
1128
1129	wc_lock(wc);
1130	activate_cleaner(wc);
1131	if (!dm_suspended(wc->ti))
1132		writecache_verify_watermark(wc);
1133	wc_unlock(wc);
1134
1135	return 0;
1136}
1137
1138static int writecache_message(struct dm_target *ti, unsigned argc, char **argv,
1139			      char *result, unsigned maxlen)
1140{
1141	int r = -EINVAL;
1142	struct dm_writecache *wc = ti->private;
1143
1144	if (!strcasecmp(argv[0], "flush"))
1145		r = process_flush_mesg(argc, argv, wc);
1146	else if (!strcasecmp(argv[0], "flush_on_suspend"))
1147		r = process_flush_on_suspend_mesg(argc, argv, wc);
1148	else if (!strcasecmp(argv[0], "cleaner"))
1149		r = process_cleaner_mesg(argc, argv, wc);
1150	else
1151		DMERR("unrecognised message received: %s", argv[0]);
1152
1153	return r;
1154}
1155
1156static void memcpy_flushcache_optimized(void *dest, void *source, size_t size)
1157{
1158	/*
1159	 * clflushopt performs better with block size 1024, 2048, 4096
1160	 * non-temporal stores perform better with block size 512
1161	 *
1162	 * block size   512             1024            2048            4096
1163	 * movnti       496 MB/s        642 MB/s        725 MB/s        744 MB/s
1164	 * clflushopt   373 MB/s        688 MB/s        1.1 GB/s        1.2 GB/s
1165	 *
1166	 * We see that movnti performs better for 512-byte blocks, and
1167	 * clflushopt performs better for 1024-byte and larger blocks. So, we
1168	 * prefer clflushopt for sizes >= 768.
1169	 *
1170	 * NOTE: this happens to be the case now (with dm-writecache's single
1171	 * threaded model) but re-evaluate this once memcpy_flushcache() is
1172	 * enabled to use movdir64b which might invalidate this performance
1173	 * advantage seen with cache-allocating-writes plus flushing.
1174	 */
1175#ifdef CONFIG_X86
1176	if (static_cpu_has(X86_FEATURE_CLFLUSHOPT) &&
1177	    likely(boot_cpu_data.x86_clflush_size == 64) &&
1178	    likely(size >= 768)) {
1179		do {
1180			memcpy((void *)dest, (void *)source, 64);
1181			clflushopt((void *)dest);
1182			dest += 64;
1183			source += 64;
1184			size -= 64;
1185		} while (size >= 64);
1186		return;
1187	}
1188#endif
1189	memcpy_flushcache(dest, source, size);
1190}
1191
1192static void bio_copy_block(struct dm_writecache *wc, struct bio *bio, void *data)
1193{
1194	void *buf;
1195	unsigned long flags;
1196	unsigned size;
1197	int rw = bio_data_dir(bio);
1198	unsigned remaining_size = wc->block_size;
1199
1200	do {
1201		struct bio_vec bv = bio_iter_iovec(bio, bio->bi_iter);
1202		buf = bvec_kmap_irq(&bv, &flags);
1203		size = bv.bv_len;
1204		if (unlikely(size > remaining_size))
1205			size = remaining_size;
1206
1207		if (rw == READ) {
1208			int r;
1209			r = memcpy_mcsafe(buf, data, size);
1210			flush_dcache_page(bio_page(bio));
1211			if (unlikely(r)) {
1212				writecache_error(wc, r, "hardware memory error when reading data: %d", r);
1213				bio->bi_status = BLK_STS_IOERR;
1214			}
1215		} else {
1216			flush_dcache_page(bio_page(bio));
1217			memcpy_flushcache_optimized(data, buf, size);
1218		}
1219
1220		bvec_kunmap_irq(buf, &flags);
1221
1222		data = (char *)data + size;
1223		remaining_size -= size;
1224		bio_advance(bio, size);
1225	} while (unlikely(remaining_size));
1226}
1227
1228static int writecache_flush_thread(void *data)
1229{
1230	struct dm_writecache *wc = data;
1231
1232	while (1) {
1233		struct bio *bio;
1234
1235		wc_lock(wc);
1236		bio = bio_list_pop(&wc->flush_list);
1237		if (!bio) {
1238			set_current_state(TASK_INTERRUPTIBLE);
1239			wc_unlock(wc);
1240
1241			if (unlikely(kthread_should_stop())) {
1242				set_current_state(TASK_RUNNING);
1243				break;
1244			}
1245
1246			schedule();
1247			continue;
1248		}
1249
1250		if (bio_op(bio) == REQ_OP_DISCARD) {
1251			writecache_discard(wc, bio->bi_iter.bi_sector,
1252					   bio_end_sector(bio));
1253			wc_unlock(wc);
1254			bio_set_dev(bio, wc->dev->bdev);
1255			submit_bio_noacct(bio);
1256		} else {
1257			writecache_flush(wc);
1258			wc_unlock(wc);
1259			if (writecache_has_error(wc))
1260				bio->bi_status = BLK_STS_IOERR;
1261			bio_endio(bio);
1262		}
1263	}
1264
1265	return 0;
1266}
1267
1268static void writecache_offload_bio(struct dm_writecache *wc, struct bio *bio)
1269{
1270	if (bio_list_empty(&wc->flush_list))
1271		wake_up_process(wc->flush_thread);
1272	bio_list_add(&wc->flush_list, bio);
1273}
1274
1275static int writecache_map(struct dm_target *ti, struct bio *bio)
1276{
1277	struct wc_entry *e;
1278	struct dm_writecache *wc = ti->private;
1279
1280	bio->bi_private = NULL;
1281
1282	wc_lock(wc);
1283
1284	if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1285		if (writecache_has_error(wc))
1286			goto unlock_error;
1287		if (WC_MODE_PMEM(wc)) {
1288			writecache_flush(wc);
1289			if (writecache_has_error(wc))
1290				goto unlock_error;
1291			goto unlock_submit;
1292		} else {
1293			writecache_offload_bio(wc, bio);
1294			goto unlock_return;
1295		}
1296	}
1297
1298	bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
1299
1300	if (unlikely((((unsigned)bio->bi_iter.bi_sector | bio_sectors(bio)) &
1301				(wc->block_size / 512 - 1)) != 0)) {
1302		DMERR("I/O is not aligned, sector %llu, size %u, block size %u",
1303		      (unsigned long long)bio->bi_iter.bi_sector,
1304		      bio->bi_iter.bi_size, wc->block_size);
1305		goto unlock_error;
1306	}
1307
1308	if (unlikely(bio_op(bio) == REQ_OP_DISCARD)) {
1309		if (writecache_has_error(wc))
1310			goto unlock_error;
1311		if (WC_MODE_PMEM(wc)) {
