drivers/md/dm-writecache.c

   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 #include <linux/delay.h>
  19 #include "dm-io-tracker.h"
  20
  21 #define DM_MSG_PREFIX "writecache"
  22
  23 #define HIGH_WATERMARK                  50
  24 #define LOW_WATERMARK                   45
  25 #define MAX_WRITEBACK_JOBS              min(0x10000000 / PAGE_SIZE, totalram_pages() / 16)
  26 #define ENDIO_LATENCY                   16
  27 #define WRITEBACK_LATENCY               64
  28 #define AUTOCOMMIT_BLOCKS_SSD           65536
  29 #define AUTOCOMMIT_BLOCKS_PMEM          64
  30 #define AUTOCOMMIT_MSEC                 1000
  31 #define MAX_AGE_DIV                     16
  32 #define MAX_AGE_UNSPECIFIED             -1UL
  33 #define PAUSE_WRITEBACK                 (HZ * 3)
  34
  35 #define BITMAP_GRANULARITY      65536
  36 #if BITMAP_GRANULARITY < PAGE_SIZE
  37 #undef BITMAP_GRANULARITY
  38 #define BITMAP_GRANULARITY      PAGE_SIZE
  39 #endif
  40
  41 #if IS_ENABLED(CONFIG_ARCH_HAS_PMEM_API) && IS_ENABLED(CONFIG_FS_DAX)
  42 #define DM_WRITECACHE_HAS_PMEM
  43 #endif
  44
  45 #ifdef DM_WRITECACHE_HAS_PMEM
  46 #define pmem_assign(dest, src)                                  \
  47 do {                                                            \
  48         typeof(dest) uniq = (src);                              \
  49         memcpy_flushcache(&(dest), &uniq, sizeof(dest));        \
  50 } while (0)
  51 #else
  52 #define pmem_assign(dest, src)  ((dest) = (src))
  53 #endif
  54
  55 #if IS_ENABLED(CONFIG_ARCH_HAS_COPY_MC) && defined(DM_WRITECACHE_HAS_PMEM)
  56 #define DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
  57 #endif
  58
  59 #define MEMORY_SUPERBLOCK_MAGIC         0x23489321
  60 #define MEMORY_SUPERBLOCK_VERSION       1
  61
  62 struct wc_memory_entry {
  63         __le64 original_sector;
  64         __le64 seq_count;
  65 };
  66
  67 struct wc_memory_superblock {
  68         union {
  69                 struct {
  70                         __le32 magic;
  71                         __le32 version;
  72                         __le32 block_size;
  73                         __le32 pad;
  74                         __le64 n_blocks;
  75                         __le64 seq_count;
  76                 };
  77                 __le64 padding[8];
  78         };
  79         struct wc_memory_entry entries[];
  80 };
  81
  82 struct wc_entry {
  83         struct rb_node rb_node;
  84         struct list_head lru;
  85         unsigned short wc_list_contiguous;
  86         bool write_in_progress
  87 #if BITS_PER_LONG == 64
  88                 :1
  89 #endif
  90         ;
  91         unsigned long index
  92 #if BITS_PER_LONG == 64
  93                 :47
  94 #endif
  95         ;
  96         unsigned long age;
  97 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
  98         uint64_t original_sector;
  99         uint64_t seq_count;
 100 #endif
 101 };
 102
 103 #ifdef DM_WRITECACHE_HAS_PMEM
 104 #define WC_MODE_PMEM(wc)                        ((wc)->pmem_mode)
 105 #define WC_MODE_FUA(wc)                         ((wc)->writeback_fua)
 106 #else
 107 #define WC_MODE_PMEM(wc)                        false
 108 #define WC_MODE_FUA(wc)                         false
 109 #endif
 110 #define WC_MODE_SORT_FREELIST(wc)               (!WC_MODE_PMEM(wc))
 111
 112 struct dm_writecache {
 113         struct mutex lock;
 114         struct list_head lru;
 115         union {
 116                 struct list_head freelist;
 117                 struct {
 118                         struct rb_root freetree;
 119                         struct wc_entry *current_free;
 120                 };
 121         };
 122         struct rb_root tree;
 123
 124         size_t freelist_size;
 125         size_t writeback_size;
 126         size_t freelist_high_watermark;
 127         size_t freelist_low_watermark;
 128         unsigned long max_age;
 129         unsigned long pause;
 130
 131         unsigned uncommitted_blocks;
 132         unsigned autocommit_blocks;
 133         unsigned max_writeback_jobs;
 134
 135         int error;
 136
 137         unsigned long autocommit_jiffies;
 138         struct timer_list autocommit_timer;
 139         struct wait_queue_head freelist_wait;
 140
 141         struct timer_list max_age_timer;
 142
 143         atomic_t bio_in_progress[2];
 144         struct wait_queue_head bio_in_progress_wait[2];
 145
 146         struct dm_target *ti;
 147         struct dm_dev *dev;
 148         struct dm_dev *ssd_dev;
 149         sector_t start_sector;
 150         void *memory_map;
 151         uint64_t memory_map_size;
 152         size_t metadata_sectors;
 153         size_t n_blocks;
 154         uint64_t seq_count;
 155         sector_t data_device_sectors;
 156         void *block_start;
 157         struct wc_entry *entries;
 158         unsigned block_size;
 159         unsigned char block_size_bits;
 160
 161         bool pmem_mode:1;
 162         bool writeback_fua:1;
 163
 164         bool overwrote_committed:1;
 165         bool memory_vmapped:1;
 166
 167         bool start_sector_set:1;
 168         bool high_wm_percent_set:1;
 169         bool low_wm_percent_set:1;
 170         bool max_writeback_jobs_set:1;
 171         bool autocommit_blocks_set:1;
 172         bool autocommit_time_set:1;