1312			writecache_discard(wc, bio->bi_iter.bi_sector, bio_end_sector(bio));
1313			goto unlock_remap_origin;
1314		} else {
1315			writecache_offload_bio(wc, bio);
1316			goto unlock_return;
1317		}
1318	}
1319
1320	if (bio_data_dir(bio) == READ) {
1321read_next_block:
1322		e = writecache_find_entry(wc, bio->bi_iter.bi_sector, WFE_RETURN_FOLLOWING);
1323		if (e && read_original_sector(wc, e) == bio->bi_iter.bi_sector) {
1324			if (WC_MODE_PMEM(wc)) {
1325				bio_copy_block(wc, bio, memory_data(wc, e));
1326				if (bio->bi_iter.bi_size)
1327					goto read_next_block;
1328				goto unlock_submit;
1329			} else {
1330				dm_accept_partial_bio(bio, wc->block_size >> SECTOR_SHIFT);
1331				bio_set_dev(bio, wc->ssd_dev->bdev);
1332				bio->bi_iter.bi_sector = cache_sector(wc, e);
1333				if (!writecache_entry_is_committed(wc, e))
1334					writecache_wait_for_ios(wc, WRITE);
1335				goto unlock_remap;
1336			}
1337		} else {
1338			if (e) {
1339				sector_t next_boundary =
1340					read_original_sector(wc, e) - bio->bi_iter.bi_sector;
1341				if (next_boundary < bio->bi_iter.bi_size >> SECTOR_SHIFT) {
1342					dm_accept_partial_bio(bio, next_boundary);
1343				}
1344			}
1345			goto unlock_remap_origin;
1346		}
1347	} else {
1348		do {
1349			bool found_entry = false;
1350			if (writecache_has_error(wc))
1351				goto unlock_error;
1352			e = writecache_find_entry(wc, bio->bi_iter.bi_sector, 0);
1353			if (e) {
1354				if (!writecache_entry_is_committed(wc, e))
1355					goto bio_copy;
1356				if (!WC_MODE_PMEM(wc) && !e->write_in_progress) {
1357					wc->overwrote_committed = true;
1358					goto bio_copy;
1359				}
1360				found_entry = true;
1361			} else {
1362				if (unlikely(wc->cleaner))
1363					goto direct_write;
1364			}
1365			e = writecache_pop_from_freelist(wc, (sector_t)-1);
1366			if (unlikely(!e)) {
1367				if (!found_entry) {
1368direct_write:
1369					e = writecache_find_entry(wc, bio->bi_iter.bi_sector, WFE_RETURN_FOLLOWING);
1370					if (e) {
1371						sector_t next_boundary = read_original_sector(wc, e) - bio->bi_iter.bi_sector;
1372						BUG_ON(!next_boundary);
1373						if (next_boundary < bio->bi_iter.bi_size >> SECTOR_SHIFT) {
1374							dm_accept_partial_bio(bio, next_boundary);
1375						}
1376					}
1377					goto unlock_remap_origin;
1378				}
1379				writecache_wait_on_freelist(wc);
1380				continue;
1381			}
1382			write_original_sector_seq_count(wc, e, bio->bi_iter.bi_sector, wc->seq_count);
1383			writecache_insert_entry(wc, e);
1384			wc->uncommitted_blocks++;
1385bio_copy:
1386			if (WC_MODE_PMEM(wc)) {
1387				bio_copy_block(wc, bio, memory_data(wc, e));
1388			} else {
1389				unsigned bio_size = wc->block_size;
1390				sector_t start_cache_sec = cache_sector(wc, e);
1391				sector_t current_cache_sec = start_cache_sec + (bio_size >> SECTOR_SHIFT);
1392
1393				while (bio_size < bio->bi_iter.bi_size) {
1394					struct wc_entry *f = writecache_pop_from_freelist(wc, current_cache_sec);
1395					if (!f)
1396						break;
1397					write_original_sector_seq_count(wc, f, bio->bi_iter.bi_sector +
1398									(bio_size >> SECTOR_SHIFT), wc->seq_count);
1399					writecache_insert_entry(wc, f);
1400					wc->uncommitted_blocks++;
1401					bio_size += wc->block_size;
1402					current_cache_sec += wc->block_size >> SECTOR_SHIFT;
1403				}
1404
1405				bio_set_dev(bio, wc->ssd_dev->bdev);
1406				bio->bi_iter.bi_sector = start_cache_sec;
1407				dm_accept_partial_bio(bio, bio_size >> SECTOR_SHIFT);
1408
1409				if (unlikely(wc->uncommitted_blocks >= wc->autocommit_blocks)) {
1410					wc->uncommitted_blocks = 0;
1411					queue_work(wc->writeback_wq, &wc->flush_work);
1412				} else {
1413					writecache_schedule_autocommit(wc);
1414				}
1415				goto unlock_remap;
1416			}
1417		} while (bio->bi_iter.bi_size);
1418
1419		if (unlikely(bio->bi_opf & REQ_FUA ||
1420			     wc->uncommitted_blocks >= wc->autocommit_blocks))
1421			writecache_flush(wc);
1422		else
1423			writecache_schedule_autocommit(wc);
1424		goto unlock_submit;
1425	}
1426
1427unlock_remap_origin:
1428	bio_set_dev(bio, wc->dev->bdev);
1429	wc_unlock(wc);
1430	return DM_MAPIO_REMAPPED;
1431
1432unlock_remap:
1433	/* make sure that writecache_end_io decrements bio_in_progress: */
1434	bio->bi_private = (void *)1;
1435	atomic_inc(&wc->bio_in_progress[bio_data_dir(bio)]);
1436	wc_unlock(wc);
1437	return DM_MAPIO_REMAPPED;
1438
1439unlock_submit:
1440	wc_unlock(wc);
1441	bio_endio(bio);
1442	return DM_MAPIO_SUBMITTED;
1443
1444unlock_return:
1445	wc_unlock(wc);
1446	return DM_MAPIO_SUBMITTED;
1447
1448unlock_error:
1449	wc_unlock(wc);
1450	bio_io_error(bio);
1451	return DM_MAPIO_SUBMITTED;
1452}
1453
1454static int writecache_end_io(struct dm_target *ti, struct bio *bio, blk_status_t *status)
1455{
1456	struct dm_writecache *wc = ti->private;
1457
1458	if (bio->bi_private != NULL) {
1459		int dir = bio_data_dir(bio);
1460		if (atomic_dec_and_test(&wc->bio_in_progress[dir]))
1461			if (unlikely(waitqueue_active(&wc->bio_in_progress_wait[dir])))
1462				wake_up(&wc->bio_in_progress_wait[dir]);
1463	}
1464	return 0;
1465}
1466
1467static int writecache_iterate_devices(struct dm_target *ti,
1468				      iterate_devices_callout_fn fn, void *data)
1469{
1470	struct dm_writecache *wc = ti->private;
1471
1472	return fn(ti, wc->dev, 0, ti->len, data);
1473}
1474
1475static void writecache_io_hints(struct dm_target *ti, struct queue_limits *limits)
1476{
1477	struct dm_writecache *wc = ti->private;
1478
1479	if (limits->logical_block_size < wc->block_size)
1480		limits->logical_block_size = wc->block_size;
1481
1482	if (limits->physical_block_size < wc->block_size)
1483		limits->physical_block_size = wc->block_size;
1484
1485	if (limits->io_min < wc->block_size)
1486		limits->io_min = wc->block_size;
1487}
1488
1489
1490static void writecache_writeback_endio(struct bio *bio)
1491{
1492	struct writeback_struct *wb = container_of(bio, struct writeback_struct, bio);
1493	struct dm_writecache *wc = wb->wc;
1494	unsigned long flags;
1495
1496	raw_spin_lock_irqsave(&wc->endio_list_lock, flags);
1497	if (unlikely(list_empty(&wc->endio_list)))
1498		wake_up_process(wc->endio_thread);
1499	list_add_tail(&wb->endio_entry, &wc->endio_list);
1500	raw_spin_unlock_irqrestore(&wc->endio_list_lock, flags);
1501}
1502
1503static void writecache_copy_endio(int read_err, unsigned long write_err, void *ptr)
1504{
1505	struct copy_struct *c = ptr;
1506	struct dm_writecache *wc = c->wc;
1507
1508	c->error = likely(!(read_err | write_err)) ? 0 : -EIO;
1509
1510	raw_spin_lock_irq(&wc->endio_list_lock);
1511	if (unlikely(list_empty(&wc->endio_list)))
1512		wake_up_process(wc->endio_thread);
1513	list_add_tail(&c->endio_entry, &wc->endio_list);
1514	raw_spin_unlock_irq(&wc->endio_list_lock);
1515}
1516
1517static void __writecache_endio_pmem(struct dm_writecache *wc, struct list_head *list)
1518{
1519	unsigned i;
1520	struct writeback_struct *wb;
1521	struct wc_entry *e;
1522	unsigned long n_walked = 0;
1523
1524	do {
1525		wb = list_entry(list->next, struct writeback_struct, endio_entry);
1526		list_del(&wb->endio_entry);
1527
1528		if (unlikely(wb->bio.bi_status != BLK_STS_OK))
1529			writecache_error(wc, blk_status_to_errno(wb->bio.bi_status),
1530					"write error %d", wb->bio.bi_status);
1531		i = 0;
1532		do {
1533			e = wb->wc_list[i];
1534			BUG_ON(!e->write_in_progress);
1535			e->write_in_progress = false;
1536			INIT_LIST_HEAD(&e->lru);
1537			if (!writecache_has_error(wc))
1538				writecache_free_entry(wc, e);
1539			BUG_ON(!wc->writeback_size);
1540			wc->writeback_size--;
1541			n_walked++;
1542			if (unlikely(n_walked >= ENDIO_LATENCY)) {
1543				writecache_commit_flushed(wc, false);
1544				wc_unlock(wc);
1545				wc_lock(wc);
1546				n_walked = 0;
1547			}
1548		} while (++i < wb->wc_list_n);
1549
1550		if (wb->wc_list != wb->wc_list_inline)
1551			kfree(wb->wc_list);
1552		bio_put(&wb->bio);
1553	} while (!list_empty(list));
1554}
1555
1556static void __writecache_endio_ssd(struct dm_writecache *wc, struct list_head *list)
1557{
1558	struct copy_struct *c;
1559	struct wc_entry *e;
1560
1561	do {
1562		c = list_entry(list->next, struct copy_struct, endio_entry);
1563		list_del(&c->endio_entry);
1564
1565		if (unlikely(c->error))
1566			writecache_error(wc, c->error, "copy error");
1567
1568		e = c->e;
1569		do {
1570			BUG_ON(!e->write_in_progress);
1571			e->write_in_progress = false;
1572			INIT_LIST_HEAD(&e->lru);
1573			if (!writecache_has_error(wc))
1574				writecache_free_entry(wc, e);
1575
1576			BUG_ON(!wc->writeback_size);
1577			wc->writeback_size--;
1578			e++;
1579		} while (--c->n_entries);
1580		mempool_free(c, &wc->copy_pool);
1581	} while (!list_empty(list));
1582}
1583
1584static int writecache_endio_thread(void *data)
1585{
1586	struct dm_writecache *wc = data;
1587
1588	while (1) {
1589		struct list_head list;
1590
1591		raw_spin_lock_irq(&wc->endio_list_lock);
1592		if (!list_empty(&wc->endio_list))
1593			goto pop_from_list;
1594		set_current_state(TASK_INTERRUPTIBLE);
1595		raw_spin_unlock_irq(&wc->endio_list_lock);
1596
1597		if (unlikely(kthread_should_stop())) {
1598			set_current_state(TASK_RUNNING);
1599			break;
1600		}
1601
1602		schedule();
1603
1604		continue;
1605
1606pop_from_list:
1607		list = wc->endio_list;
1608		list.next->prev = list.prev->next = &list;
1609		INIT_LIST_HEAD(&wc->endio_list);
1610		raw_spin_unlock_irq(&wc->endio_list_lock);
1611
1612		if (!WC_MODE_FUA(wc))
1613			writecache_disk_flush(wc, wc->dev);
1614
1615		wc_lock(wc);
1616
1617		if (WC_MODE_PMEM(wc)) {
1618			__writecache_endio_pmem(wc, &list);
1619		} else {
1620			__writecache_endio_ssd(wc, &list);
1621			writecache_wait_for_ios(wc, READ);
1622		}
1623
1624		writecache_commit_flushed(wc, false);
1625
1626		wc_unlock(wc);
1627	}
1628
1629	return 0;
1630}
1631
1632static bool wc_add_block(struct writeback_struct *wb, struct wc_entry *e, gfp_t gfp)
1633{
1634	struct dm_writecache *wc = wb->wc;
1635	unsigned block_size = wc->block_size;
1636	void *address = memory_data(wc, e);
1637
1638	persistent_memory_flush_cache(address, block_size);
1639	return bio_add_page(&wb->bio, persistent_memory_page(address),
1640			    block_size, persistent_memory_page_offset(address)) != 0;
1641}
1642
1643struct writeback_list {
1644	struct list_head list;
1645	size_t size;
1646};
1647
1648static void __writeback_throttle(struct dm_writecache *wc, struct writeback_list *wbl)
1649{
1650	if (unlikely(wc->max_writeback_jobs)) {
1651		if (READ_ONCE(wc->writeback_size) - wbl->size >= wc->max_writeback_jobs) {
1652			wc_lock(wc);
1653			while (wc->writeback_size - wbl->size >= wc->max_writeback_jobs)
1654				writecache_wait_on_freelist(wc);
1655			wc_unlock(wc);
1656		}
1657	}
1658	cond_resched();
1659}
1660
1661static void __writecache_writeback_pmem(struct dm_writecache *wc, struct writeback_list *wbl)
1662{
1663	struct wc_entry *e, *f;
1664	struct bio *bio;
1665	struct writeback_struct *wb;
1666	unsigned max_pages;
1667