 173         bool max_age_set:1;
 174         bool writeback_fua_set:1;
 175         bool flush_on_suspend:1;
 176         bool cleaner:1;
 177         bool cleaner_set:1;
 178         bool metadata_only:1;
 179         bool pause_set:1;
 180
 181         unsigned high_wm_percent_value;
 182         unsigned low_wm_percent_value;
 183         unsigned autocommit_time_value;
 184         unsigned max_age_value;
 185         unsigned pause_value;
 186
 187         unsigned writeback_all;
 188         struct workqueue_struct *writeback_wq;
 189         struct work_struct writeback_work;
 190         struct work_struct flush_work;
 191
 192         struct dm_io_tracker iot;
 193
 194         struct dm_io_client *dm_io;
 195
 196         raw_spinlock_t endio_list_lock;
 197         struct list_head endio_list;
 198         struct task_struct *endio_thread;
 199
 200         struct task_struct *flush_thread;
 201         struct bio_list flush_list;
 202
 203         struct dm_kcopyd_client *dm_kcopyd;
 204         unsigned long *dirty_bitmap;
 205         unsigned dirty_bitmap_size;
 206
 207         struct bio_set bio_set;
 208         mempool_t copy_pool;
 209
 210         struct {
 211                 unsigned long long reads;
 212                 unsigned long long read_hits;
 213                 unsigned long long writes;
 214                 unsigned long long write_hits_uncommitted;
 215                 unsigned long long write_hits_committed;
 216                 unsigned long long writes_around;
 217                 unsigned long long writes_allocate;
 218                 unsigned long long writes_blocked_on_freelist;
 219                 unsigned long long flushes;
 220                 unsigned long long discards;
 221         } stats;
 222 };
 223
 224 #define WB_LIST_INLINE          16
 225
 226 struct writeback_struct {
 227         struct list_head endio_entry;
 228         struct dm_writecache *wc;
 229         struct wc_entry **wc_list;
 230         unsigned wc_list_n;
 231         struct wc_entry *wc_list_inline[WB_LIST_INLINE];
 232         struct bio bio;
 233 };
 234
 235 struct copy_struct {
 236         struct list_head endio_entry;
 237         struct dm_writecache *wc;
 238         struct wc_entry *e;
 239         unsigned n_entries;
 240         int error;
 241 };
 242
 243 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(dm_writecache_throttle,
 244                                             "A percentage of time allocated for data copying");
 245
 246 static void wc_lock(struct dm_writecache *wc)
 247 {
 248         mutex_lock(&wc->lock);
 249 }
 250
 251 static void wc_unlock(struct dm_writecache *wc)
 252 {
 253         mutex_unlock(&wc->lock);
 254 }
 255
 256 #ifdef DM_WRITECACHE_HAS_PMEM
 257 static int persistent_memory_claim(struct dm_writecache *wc)
 258 {
 259         int r;
 260         loff_t s;
 261         long p, da;
 262         pfn_t pfn;
 263         int id;
 264         struct page **pages;
 265         sector_t offset;
 266
 267         wc->memory_vmapped = false;
 268
 269         s = wc->memory_map_size;
 270         p = s >> PAGE_SHIFT;
 271         if (!p) {
 272                 r = -EINVAL;
 273                 goto err1;
 274         }
 275         if (p != s >> PAGE_SHIFT) {
 276                 r = -EOVERFLOW;
 277                 goto err1;
 278         }
 279
 280         offset = get_start_sect(wc->ssd_dev->bdev);
 281         if (offset & (PAGE_SIZE / 512 - 1)) {
 282                 r = -EINVAL;
 283                 goto err1;
 284         }
 285         offset >>= PAGE_SHIFT - 9;
 286
 287         id = dax_read_lock();
 288
 289         da = dax_direct_access(wc->ssd_dev->dax_dev, offset, p, DAX_ACCESS,
 290                         &wc->memory_map, &pfn);
 291         if (da < 0) {
 292                 wc->memory_map = NULL;
 293                 r = da;
 294                 goto err2;
 295         }
 296         if (!pfn_t_has_page(pfn)) {
 297                 wc->memory_map = NULL;
 298                 r = -EOPNOTSUPP;
 299                 goto err2;
 300         }
 301         if (da != p) {
 302                 long i;
 303                 wc->memory_map = NULL;
 304                 pages = kvmalloc_array(p, sizeof(struct page *), GFP_KERNEL);
 305                 if (!pages) {
 306                         r = -ENOMEM;
 307                         goto err2;
 308                 }
 309                 i = 0;
 310                 do {
 311                         long daa;
 312                         daa = dax_direct_access(wc->ssd_dev->dax_dev, offset + i,
 313                                         p - i, DAX_ACCESS, NULL, &pfn);
 314                         if (daa <= 0) {
 315                                 r = daa ? daa : -EINVAL;
 316                                 goto err3;
 317                         }
 318                         if (!pfn_t_has_page(pfn)) {
 319                                 r = -EOPNOTSUPP;
 320                                 goto err3;
 321                         }
 322                         while (daa-- && i < p) {
 323                                 pages[i++] = pfn_t_to_page(pfn);
 324                                 pfn.val++;
 325                                 if (!(i & 15))
 326                                         cond_resched();
 327                         }