1668	while (wbl->size) {
1669		wbl->size--;
1670		e = container_of(wbl->list.prev, struct wc_entry, lru);
1671		list_del(&e->lru);
1672
1673		max_pages = e->wc_list_contiguous;
1674
1675		bio = bio_alloc_bioset(GFP_NOIO, max_pages, &wc->bio_set);
1676		wb = container_of(bio, struct writeback_struct, bio);
1677		wb->wc = wc;
1678		bio->bi_end_io = writecache_writeback_endio;
1679		bio_set_dev(bio, wc->dev->bdev);
1680		bio->bi_iter.bi_sector = read_original_sector(wc, e);
1681		if (max_pages <= WB_LIST_INLINE ||
1682		    unlikely(!(wb->wc_list = kmalloc_array(max_pages, sizeof(struct wc_entry *),
1683							   GFP_NOIO | __GFP_NORETRY |
1684							   __GFP_NOMEMALLOC | __GFP_NOWARN)))) {
1685			wb->wc_list = wb->wc_list_inline;
1686			max_pages = WB_LIST_INLINE;
1687		}
1688
1689		BUG_ON(!wc_add_block(wb, e, GFP_NOIO));
1690
1691		wb->wc_list[0] = e;
1692		wb->wc_list_n = 1;
1693
1694		while (wbl->size && wb->wc_list_n < max_pages) {
1695			f = container_of(wbl->list.prev, struct wc_entry, lru);
1696			if (read_original_sector(wc, f) !=
1697			    read_original_sector(wc, e) + (wc->block_size >> SECTOR_SHIFT))
1698				break;
1699			if (!wc_add_block(wb, f, GFP_NOWAIT | __GFP_NOWARN))
1700				break;
1701			wbl->size--;
1702			list_del(&f->lru);
1703			wb->wc_list[wb->wc_list_n++] = f;
1704			e = f;
1705		}
1706		bio_set_op_attrs(bio, REQ_OP_WRITE, WC_MODE_FUA(wc) * REQ_FUA);
1707		if (writecache_has_error(wc)) {
1708			bio->bi_status = BLK_STS_IOERR;
1709			bio_endio(bio);
1710		} else {
1711			submit_bio(bio);
1712		}
1713
1714		__writeback_throttle(wc, wbl);
1715	}
1716}
1717
1718static void __writecache_writeback_ssd(struct dm_writecache *wc, struct writeback_list *wbl)
1719{
1720	struct wc_entry *e, *f;
1721	struct dm_io_region from, to;
1722	struct copy_struct *c;
1723
1724	while (wbl->size) {
1725		unsigned n_sectors;
1726
1727		wbl->size--;
1728		e = container_of(wbl->list.prev, struct wc_entry, lru);
1729		list_del(&e->lru);
1730
1731		n_sectors = e->wc_list_contiguous << (wc->block_size_bits - SECTOR_SHIFT);
1732
1733		from.bdev = wc->ssd_dev->bdev;
1734		from.sector = cache_sector(wc, e);
1735		from.count = n_sectors;
1736		to.bdev = wc->dev->bdev;
1737		to.sector = read_original_sector(wc, e);
1738		to.count = n_sectors;
1739
1740		c = mempool_alloc(&wc->copy_pool, GFP_NOIO);
1741		c->wc = wc;
1742		c->e = e;
1743		c->n_entries = e->wc_list_contiguous;
1744
1745		while ((n_sectors -= wc->block_size >> SECTOR_SHIFT)) {
1746			wbl->size--;
1747			f = container_of(wbl->list.prev, struct wc_entry, lru);
1748			BUG_ON(f != e + 1);
1749			list_del(&f->lru);
1750			e = f;
1751		}
1752
1753		dm_kcopyd_copy(wc->dm_kcopyd, &from, 1, &to, 0, writecache_copy_endio, c);
1754
1755		__writeback_throttle(wc, wbl);
1756	}
1757}
1758
1759static void writecache_writeback(struct work_struct *work)
1760{
1761	struct dm_writecache *wc = container_of(work, struct dm_writecache, writeback_work);
1762	struct blk_plug plug;
1763	struct wc_entry *f, *g, *e = NULL;
1764	struct rb_node *node, *next_node;
1765	struct list_head skipped;
1766	struct writeback_list wbl;
1767	unsigned long n_walked;
1768
1769	wc_lock(wc);
1770restart:
1771	if (writecache_has_error(wc)) {
1772		wc_unlock(wc);
1773		return;
1774	}
1775
1776	if (unlikely(wc->writeback_all)) {
1777		if (writecache_wait_for_writeback(wc))
1778			goto restart;
1779	}
1780
1781	if (wc->overwrote_committed) {
1782		writecache_wait_for_ios(wc, WRITE);
1783	}
1784
1785	n_walked = 0;
1786	INIT_LIST_HEAD(&skipped);
1787	INIT_LIST_HEAD(&wbl.list);
1788	wbl.size = 0;
1789	while (!list_empty(&wc->lru) &&
1790	       (wc->writeback_all ||
1791		wc->freelist_size + wc->writeback_size <= wc->freelist_low_watermark ||
1792		(jiffies - container_of(wc->lru.prev, struct wc_entry, lru)->age >=
1793		 wc->max_age - wc->max_age / MAX_AGE_DIV))) {
1794
1795		n_walked++;
1796		if (unlikely(n_walked > WRITEBACK_LATENCY) &&
1797		    likely(!wc->writeback_all) && likely(!dm_suspended(wc->ti))) {
1798			queue_work(wc->writeback_wq, &wc->writeback_work);
1799			break;
1800		}
1801
1802		if (unlikely(wc->writeback_all)) {
1803			if (unlikely(!e)) {
1804				writecache_flush(wc);
1805				e = container_of(rb_first(&wc->tree), struct wc_entry, rb_node);
1806			} else
1807				e = g;
1808		} else
1809			e = container_of(wc->lru.prev, struct wc_entry, lru);
1810		BUG_ON(e->write_in_progress);
1811		if (unlikely(!writecache_entry_is_committed(wc, e))) {
1812			writecache_flush(wc);
1813		}
1814		node = rb_prev(&e->rb_node);
1815		if (node) {
1816			f = container_of(node, struct wc_entry, rb_node);
1817			if (unlikely(read_original_sector(wc, f) ==
1818				     read_original_sector(wc, e))) {
1819				BUG_ON(!f->write_in_progress);
1820				list_del(&e->lru);
1821				list_add(&e->lru, &skipped);
1822				cond_resched();
1823				continue;
1824			}
1825		}
1826		wc->writeback_size++;
1827		list_del(&e->lru);
1828		list_add(&e->lru, &wbl.list);
1829		wbl.size++;
1830		e->write_in_progress = true;
1831		e->wc_list_contiguous = 1;
1832
1833		f = e;
1834
1835		while (1) {
1836			next_node = rb_next(&f->rb_node);
1837			if (unlikely(!next_node))
1838				break;
1839			g = container_of(next_node, struct wc_entry, rb_node);
1840			if (unlikely(read_original_sector(wc, g) ==
1841			    read_original_sector(wc, f))) {
1842				f = g;
1843				continue;
1844			}
1845			if (read_original_sector(wc, g) !=
1846			    read_original_sector(wc, f) + (wc->block_size >> SECTOR_SHIFT))
1847				break;
1848			if (unlikely(g->write_in_progress))
1849				break;
1850			if (unlikely(!writecache_entry_is_committed(wc, g)))
1851				break;
1852
1853			if (!WC_MODE_PMEM(wc)) {
1854				if (g != f + 1)
1855					break;
1856			}
1857
1858			n_walked++;
1859			//if (unlikely(n_walked > WRITEBACK_LATENCY) && likely(!wc->writeback_all))
1860			//	break;
1861
1862			wc->writeback_size++;
1863			list_del(&g->lru);
1864			list_add(&g->lru, &wbl.list);
1865			wbl.size++;
1866			g->write_in_progress = true;
1867			g->wc_list_contiguous = BIO_MAX_PAGES;
1868			f = g;
1869			e->wc_list_contiguous++;
1870			if (unlikely(e->wc_list_contiguous == BIO_MAX_PAGES)) {
1871				if (unlikely(wc->writeback_all)) {
1872					next_node = rb_next(&f->rb_node);
1873					if (likely(next_node))
1874						g = container_of(next_node, struct wc_entry, rb_node);
1875				}
1876				break;
1877			}
1878		}
1879		cond_resched();
1880	}
1881
1882	if (!list_empty(&skipped)) {
1883		list_splice_tail(&skipped, &wc->lru);
1884		/*
1885		 * If we didn't do any progress, we must wait until some
1886		 * writeback finishes to avoid burning CPU in a loop
1887		 */
1888		if (unlikely(!wbl.size))
1889			writecache_wait_for_writeback(wc);
1890	}
1891
1892	wc_unlock(wc);
1893
1894	blk_start_plug(&plug);
1895
1896	if (WC_MODE_PMEM(wc))
1897		__writecache_writeback_pmem(wc, &wbl);
1898	else
1899		__writecache_writeback_ssd(wc, &wbl);
1900
1901	blk_finish_plug(&plug);
1902
1903	if (unlikely(wc->writeback_all)) {
1904		wc_lock(wc);
1905		while (writecache_wait_for_writeback(wc));
1906		wc_unlock(wc);
1907	}
1908}
1909
1910static int calculate_memory_size(uint64_t device_size, unsigned block_size,
1911				 size_t *n_blocks_p, size_t *n_metadata_blocks_p)
1912{
1913	uint64_t n_blocks, offset;
1914	struct wc_entry e;
1915
1916	n_blocks = device_size;
1917	do_div(n_blocks, block_size + sizeof(struct wc_memory_entry));
1918
1919	while (1) {
1920		if (!n_blocks)
1921			return -ENOSPC;
1922		/* Verify the following entries[n_blocks] won't overflow */
1923		if (n_blocks >= ((size_t)-sizeof(struct wc_memory_superblock) /
1924				 sizeof(struct wc_memory_entry)))
1925			return -EFBIG;
1926		offset = offsetof(struct wc_memory_superblock, entries[n_blocks]);
1927		offset = (offset + block_size - 1) & ~(uint64_t)(block_size - 1);
1928		if (offset + n_blocks * block_size <= device_size)
1929			break;
1930		n_blocks--;
1931	}
1932
1933	/* check if the bit field overflows */
1934	e.index = n_blocks;
1935	if (e.index != n_blocks)
1936		return -EFBIG;
1937
1938	if (n_blocks_p)
1939		*n_blocks_p = n_blocks;
1940	if (n_metadata_blocks_p)
1941		*n_metadata_blocks_p = offset >> __ffs(block_size);
1942	return 0;
1943}
1944
1945static int init_memory(struct dm_writecache *wc)
1946{
1947	size_t b;
1948	int r;
1949
1950	r = calculate_memory_size(wc->memory_map_size, wc->block_size, &wc->n_blocks, NULL);
1951	if (r)
1952		return r;
1953
1954	r = writecache_alloc_entries(wc);
1955	if (r)
1956		return r;
1957
1958	for (b = 0; b < ARRAY_SIZE(sb(wc)->padding); b++)
1959		pmem_assign(sb(wc)->padding[b], cpu_to_le64(0));
1960	pmem_assign(sb(wc)->version, cpu_to_le32(MEMORY_SUPERBLOCK_VERSION));
1961	pmem_assign(sb(wc)->block_size, cpu_to_le32(wc->block_size));
1962	pmem_assign(sb(wc)->n_blocks, cpu_to_le64(wc->n_blocks));
1963	pmem_assign(sb(wc)->seq_count, cpu_to_le64(0));
1964
1965	for (b = 0; b < wc->n_blocks; b++) {
1966		write_original_sector_seq_count(wc, &wc->entries[b], -1, -1);
1967		cond_resched();
1968	}
1969
1970	writecache_flush_all_metadata(wc);
1971	writecache_commit_flushed(wc, false);
1972	pmem_assign(sb(wc)->magic, cpu_to_le32(MEMORY_SUPERBLOCK_MAGIC));
1973	writecache_flush_region(wc, &sb(wc)->magic, sizeof sb(wc)->magic);
1974	writecache_commit_flushed(wc, false);
1975
1976	return 0;
1977}
1978
1979static void writecache_dtr(struct dm_target *ti)
1980{
1981	struct dm_writecache *wc = ti->private;
1982
1983	if (!wc)
1984		return;
1985
1986	if (wc->endio_thread)
1987		kthread_stop(wc->endio_thread);
1988
1989	if (wc->flush_thread)
1990		kthread_stop(wc->flush_thread);
1991
1992	bioset_exit(&wc->bio_set);
1993
1994	mempool_exit(&wc->copy_pool);
1995
1996	if (wc->writeback_wq)
1997		destroy_workqueue(wc->writeback_wq);
1998
1999	if (wc->dev)
2000		dm_put_device(ti, wc->dev);
2001
2002	if (wc->ssd_dev)
2003		dm_put_device(ti, wc->ssd_dev);
2004
2005	if (wc->entries)
2006		vfree(wc->entries);
2007
2008	if (wc->memory_map) {
2009		if (WC_MODE_PMEM(wc))
2010			persistent_memory_release(wc);
2011		else
2012			vfree(wc->memory_map);
2013	}
2014
2015	if (wc->dm_kcopyd)
2016		dm_kcopyd_client_destroy(wc->dm_kcopyd);
2017
2018	if (wc->dm_io)
2019		dm_io_client_destroy(wc->dm_io);
2020
2021	if (wc->dirty_bitmap)
2022		vfree(wc->dirty_bitmap);
2023
2024	kfree(wc);
2025}
2026
2027static int writecache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2028{
2029	struct dm_writecache *wc;
2030	struct dm_arg_set as;
2031	const char *string;
2032	unsigned opt_params;
2033	size_t offset, data_size;
2034	int i, r;
2035	char dummy;
2036	int high_wm_percent = HIGH_WATERMARK;
2037	int low_wm_percent = LOW_WATERMARK;
2038	uint64_t x;
2039	struct wc_memory_superblock s;
2040
2041	static struct dm_arg _args[] = {
2042		{0, 10, "Invalid number of feature args"},
2043	};
2044
2045	as.argc = argc;
2046	as.argv = argv;
2047
2048	wc = kzalloc(sizeof(struct dm_writecache), GFP_KERNEL);
2049	if (!wc) {
2050		ti->error = "Cannot allocate writecache structure";
2051		r = -ENOMEM;