 328                 } while (i < p);
 329                 wc->memory_map = vmap(pages, p, VM_MAP, PAGE_KERNEL);
 330                 if (!wc->memory_map) {
 331                         r = -ENOMEM;
 332                         goto err3;
 333                 }
 334                 kvfree(pages);
 335                 wc->memory_vmapped = true;
 336         }
 337
 338         dax_read_unlock(id);
 339
 340         wc->memory_map += (size_t)wc->start_sector << SECTOR_SHIFT;
 341         wc->memory_map_size -= (size_t)wc->start_sector << SECTOR_SHIFT;
 342
 343         return 0;
 344 err3:
 345         kvfree(pages);
 346 err2:
 347         dax_read_unlock(id);
 348 err1:
 349         return r;
 350 }
 351 #else
 352 static int persistent_memory_claim(struct dm_writecache *wc)
 353 {
 354         return -EOPNOTSUPP;
 355 }
 356 #endif
 357
 358 static void persistent_memory_release(struct dm_writecache *wc)
 359 {
 360         if (wc->memory_vmapped)
 361                 vunmap(wc->memory_map - ((size_t)wc->start_sector << SECTOR_SHIFT));
 362 }
 363
 364 static struct page *persistent_memory_page(void *addr)
 365 {
 366         if (is_vmalloc_addr(addr))
 367                 return vmalloc_to_page(addr);
 368         else
 369                 return virt_to_page(addr);
 370 }
 371
 372 static unsigned persistent_memory_page_offset(void *addr)
 373 {
 374         return (unsigned long)addr & (PAGE_SIZE - 1);
 375 }
 376
 377 static void persistent_memory_flush_cache(void *ptr, size_t size)
 378 {
 379         if (is_vmalloc_addr(ptr))
 380                 flush_kernel_vmap_range(ptr, size);
 381 }
 382
 383 static void persistent_memory_invalidate_cache(void *ptr, size_t size)
 384 {
 385         if (is_vmalloc_addr(ptr))
 386                 invalidate_kernel_vmap_range(ptr, size);
 387 }
 388
 389 static struct wc_memory_superblock *sb(struct dm_writecache *wc)
 390 {
 391         return wc->memory_map;
 392 }
 393
 394 static struct wc_memory_entry *memory_entry(struct dm_writecache *wc, struct wc_entry *e)
 395 {
 396         return &sb(wc)->entries[e->index];
 397 }
 398
 399 static void *memory_data(struct dm_writecache *wc, struct wc_entry *e)
 400 {
 401         return (char *)wc->block_start + (e->index << wc->block_size_bits);
 402 }
 403
 404 static sector_t cache_sector(struct dm_writecache *wc, struct wc_entry *e)
 405 {
 406         return wc->start_sector + wc->metadata_sectors +
 407                 ((sector_t)e->index << (wc->block_size_bits - SECTOR_SHIFT));
 408 }
 409
 410 static uint64_t read_original_sector(struct dm_writecache *wc, struct wc_entry *e)
 411 {
 412 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 413         return e->original_sector;
 414 #else
 415         return le64_to_cpu(memory_entry(wc, e)->original_sector);
 416 #endif
 417 }
 418
 419 static uint64_t read_seq_count(struct dm_writecache *wc, struct wc_entry *e)
 420 {
 421 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 422         return e->seq_count;
 423 #else
 424         return le64_to_cpu(memory_entry(wc, e)->seq_count);
 425 #endif
 426 }
 427
 428 static void clear_seq_count(struct dm_writecache *wc, struct wc_entry *e)
 429 {
 430 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 431         e->seq_count = -1;
 432 #endif
 433         pmem_assign(memory_entry(wc, e)->seq_count, cpu_to_le64(-1));
 434 }
 435
 436 static void write_original_sector_seq_count(struct dm_writecache *wc, struct wc_entry *e,
 437                                             uint64_t original_sector, uint64_t seq_count)
 438 {
 439         struct wc_memory_entry me;
 440 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 441         e->original_sector = original_sector;
 442         e->seq_count = seq_count;
 443 #endif
 444         me.original_sector = cpu_to_le64(original_sector);
 445         me.seq_count = cpu_to_le64(seq_count);
 446         pmem_assign(*memory_entry(wc, e), me);
 447 }
 448
 449 #define writecache_error(wc, err, msg, arg...)                          \
 450 do {                                                                    \
 451         if (!cmpxchg(&(wc)->error, 0, err))                             \
 452                 DMERR(msg, ##arg);                                      \
 453         wake_up(&(wc)->freelist_wait);                                  \
 454 } while (0)
 455
 456 #define writecache_has_error(wc)        (unlikely(READ_ONCE((wc)->error)))
 457
 458 static void writecache_flush_all_metadata(struct dm_writecache *wc)
 459 {
 460         if (!WC_MODE_PMEM(wc))
 461                 memset(wc->dirty_bitmap, -1, wc->dirty_bitmap_size);
 462 }
 463
 464 static void writecache_flush_region(struct dm_writecache *wc, void *ptr, size_t size)
 465 {
 466         if (!WC_MODE_PMEM(wc))
 467                 __set_bit(((char *)ptr - (char *)wc->memory_map) / BITMAP_GRANULARITY,
 468                           wc->dirty_bitmap);
 469 }
 470
 471 static void writecache_disk_flush(struct dm_writecache *wc, struct dm_dev *dev);
 472
 473 struct io_notify {
 474         struct dm_writecache *wc;
 475         struct completion c;
 476         atomic_t count;
 477 };
 478
 479 static void writecache_notify_io(unsigned long error, void *context)
 480 {
 481         struct io_notify *endio = context;
 482
 483         if (unlikely(error != 0))
 484                 writecache_error(endio->wc, -EIO, "error writing metadata");
 485         BUG_ON(atomic_read(&endio->count) <= 0);
 486         if (atomic_dec_and_test(&endio->count))
 487                 complete(&endio->c);
 488 }
 489
 490 static void writecache_wait_for_ios(struct dm_writecache *wc, int direction)
 491 {
 492         wait_event(wc->bio_in_progress_wait[direction],
 493                    !atomic_read(&wc->bio_in_progress[direction]));
 494 }
 495
 496 static void ssd_commit_flushed(struct dm_writecache *wc, bool wait_for_ios)
 497 {
 498         struct dm_io_region region;
 499         struct dm_io_request req;
 500         struct io_notify endio = {
 501                 wc,
 502                 COMPLETION_INITIALIZER_ONSTACK(endio.c),
 503                 ATOMIC_INIT(1),
 504         };
 505         unsigned bitmap_bits = wc->dirty_bitmap_size * 8;
 506         unsigned i = 0;
 507
 508         while (1) {
 509                 unsigned j;
 510                 i = find_next_bit(wc->dirty_bitmap, bitmap_bits, i);
 511                 if (unlikely(i == bitmap_bits))
 512                         break;
 513                 j = find_next_zero_bit(wc->dirty_bitmap, bitmap_bits, i);
 514
 515                 region.bdev = wc->ssd_dev->bdev;
 516                 region.sector = (sector_t)i * (BITMAP_GRANULARITY >> SECTOR_SHIFT);
 517                 region.count = (sector_t)(j - i) * (BITMAP_GRANULARITY >> SECTOR_SHIFT);
 518
 519                 if (unlikely(region.sector >= wc->metadata_sectors))
 520                         break;
 521                 if (unlikely(region.sector + region.count > wc->metadata_sectors))
 522                         region.count = wc->metadata_sectors - region.sector;
 523
 524                 region.sector += wc->start_sector;
 525                 atomic_inc(&endio.count);
 526                 req.bi_opf = REQ_OP_WRITE | REQ_SYNC;
 527                 req.mem.type = DM_IO_VMA;
 528                 req.mem.ptr.vma = (char *)wc->memory_map + (size_t)i * BITMAP_GRANULARITY;
 529                 req.client = wc->dm_io;
 530                 req.notify.fn = writecache_notify_io;
 531                 req.notify.context = &endio;
 532
 533                 /* writing via async dm-io (implied by notify.fn above) won't return an error */
 534                 (void) dm_io(&req, 1, &region, NULL);
 535                 i = j;
 536         }
 537
 538         writecache_notify_io(0, &endio);
 539         wait_for_completion_io(&endio.c);
 540
 541         if (wait_for_ios)
 542                 writecache_wait_for_ios(wc, WRITE);
 543
 544         writecache_disk_flush(wc, wc->ssd_dev);
 545
 546         memset(wc->dirty_bitmap, 0, wc->dirty_bitmap_size);
 547 }
 548
 549 static void ssd_commit_superblock(struct dm_writecache *wc)
 550 {
 551         int r;
 552         struct dm_io_region region;
 553         struct dm_io_request req;
 554
 555         region.bdev = wc->ssd_dev->bdev;
 556         region.sector = 0;
 557         region.count = max(4096U, wc->block_size) >> SECTOR_SHIFT;
 558
 559         if (unlikely(region.sector + region.count > wc->metadata_sectors))
 560                 region.count = wc->metadata_sectors - region.sector;
 561
 562         region.sector += wc->start_sector;
 563
 564         req.bi_opf = REQ_OP_WRITE | REQ_SYNC | REQ_FUA;
 565         req.mem.type = DM_IO_VMA;
 566         req.mem.ptr.vma = (char *)wc->memory_map;
 567         req.client = wc->dm_io;
 568         req.notify.fn = NULL;
 569         req.notify.context = NULL;
 570
 571         r = dm_io(&req, 1, &region, NULL);
 572         if (unlikely(r))
 573                 writecache_error(wc, r, "error writing superblock");
 574 }
 575
 576 static void writecache_commit_flushed(struct dm_writecache *wc, bool wait_for_ios)
 577 {
 578         if (WC_MODE_PMEM(wc))
 579                 pmem_wmb();
 580         else
 581                 ssd_commit_flushed(wc, wait_for_ios);
 582 }
 583
 584 static void writecache_disk_flush(struct dm_writecache *wc, struct dm_dev *dev)
 585 {
 586         int r;
 587         struct dm_io_region region;
 588         struct dm_io_request req;
 589
 590         region.bdev = dev->bdev;
 591         region.sector = 0;
 592         region.count = 0;
 593         req.bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
 594         req.mem.type = DM_IO_KMEM;
 595         req.mem.ptr.addr = NULL;
 596         req.client = wc->dm_io;
 597         req.notify.fn = NULL;
 598
 599         r = dm_io(&req, 1, &region, NULL);
 600         if (unlikely(r))
 601                 writecache_error(wc, r, "error flushing metadata: %d", r);
 602 }
 603
 604 #define WFE_RETURN_FOLLOWING    1
 605 #define WFE_LOWEST_SEQ          2
 606
 607 static struct wc_entry *writecache_find_entry(struct dm_writecache *wc,
 608                                               uint64_t block, int flags)
 609 {
 610         struct wc_entry *e;
 611         struct rb_node *node = wc->tree.rb_node;
 612
 613         if (unlikely(!node))
 614                 return NULL;
 615
 616         while (1) {
 617                 e = container_of(node, struct wc_entry, rb_node);
 618                 if (read_original_sector(wc, e) == block)
 619                         break;
 620
 621                 node = (read_original_sector(wc, e) >= block ?