2052		goto bad;
2053	}
2054	ti->private = wc;
2055	wc->ti = ti;
2056
2057	mutex_init(&wc->lock);
2058	wc->max_age = MAX_AGE_UNSPECIFIED;
2059	writecache_poison_lists(wc);
2060	init_waitqueue_head(&wc->freelist_wait);
2061	timer_setup(&wc->autocommit_timer, writecache_autocommit_timer, 0);
2062	timer_setup(&wc->max_age_timer, writecache_max_age_timer, 0);
2063
2064	for (i = 0; i < 2; i++) {
2065		atomic_set(&wc->bio_in_progress[i], 0);
2066		init_waitqueue_head(&wc->bio_in_progress_wait[i]);
2067	}
2068
2069	wc->dm_io = dm_io_client_create();
2070	if (IS_ERR(wc->dm_io)) {
2071		r = PTR_ERR(wc->dm_io);
2072		ti->error = "Unable to allocate dm-io client";
2073		wc->dm_io = NULL;
2074		goto bad;
2075	}
2076
2077	wc->writeback_wq = alloc_workqueue("writecache-writeback", WQ_MEM_RECLAIM, 1);
2078	if (!wc->writeback_wq) {
2079		r = -ENOMEM;
2080		ti->error = "Could not allocate writeback workqueue";
2081		goto bad;
2082	}
2083	INIT_WORK(&wc->writeback_work, writecache_writeback);
2084	INIT_WORK(&wc->flush_work, writecache_flush_work);
2085
2086	raw_spin_lock_init(&wc->endio_list_lock);
2087	INIT_LIST_HEAD(&wc->endio_list);
2088	wc->endio_thread = kthread_create(writecache_endio_thread, wc, "writecache_endio");
2089	if (IS_ERR(wc->endio_thread)) {
2090		r = PTR_ERR(wc->endio_thread);
2091		wc->endio_thread = NULL;
2092		ti->error = "Couldn't spawn endio thread";
2093		goto bad;
2094	}
2095	wake_up_process(wc->endio_thread);
2096
2097	/*
2098	 * Parse the mode (pmem or ssd)
2099	 */
2100	string = dm_shift_arg(&as);
2101	if (!string)
2102		goto bad_arguments;
2103
2104	if (!strcasecmp(string, "s")) {
2105		wc->pmem_mode = false;
2106	} else if (!strcasecmp(string, "p")) {
2107#ifdef DM_WRITECACHE_HAS_PMEM
2108		wc->pmem_mode = true;
2109		wc->writeback_fua = true;
2110#else
2111		/*
2112		 * If the architecture doesn't support persistent memory or
2113		 * the kernel doesn't support any DAX drivers, this driver can
2114		 * only be used in SSD-only mode.
2115		 */
2116		r = -EOPNOTSUPP;
2117		ti->error = "Persistent memory or DAX not supported on this system";
2118		goto bad;
2119#endif
2120	} else {
2121		goto bad_arguments;
2122	}
2123
2124	if (WC_MODE_PMEM(wc)) {
2125		r = bioset_init(&wc->bio_set, BIO_POOL_SIZE,
2126				offsetof(struct writeback_struct, bio),
2127				BIOSET_NEED_BVECS);
2128		if (r) {
2129			ti->error = "Could not allocate bio set";
2130			goto bad;
2131		}
2132	} else {
2133		r = mempool_init_kmalloc_pool(&wc->copy_pool, 1, sizeof(struct copy_struct));
2134		if (r) {
2135			ti->error = "Could not allocate mempool";
2136			goto bad;
2137		}
2138	}
2139
2140	/*
2141	 * Parse the origin data device
2142	 */
2143	string = dm_shift_arg(&as);
2144	if (!string)
2145		goto bad_arguments;
2146	r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->dev);
2147	if (r) {
2148		ti->error = "Origin data device lookup failed";
2149		goto bad;
2150	}
2151
2152	/*
2153	 * Parse cache data device (be it pmem or ssd)
2154	 */
2155	string = dm_shift_arg(&as);
2156	if (!string)
2157		goto bad_arguments;
2158
2159	r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->ssd_dev);
2160	if (r) {
2161		ti->error = "Cache data device lookup failed";
2162		goto bad;
2163	}
2164	wc->memory_map_size = i_size_read(wc->ssd_dev->bdev->bd_inode);
2165
2166	/*
2167	 * Parse the cache block size
2168	 */
2169	string = dm_shift_arg(&as);
2170	if (!string)
2171		goto bad_arguments;
2172	if (sscanf(string, "%u%c", &wc->block_size, &dummy) != 1 ||
2173	    wc->block_size < 512 || wc->block_size > PAGE_SIZE ||
2174	    (wc->block_size & (wc->block_size - 1))) {
2175		r = -EINVAL;
2176		ti->error = "Invalid block size";
2177		goto bad;
2178	}
2179	if (wc->block_size < bdev_logical_block_size(wc->dev->bdev) ||
2180	    wc->block_size < bdev_logical_block_size(wc->ssd_dev->bdev)) {
2181		r = -EINVAL;
2182		ti->error = "Block size is smaller than device logical block size";
2183		goto bad;
2184	}
2185	wc->block_size_bits = __ffs(wc->block_size);
2186
2187	wc->max_writeback_jobs = MAX_WRITEBACK_JOBS;
2188	wc->autocommit_blocks = !WC_MODE_PMEM(wc) ? AUTOCOMMIT_BLOCKS_SSD : AUTOCOMMIT_BLOCKS_PMEM;
2189	wc->autocommit_jiffies = msecs_to_jiffies(AUTOCOMMIT_MSEC);
2190
2191	/*
2192	 * Parse optional arguments
2193	 */
2194	r = dm_read_arg_group(_args, &as, &opt_params, &ti->error);
2195	if (r)
2196		goto bad;
2197
2198	while (opt_params) {
2199		string = dm_shift_arg(&as), opt_params--;
2200		if (!strcasecmp(string, "start_sector") && opt_params >= 1) {
2201			unsigned long long start_sector;
2202			string = dm_shift_arg(&as), opt_params--;
2203			if (sscanf(string, "%llu%c", &start_sector, &dummy) != 1)
2204				goto invalid_optional;
2205			wc->start_sector = start_sector;
2206			if (wc->start_sector != start_sector ||
2207			    wc->start_sector >= wc->memory_map_size >> SECTOR_SHIFT)
2208				goto invalid_optional;
2209		} else if (!strcasecmp(string, "high_watermark") && opt_params >= 1) {
2210			string = dm_shift_arg(&as), opt_params--;
2211			if (sscanf(string, "%d%c", &high_wm_percent, &dummy) != 1)
2212				goto invalid_optional;
2213			if (high_wm_percent < 0 || high_wm_percent > 100)
2214				goto invalid_optional;
2215			wc->high_wm_percent_set = true;
2216		} else if (!strcasecmp(string, "low_watermark") && opt_params >= 1) {
2217			string = dm_shift_arg(&as), opt_params--;
2218			if (sscanf(string, "%d%c", &low_wm_percent, &dummy) != 1)
2219				goto invalid_optional;
2220			if (low_wm_percent < 0 || low_wm_percent > 100)
2221				goto invalid_optional;
2222			wc->low_wm_percent_set = true;
2223		} else if (!strcasecmp(string, "writeback_jobs") && opt_params >= 1) {
2224			string = dm_shift_arg(&as), opt_params--;
2225			if (sscanf(string, "%u%c", &wc->max_writeback_jobs, &dummy) != 1)
2226				goto invalid_optional;
2227			wc->max_writeback_jobs_set = true;
2228		} else if (!strcasecmp(string, "autocommit_blocks") && opt_params >= 1) {
2229			string = dm_shift_arg(&as), opt_params--;
2230			if (sscanf(string, "%u%c", &wc->autocommit_blocks, &dummy) != 1)
2231				goto invalid_optional;
2232			wc->autocommit_blocks_set = true;
2233		} else if (!strcasecmp(string, "autocommit_time") && opt_params >= 1) {
2234			unsigned autocommit_msecs;
2235			string = dm_shift_arg(&as), opt_params--;
2236			if (sscanf(string, "%u%c", &autocommit_msecs, &dummy) != 1)
2237				goto invalid_optional;
2238			if (autocommit_msecs > 3600000)
2239				goto invalid_optional;
2240			wc->autocommit_jiffies = msecs_to_jiffies(autocommit_msecs);
2241			wc->autocommit_time_set = true;
2242		} else if (!strcasecmp(string, "max_age") && opt_params >= 1) {
2243			unsigned max_age_msecs;
2244			string = dm_shift_arg(&as), opt_params--;
2245			if (sscanf(string, "%u%c", &max_age_msecs, &dummy) != 1)
2246				goto invalid_optional;
2247			if (max_age_msecs > 86400000)
2248				goto invalid_optional;
2249			wc->max_age = msecs_to_jiffies(max_age_msecs);
2250		} else if (!strcasecmp(string, "cleaner")) {
2251			wc->cleaner = true;
2252		} else if (!strcasecmp(string, "fua")) {
2253			if (WC_MODE_PMEM(wc)) {
2254				wc->writeback_fua = true;
2255				wc->writeback_fua_set = true;
2256			} else goto invalid_optional;
2257		} else if (!strcasecmp(string, "nofua")) {
2258			if (WC_MODE_PMEM(wc)) {
2259				wc->writeback_fua = false;
2260				wc->writeback_fua_set = true;
2261			} else goto invalid_optional;
2262		} else {
2263invalid_optional:
2264			r = -EINVAL;
2265			ti->error = "Invalid optional argument";
2266			goto bad;
2267		}
2268	}
2269
2270	if (high_wm_percent < low_wm_percent) {
2271		r = -EINVAL;
2272		ti->error = "High watermark must be greater than or equal to low watermark";
2273		goto bad;
2274	}
2275
2276	if (WC_MODE_PMEM(wc)) {
2277		if (!dax_synchronous(wc->ssd_dev->dax_dev)) {
2278			r = -EOPNOTSUPP;
2279			ti->error = "Asynchronous persistent memory not supported as pmem cache";
2280			goto bad;
2281		}
2282
2283		r = persistent_memory_claim(wc);
2284		if (r) {
2285			ti->error = "Unable to map persistent memory for cache";
2286			goto bad;
2287		}
2288	} else {
 
 
2289		size_t n_blocks, n_metadata_blocks;
2290		uint64_t n_bitmap_bits;
2291
2292		wc->memory_map_size -= (uint64_t)wc->start_sector << SECTOR_SHIFT;
2293
2294		bio_list_init(&wc->flush_list);
2295		wc->flush_thread = kthread_create(writecache_flush_thread, wc, "dm_writecache_flush");
2296		if (IS_ERR(wc->flush_thread)) {
2297			r = PTR_ERR(wc->flush_thread);
2298			wc->flush_thread = NULL;
2299			ti->error = "Couldn't spawn flush thread";
2300			goto bad;
2301		}
2302		wake_up_process(wc->flush_thread);
2303
2304		r = calculate_memory_size(wc->memory_map_size, wc->block_size,
2305					  &n_blocks, &n_metadata_blocks);
2306		if (r) {
2307			ti->error = "Invalid device size";
2308			goto bad;
2309		}
2310
2311		n_bitmap_bits = (((uint64_t)n_metadata_blocks << wc->block_size_bits) +
2312				 BITMAP_GRANULARITY - 1) / BITMAP_GRANULARITY;
2313		/* this is limitation of test_bit functions */
2314		if (n_bitmap_bits > 1U << 31) {
2315			r = -EFBIG;
2316			ti->error = "Invalid device size";
2317			goto bad;
2318		}
2319
2320		wc->memory_map = vmalloc(n_metadata_blocks << wc->block_size_bits);
2321		if (!wc->memory_map) {
2322			r = -ENOMEM;
2323			ti->error = "Unable to allocate memory for metadata";
2324			goto bad;
2325		}
2326
2327		wc->dm_kcopyd = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2328		if (IS_ERR(wc->dm_kcopyd)) {
2329			r = PTR_ERR(wc->dm_kcopyd);
2330			ti->error = "Unable to allocate dm-kcopyd client";
2331			wc->dm_kcopyd = NULL;
2332			goto bad;
2333		}
2334
2335		wc->metadata_sectors = n_metadata_blocks << (wc->block_size_bits - SECTOR_SHIFT);
2336		wc->dirty_bitmap_size = (n_bitmap_bits + BITS_PER_LONG - 1) /
2337			BITS_PER_LONG * sizeof(unsigned long);
2338		wc->dirty_bitmap = vzalloc(wc->dirty_bitmap_size);
2339		if (!wc->dirty_bitmap) {
2340			r = -ENOMEM;
2341			ti->error = "Unable to allocate dirty bitmap";
2342			goto bad;
2343		}
2344
2345		r = writecache_read_metadata(wc, wc->block_size >> SECTOR_SHIFT);
 
 
 
 
 
 
 
 
 
 
2346		if (r) {
2347			ti->error = "Unable to read first block of metadata";
2348			goto bad;
2349		}
2350	}
2351
2352	r = memcpy_mcsafe(&s, sb(wc), sizeof(struct wc_memory_superblock));
2353	if (r) {
2354		ti->error = "Hardware memory error when reading superblock";
2355		goto bad;
2356	}
2357	if (!le32_to_cpu(s.magic) && !le32_to_cpu(s.version)) {
2358		r = init_memory(wc);
2359		if (r) {
2360			ti->error = "Unable to initialize device";
2361			goto bad;
2362		}
2363		r = memcpy_mcsafe(&s, sb(wc), sizeof(struct wc_memory_superblock));
2364		if (r) {
2365			ti->error = "Hardware memory error when reading superblock";
2366			goto bad;
2367		}
2368	}
2369
2370	if (le32_to_cpu(s.magic) != MEMORY_SUPERBLOCK_MAGIC) {
2371		ti->error = "Invalid magic in the superblock";
2372		r = -EINVAL;
2373		goto bad;
2374	}
2375
2376	if (le32_to_cpu(s.version) != MEMORY_SUPERBLOCK_VERSION) {