 622                         e->rb_node.rb_left : e->rb_node.rb_right);
 623                 if (unlikely(!node)) {
 624                         if (!(flags & WFE_RETURN_FOLLOWING))
 625                                 return NULL;
 626                         if (read_original_sector(wc, e) >= block) {
 627                                 return e;
 628                         } else {
 629                                 node = rb_next(&e->rb_node);
 630                                 if (unlikely(!node))
 631                                         return NULL;
 632                                 e = container_of(node, struct wc_entry, rb_node);
 633                                 return e;
 634                         }
 635                 }
 636         }
 637
 638         while (1) {
 639                 struct wc_entry *e2;
 640                 if (flags & WFE_LOWEST_SEQ)
 641                         node = rb_prev(&e->rb_node);
 642                 else
 643                         node = rb_next(&e->rb_node);
 644                 if (unlikely(!node))
 645                         return e;
 646                 e2 = container_of(node, struct wc_entry, rb_node);
 647                 if (read_original_sector(wc, e2) != block)
 648                         return e;
 649                 e = e2;
 650         }
 651 }
 652
 653 static void writecache_insert_entry(struct dm_writecache *wc, struct wc_entry *ins)
 654 {
 655         struct wc_entry *e;
 656         struct rb_node **node = &wc->tree.rb_node, *parent = NULL;
 657
 658         while (*node) {
 659                 e = container_of(*node, struct wc_entry, rb_node);
 660                 parent = &e->rb_node;
 661                 if (read_original_sector(wc, e) > read_original_sector(wc, ins))
 662                         node = &parent->rb_left;
 663                 else
 664                         node = &parent->rb_right;
 665         }
 666         rb_link_node(&ins->rb_node, parent, node);
 667         rb_insert_color(&ins->rb_node, &wc->tree);
 668         list_add(&ins->lru, &wc->lru);
 669         ins->age = jiffies;
 670 }
 671
 672 static void writecache_unlink(struct dm_writecache *wc, struct wc_entry *e)
 673 {
 674         list_del(&e->lru);
 675         rb_erase(&e->rb_node, &wc->tree);
 676 }
 677
 678 static void writecache_add_to_freelist(struct dm_writecache *wc, struct wc_entry *e)
 679 {
 680         if (WC_MODE_SORT_FREELIST(wc)) {
 681                 struct rb_node **node = &wc->freetree.rb_node, *parent = NULL;
 682                 if (unlikely(!*node))
 683                         wc->current_free = e;
 684                 while (*node) {
 685                         parent = *node;
 686                         if (&e->rb_node < *node)
 687                                 node = &parent->rb_left;
 688                         else
 689                                 node = &parent->rb_right;
 690                 }
 691                 rb_link_node(&e->rb_node, parent, node);
 692                 rb_insert_color(&e->rb_node, &wc->freetree);
 693         } else {
 694                 list_add_tail(&e->lru, &wc->freelist);
 695         }
 696         wc->freelist_size++;
 697 }
 698
 699 static inline void writecache_verify_watermark(struct dm_writecache *wc)
 700 {
 701         if (unlikely(wc->freelist_size + wc->writeback_size <= wc->freelist_high_watermark))
 702                 queue_work(wc->writeback_wq, &wc->writeback_work);
 703 }
 704
 705 static void writecache_max_age_timer(struct timer_list *t)
 706 {
 707         struct dm_writecache *wc = from_timer(wc, t, max_age_timer);
 708
 709         if (!dm_suspended(wc->ti) && !writecache_has_error(wc)) {
 710                 queue_work(wc->writeback_wq, &wc->writeback_work);
 711                 mod_timer(&wc->max_age_timer, jiffies + wc->max_age / MAX_AGE_DIV);
 712         }
 713 }
 714
 715 static struct wc_entry *writecache_pop_from_freelist(struct dm_writecache *wc, sector_t expected_sector)
 716 {
 717         struct wc_entry *e;
 718
 719         if (WC_MODE_SORT_FREELIST(wc)) {
 720                 struct rb_node *next;
 721                 if (unlikely(!wc->current_free))
 722                         return NULL;
 723                 e = wc->current_free;
 724                 if (expected_sector != (sector_t)-1 && unlikely(cache_sector(wc, e) != expected_sector))
 725                         return NULL;
 726                 next = rb_next(&e->rb_node);
 727                 rb_erase(&e->rb_node, &wc->freetree);
 728                 if (unlikely(!next))
 729                         next = rb_first(&wc->freetree);
 730                 wc->current_free = next ? container_of(next, struct wc_entry, rb_node) : NULL;
 731         } else {
 732                 if (unlikely(list_empty(&wc->freelist)))
 733                         return NULL;
 734                 e = container_of(wc->freelist.next, struct wc_entry, lru);
 735                 if (expected_sector != (sector_t)-1 && unlikely(cache_sector(wc, e) != expected_sector))
 736                         return NULL;
 737                 list_del(&e->lru);
 738         }
 739         wc->freelist_size--;
 740
 741         writecache_verify_watermark(wc);
 742
 743         return e;
 744 }
 745
 746 static void writecache_free_entry(struct dm_writecache *wc, struct wc_entry *e)
 747 {
 748         writecache_unlink(wc, e);
 749         writecache_add_to_freelist(wc, e);
 750         clear_seq_count(wc, e);
 751         writecache_flush_region(wc, memory_entry(wc, e), sizeof(struct wc_memory_entry));
 752         if (unlikely(waitqueue_active(&wc->freelist_wait)))
 753                 wake_up(&wc->freelist_wait);
 754 }
 755
 756 static void writecache_wait_on_freelist(struct dm_writecache *wc)
 757 {
 758         DEFINE_WAIT(wait);
 759
 760         prepare_to_wait(&wc->freelist_wait, &wait, TASK_UNINTERRUPTIBLE);
 761         wc_unlock(wc);
 762         io_schedule();
 763         finish_wait(&wc->freelist_wait, &wait);
 764         wc_lock(wc);
 765 }
 766
 767 static void writecache_poison_lists(struct dm_writecache *wc)
 768 {
 769         /*
 770          * Catch incorrect access to these values while the device is suspended.
 771          */
 772         memset(&wc->tree, -1, sizeof wc->tree);
 773         wc->lru.next = LIST_POISON1;
 774         wc->lru.prev = LIST_POISON2;
 775         wc->freelist.next = LIST_POISON1;
 776         wc->freelist.prev = LIST_POISON2;
 777 }
 778
 779 static void writecache_flush_entry(struct dm_writecache *wc, struct wc_entry *e)
 780 {
 781         writecache_flush_region(wc, memory_entry(wc, e), sizeof(struct wc_memory_entry));
 782         if (WC_MODE_PMEM(wc))
 783                 writecache_flush_region(wc, memory_data(wc, e), wc->block_size);
 784 }
 785
 786 static bool writecache_entry_is_committed(struct dm_writecache *wc, struct wc_entry *e)
 787 {
 788         return read_seq_count(wc, e) < wc->seq_count;
 789 }
 790
 791 static void writecache_flush(struct dm_writecache *wc)
 792 {
 793         struct wc_entry *e, *e2;
 794         bool need_flush_after_free;
 795
 796         wc->uncommitted_blocks = 0;
 797         del_timer(&wc->autocommit_timer);
 798
 799         if (list_empty(&wc->lru))
 800                 return;
 801
 802         e = container_of(wc->lru.next, struct wc_entry, lru);
 803         if (writecache_entry_is_committed(wc, e)) {
 804                 if (wc->overwrote_committed) {
 805                         writecache_wait_for_ios(wc, WRITE);
 806                         writecache_disk_flush(wc, wc->ssd_dev);
 807                         wc->overwrote_committed = false;
 808                 }
 809                 return;
 810         }
 811         while (1) {
 812                 writecache_flush_entry(wc, e);
 813                 if (unlikely(e->lru.next == &wc->lru))
 814                         break;
 815                 e2 = container_of(e->lru.next, struct wc_entry, lru);
 816                 if (writecache_entry_is_committed(wc, e2))
 817                         break;
 818                 e = e2;
 819                 cond_resched();
 820         }
 821         writecache_commit_flushed(wc, true);
 822
 823         wc->seq_count++;
 824         pmem_assign(sb(wc)->seq_count, cpu_to_le64(wc->seq_count));
 825         if (WC_MODE_PMEM(wc))
 826                 writecache_commit_flushed(wc, false);
 827         else
 828                 ssd_commit_superblock(wc);
 829
 830         wc->overwrote_committed = false;
 831
 832         need_flush_after_free = false;
 833         while (1) {
 834                 /* Free another committed entry with lower seq-count */
 835                 struct rb_node *rb_node = rb_prev(&e->rb_node);
 836
 837                 if (rb_node) {
 838                         e2 = container_of(rb_node, struct wc_entry, rb_node);
 839                         if (read_original_sector(wc, e2) == read_original_sector(wc, e) &&
 840                             likely(!e2->write_in_progress)) {
 841                                 writecache_free_entry(wc, e2);
 842                                 need_flush_after_free = true;
 843                         }
 844                 }
 845                 if (unlikely(e->lru.prev == &wc->lru))
 846                         break;
 847                 e = container_of(e->lru.prev, struct wc_entry, lru);
 848                 cond_resched();
 849         }
 850
 851         if (need_flush_after_free)
 852                 writecache_commit_flushed(wc, false);
 853 }
 854
 855 static void writecache_flush_work(struct work_struct *work)
 856 {
 857         struct dm_writecache *wc = container_of(work, struct dm_writecache, flush_work);
 858
 859         wc_lock(wc);
 860         writecache_flush(wc);
 861         wc_unlock(wc);
 862 }
 863
 864 static void writecache_autocommit_timer(struct timer_list *t)
 865 {
 866         struct dm_writecache *wc = from_timer(wc, t, autocommit_timer);
 867         if (!writecache_has_error(wc))
 868                 queue_work(wc->writeback_wq, &wc->flush_work);
 869 }
 870
 871 static void writecache_schedule_autocommit(struct dm_writecache *wc)
 872 {
 873         if (!timer_pending(&wc->autocommit_timer))
 874                 mod_timer(&wc->autocommit_timer, jiffies + wc->autocommit_jiffies);
 875 }
 876
 877 static void writecache_discard(struct dm_writecache *wc, sector_t start, sector_t end)
 878 {
 879         struct wc_entry *e;
 880         bool discarded_something = false;
 881
 882         e = writecache_find_entry(wc, start, WFE_RETURN_FOLLOWING | WFE_LOWEST_SEQ);
 883         if (unlikely(!e))
 884                 return;
 885
 886         while (read_original_sector(wc, e) < end) {
 887                 struct rb_node *node = rb_next(&e->rb_node);
 888
 889                 if (likely(!e->write_in_progress)) {
 890                         if (!discarded_something) {
 891                                 if (!WC_MODE_PMEM(wc)) {
 892                                         writecache_wait_for_ios(wc, READ);
 893                                         writecache_wait_for_ios(wc, WRITE);
 894                                 }
 895                                 discarded_something = true;
 896                         }
 897                         if (!writecache_entry_is_committed(wc, e))
 898                                 wc->uncommitted_blocks--;
 899                         writecache_free_entry(wc, e);
 900                 }
 901
 902                 if (unlikely(!node))
 903                         break;
 904
 905                 e = container_of(node, struct wc_entry, rb_node);
 906         }
 907
 908         if (discarded_something)
 909                 writecache_commit_flushed(wc, false);
 910 }
 911
 912 static bool writecache_wait_for_writeback(struct dm_writecache *wc)
 913 {
 914         if (wc->writeback_size) {
 915                 writecache_wait_on_freelist(wc);
 916                 return true;
 917         }
 918         return false;
 919 }
 920
 921 static void writecache_suspend(struct dm_target *ti)
 922 {
 923         struct dm_writecache *wc = ti->private;
 924         bool flush_on_suspend;
 925
 926         del_timer_sync(&wc->autocommit_timer);
 927         del_timer_sync(&wc->max_age_timer);