2377		ti->error = "Invalid version in the superblock";
2378		r = -EINVAL;
2379		goto bad;
2380	}
2381
2382	if (le32_to_cpu(s.block_size) != wc->block_size) {
2383		ti->error = "Block size does not match superblock";
2384		r = -EINVAL;
2385		goto bad;
2386	}
2387
2388	wc->n_blocks = le64_to_cpu(s.n_blocks);
2389
2390	offset = wc->n_blocks * sizeof(struct wc_memory_entry);
2391	if (offset / sizeof(struct wc_memory_entry) != le64_to_cpu(sb(wc)->n_blocks)) {
2392overflow:
2393		ti->error = "Overflow in size calculation";
2394		r = -EINVAL;
2395		goto bad;
2396	}
2397	offset += sizeof(struct wc_memory_superblock);
2398	if (offset < sizeof(struct wc_memory_superblock))
2399		goto overflow;
2400	offset = (offset + wc->block_size - 1) & ~(size_t)(wc->block_size - 1);
2401	data_size = wc->n_blocks * (size_t)wc->block_size;
2402	if (!offset || (data_size / wc->block_size != wc->n_blocks) ||
2403	    (offset + data_size < offset))
2404		goto overflow;
2405	if (offset + data_size > wc->memory_map_size) {
2406		ti->error = "Memory area is too small";
2407		r = -EINVAL;
2408		goto bad;
2409	}
2410
2411	wc->metadata_sectors = offset >> SECTOR_SHIFT;
2412	wc->block_start = (char *)sb(wc) + offset;
2413
2414	x = (uint64_t)wc->n_blocks * (100 - high_wm_percent);
2415	x += 50;
2416	do_div(x, 100);
2417	wc->freelist_high_watermark = x;
2418	x = (uint64_t)wc->n_blocks * (100 - low_wm_percent);
2419	x += 50;
2420	do_div(x, 100);
2421	wc->freelist_low_watermark = x;
2422
2423	if (wc->cleaner)
2424		activate_cleaner(wc);
2425
2426	r = writecache_alloc_entries(wc);
2427	if (r) {
2428		ti->error = "Cannot allocate memory";
2429		goto bad;
2430	}
2431
2432	ti->num_flush_bios = 1;
2433	ti->flush_supported = true;
2434	ti->num_discard_bios = 1;
2435
2436	if (WC_MODE_PMEM(wc))
2437		persistent_memory_flush_cache(wc->memory_map, wc->memory_map_size);
2438
2439	return 0;
2440
2441bad_arguments:
2442	r = -EINVAL;
2443	ti->error = "Bad arguments";
2444bad:
2445	writecache_dtr(ti);
2446	return r;
2447}
2448
2449static void writecache_status(struct dm_target *ti, status_type_t type,
2450			      unsigned status_flags, char *result, unsigned maxlen)
2451{
2452	struct dm_writecache *wc = ti->private;
2453	unsigned extra_args;
2454	unsigned sz = 0;
2455	uint64_t x;
2456
2457	switch (type) {
2458	case STATUSTYPE_INFO:
2459		DMEMIT("%ld %llu %llu %llu", writecache_has_error(wc),
2460		       (unsigned long long)wc->n_blocks, (unsigned long long)wc->freelist_size,
2461		       (unsigned long long)wc->writeback_size);
2462		break;
2463	case STATUSTYPE_TABLE:
2464		DMEMIT("%c %s %s %u ", WC_MODE_PMEM(wc) ? 'p' : 's',
2465				wc->dev->name, wc->ssd_dev->name, wc->block_size);
2466		extra_args = 0;
2467		if (wc->start_sector)
2468			extra_args += 2;
2469		if (wc->high_wm_percent_set && !wc->cleaner)
2470			extra_args += 2;
2471		if (wc->low_wm_percent_set && !wc->cleaner)
2472			extra_args += 2;
2473		if (wc->max_writeback_jobs_set)
2474			extra_args += 2;
2475		if (wc->autocommit_blocks_set)
2476			extra_args += 2;
2477		if (wc->autocommit_time_set)
2478			extra_args += 2;
2479		if (wc->cleaner)
2480			extra_args++;
2481		if (wc->writeback_fua_set)
2482			extra_args++;
2483
2484		DMEMIT("%u", extra_args);
2485		if (wc->start_sector)
2486			DMEMIT(" start_sector %llu", (unsigned long long)wc->start_sector);
2487		if (wc->high_wm_percent_set && !wc->cleaner) {
2488			x = (uint64_t)wc->freelist_high_watermark * 100;
2489			x += wc->n_blocks / 2;
2490			do_div(x, (size_t)wc->n_blocks);
2491			DMEMIT(" high_watermark %u", 100 - (unsigned)x);
2492		}
2493		if (wc->low_wm_percent_set && !wc->cleaner) {
2494			x = (uint64_t)wc->freelist_low_watermark * 100;
2495			x += wc->n_blocks / 2;
2496			do_div(x, (size_t)wc->n_blocks);
2497			DMEMIT(" low_watermark %u", 100 - (unsigned)x);
2498		}
2499		if (wc->max_writeback_jobs_set)
2500			DMEMIT(" writeback_jobs %u", wc->max_writeback_jobs);
2501		if (wc->autocommit_blocks_set)
2502			DMEMIT(" autocommit_blocks %u", wc->autocommit_blocks);
2503		if (wc->autocommit_time_set)
2504			DMEMIT(" autocommit_time %u", jiffies_to_msecs(wc->autocommit_jiffies));
2505		if (wc->max_age != MAX_AGE_UNSPECIFIED)
2506			DMEMIT(" max_age %u", jiffies_to_msecs(wc->max_age));
2507		if (wc->cleaner)
2508			DMEMIT(" cleaner");
2509		if (wc->writeback_fua_set)
2510			DMEMIT(" %sfua", wc->writeback_fua ? "" : "no");
2511		break;
2512	}
2513}
2514
2515static struct target_type writecache_target = {
2516	.name			= "writecache",
2517	.version		= {1, 3, 0},
2518	.module			= THIS_MODULE,
2519	.ctr			= writecache_ctr,
2520	.dtr			= writecache_dtr,
2521	.status			= writecache_status,
2522	.postsuspend		= writecache_suspend,
2523	.resume			= writecache_resume,
2524	.message		= writecache_message,
2525	.map			= writecache_map,
2526	.end_io			= writecache_end_io,
2527	.iterate_devices	= writecache_iterate_devices,
2528	.io_hints		= writecache_io_hints,
2529};
2530
2531static int __init dm_writecache_init(void)
2532{
2533	int r;
2534
2535	r = dm_register_target(&writecache_target);
2536	if (r < 0) {
2537		DMERR("register failed %d", r);
2538		return r;
2539	}
2540
2541	return 0;
2542}
2543
2544static void __exit dm_writecache_exit(void)
2545{
2546	dm_unregister_target(&writecache_target);
2547}
2548
2549module_init(dm_writecache_init);
2550module_exit(dm_writecache_exit);
2551
2552MODULE_DESCRIPTION(DM_NAME " writecache target");
2553MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
2554MODULE_LICENSE("GPL");