 928
 929         wc_lock(wc);
 930         writecache_flush(wc);
 931         flush_on_suspend = wc->flush_on_suspend;
 932         if (flush_on_suspend) {
 933                 wc->flush_on_suspend = false;
 934                 wc->writeback_all++;
 935                 queue_work(wc->writeback_wq, &wc->writeback_work);
 936         }
 937         wc_unlock(wc);
 938
 939         drain_workqueue(wc->writeback_wq);
 940
 941         wc_lock(wc);
 942         if (flush_on_suspend)
 943                 wc->writeback_all--;
 944         while (writecache_wait_for_writeback(wc));
 945
 946         if (WC_MODE_PMEM(wc))
 947                 persistent_memory_flush_cache(wc->memory_map, wc->memory_map_size);
 948
 949         writecache_poison_lists(wc);
 950
 951         wc_unlock(wc);
 952 }
 953
 954 static int writecache_alloc_entries(struct dm_writecache *wc)
 955 {
 956         size_t b;
 957
 958         if (wc->entries)
 959                 return 0;
 960         wc->entries = vmalloc(array_size(sizeof(struct wc_entry), wc->n_blocks));
 961         if (!wc->entries)
 962                 return -ENOMEM;
 963         for (b = 0; b < wc->n_blocks; b++) {
 964                 struct wc_entry *e = &wc->entries[b];
 965                 e->index = b;
 966                 e->write_in_progress = false;
 967                 cond_resched();
 968         }
 969
 970         return 0;
 971 }
 972
 973 static int writecache_read_metadata(struct dm_writecache *wc, sector_t n_sectors)
 974 {
 975         struct dm_io_region region;
 976         struct dm_io_request req;
 977
 978         region.bdev = wc->ssd_dev->bdev;
 979         region.sector = wc->start_sector;
 980         region.count = n_sectors;
 981         req.bi_opf = REQ_OP_READ | REQ_SYNC;
 982         req.mem.type = DM_IO_VMA;
 983         req.mem.ptr.vma = (char *)wc->memory_map;
 984         req.client = wc->dm_io;
 985         req.notify.fn = NULL;
 986
 987         return dm_io(&req, 1, &region, NULL);
 988 }
 989
 990 static void writecache_resume(struct dm_target *ti)
 991 {
 992         struct dm_writecache *wc = ti->private;
 993         size_t b;
 994         bool need_flush = false;
 995         __le64 sb_seq_count;
 996         int r;
 997
 998         wc_lock(wc);
 999
1000         wc->data_device_sectors = bdev_nr_sectors(wc->dev->bdev);
1001
1002         if (WC_MODE_PMEM(wc)) {
1003                 persistent_memory_invalidate_cache(wc->memory_map, wc->memory_map_size);
1004         } else {
1005                 r = writecache_read_metadata(wc, wc->metadata_sectors);
1006                 if (r) {
1007                         size_t sb_entries_offset;
1008                         writecache_error(wc, r, "unable to read metadata: %d", r);
1009                         sb_entries_offset = offsetof(struct wc_memory_superblock, entries);
1010                         memset((char *)wc->memory_map + sb_entries_offset, -1,
1011                                (wc->metadata_sectors << SECTOR_SHIFT) - sb_entries_offset);
1012                 }
1013         }
1014
1015         wc->tree = RB_ROOT;
1016         INIT_LIST_HEAD(&wc->lru);
1017         if (WC_MODE_SORT_FREELIST(wc)) {
1018                 wc->freetree = RB_ROOT;
1019                 wc->current_free = NULL;
1020         } else {
1021                 INIT_LIST_HEAD(&wc->freelist);
1022         }
1023         wc->freelist_size = 0;
1024
1025         r = copy_mc_to_kernel(&sb_seq_count, &sb(wc)->seq_count,
1026                               sizeof(uint64_t));
1027         if (r) {
1028                 writecache_error(wc, r, "hardware memory error when reading superblock: %d", r);
1029                 sb_seq_count = cpu_to_le64(0);
1030         }
1031         wc->seq_count = le64_to_cpu(sb_seq_count);
1032
1033 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
1034         for (b = 0; b < wc->n_blocks; b++) {
1035                 struct wc_entry *e = &wc->entries[b];
1036                 struct wc_memory_entry wme;
1037                 if (writecache_has_error(wc)) {
1038                         e->original_sector = -1;
1039                         e->seq_count = -1;
1040                         continue;
1041                 }
1042                 r = copy_mc_to_kernel(&wme, memory_entry(wc, e),
1043                                       sizeof(struct wc_memory_entry));
1044                 if (r) {
1045                         writecache_error(wc, r, "hardware memory error when reading metadata entry %lu: %d",
1046                                          (unsigned long)b, r);
1047                         e->original_sector = -1;
1048                         e->seq_count = -1;
1049                 } else {
1050                         e->original_sector = le64_to_cpu(wme.original_sector);
1051                         e->seq_count = le64_to_cpu(wme.seq_count);
1052                 }
1053                 cond_resched();
1054         }
1055 #endif
1056         for (b = 0; b < wc->n_blocks; b++) {
1057                 struct wc_entry *e = &wc->entries[b];
1058                 if (!writecache_entry_is_committed(wc, e)) {
1059                         if (read_seq_count(wc, e) != -1) {
1060 erase_this:
1061                                 clear_seq_count(wc, e);
1062                                 need_flush = true;
1063                         }
1064                         writecache_add_to_freelist(wc, e);
1065                 } else {
1066                         struct wc_entry *old;
1067
1068                         old = writecache_find_entry(wc, read_original_sector(wc, e), 0);
1069                         if (!old) {
1070                                 writecache_insert_entry(wc, e);
1071                         } else {
1072                                 if (read_seq_count(wc, old) == read_seq_count(wc, e)) {
1073                                         writecache_error(wc, -EINVAL,
1074                                                  "two identical entries, position %llu, sector %llu, sequence %llu",
1075                                                  (unsigned long long)b, (unsigned long long)read_original_sector(wc, e),
1076                                                  (unsigned long long)read_seq_count(wc, e));
1077                                 }
1078                                 if (read_seq_count(wc, old) > read_seq_count(wc, e)) {
1079                                         goto erase_this;
1080                                 } else {
1081                                         writecache_free_entry(wc, old);
1082                                         writecache_insert_entry(wc, e);
1083                                         need_flush = true;
1084                                 }
1085                         }
1086                 }
1087                 cond_resched();
1088         }
1089
1090         if (need_flush) {
1091                 writecache_flush_all_metadata(wc);
1092                 writecache_commit_flushed(wc, false);
1093         }
1094
1095         writecache_verify_watermark(wc);
1096
1097         if (wc->max_age != MAX_AGE_UNSPECIFIED)
1098                 mod_timer(&wc->max_age_timer, jiffies + wc->max_age / MAX_AGE_DIV);
1099
1100         wc_unlock(wc);
1101 }
1102
1103 static int process_flush_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1104 {
1105         if (argc != 1)
1106                 return -EINVAL;
1107
1108         wc_lock(wc);
1109         if (dm_suspended(wc->ti)) {
1110                 wc_unlock(wc);
1111                 return -EBUSY;
1112         }
1113         if (writecache_has_error(wc)) {
1114                 wc_unlock(wc);
1115                 return -EIO;
1116         }
1117
1118         writecache_flush(wc);
1119         wc->writeback_all++;
1120         queue_work(wc->writeback_wq, &wc->writeback_work);
1121         wc_unlock(wc);
1122
1123         flush_workqueue(wc->writeback_wq);
1124
1125         wc_lock(wc);
1126         wc->writeback_all--;
1127         if (writecache_has_error(wc)) {
1128                 wc_unlock(wc);
1129                 return -EIO;
1130         }
1131         wc_unlock(wc);
1132
1133         return 0;
1134 }
1135
1136 static int process_flush_on_suspend_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1137 {
1138         if (argc != 1)
1139                 return -EINVAL;
1140
1141         wc_lock(wc);
1142         wc->flush_on_suspend = true;
1143         wc_unlock(wc);
1144
1145         return 0;
1146 }
1147
1148 static void activate_cleaner(struct dm_writecache *wc)
1149 {
1150         wc->flush_on_suspend = true;
1151         wc->cleaner = true;
1152         wc->freelist_high_watermark = wc->n_blocks;
1153         wc->freelist_low_watermark = wc->n_blocks;
1154 }
1155
1156 static int process_cleaner_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1157 {
1158         if (argc != 1)
1159                 return -EINVAL;
1160
1161         wc_lock(wc);
1162         activate_cleaner(wc);
1163         if (!dm_suspended(wc->ti))
1164                 writecache_verify_watermark(wc);
1165         wc_unlock(wc);
1166
1167         return 0;
1168 }
1169
1170 static int process_clear_stats_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1171 {
1172         if (argc != 1)
1173                 return -EINVAL;
1174
1175         wc_lock(wc);
1176         memset(&wc->stats, 0, sizeof wc->stats);
1177         wc_unlock(wc);
1178
1179         return 0;
1180 }
1181
1182 static int writecache_message(struct dm_target *ti, unsigned argc, char **argv,
1183                               char *result, unsigned maxlen)
1184 {
1185         int r = -EINVAL;
1186         struct dm_writecache *wc = ti->private;
1187
1188         if (!strcasecmp(argv[0], "flush"))
1189                 r = process_flush_mesg(argc, argv, wc);
1190         else if (!strcasecmp(argv[0], "flush_on_suspend"))
1191                 r = process_flush_on_suspend_mesg(argc, argv, wc);
1192         else if (!strcasecmp(argv[0], "cleaner"))
1193                 r = process_cleaner_mesg(argc, argv, wc);
1194         else if (!strcasecmp(argv[0], "clear_stats"))
1195                 r = process_clear_stats_mesg(argc, argv, wc);
1196         else
1197                 DMERR("unrecognised message received: %s", argv[0]);
1198
1199         return r;
1200 }
1201
1202 static void memcpy_flushcache_optimized(void *dest, void *source, size_t size)
1203 {
1204         /*
1205          * clflushopt performs better with block size 1024, 2048, 4096
1206          * non-temporal stores perform better with block size 512
1207          *
1208          * block size   512             1024            2048            4096
1209          * movnti       496 MB/s        642 MB/s        725 MB/s        744 MB/s
1210          * clflushopt   373 MB/s        688 MB/s        1.1 GB/s        1.2 GB/s
1211          *
1212          * We see that movnti performs better for 512-byte blocks, and
1213          * clflushopt performs better for 1024-byte and larger blocks. So, we
1214          * prefer clflushopt for sizes >= 768.
1215          *
1216          * NOTE: this happens to be the case now (with dm-writecache's single
1217          * threaded model) but re-evaluate this once memcpy_flushcache() is
1218          * enabled to use movdir64b which might invalidate this performance
1219          * advantage seen with cache-allocating-writes plus flushing.
1220          */
1221 #ifdef CONFIG_X86
1222         if (static_cpu_has(X86_FEATURE_CLFLUSHOPT) &&
1223             likely(boot_cpu_data.x86_clflush_size == 64) &&
1224             likely(size >= 768)) {
1225                 do {
1226                         memcpy((void *)dest, (void *)source, 64);
1227                         clflushopt((void *)dest);
1228                         dest += 64;
1229                         source += 64;
1230                         size -= 64;
1231                 } while (size >= 64);
1232                 return;
1233         }
1234 #endif
1235         memcpy_flushcache(dest, source, size);
1236 }
1237
1238 static void bio_copy_block(struct dm_writecache *wc, struct bio *bio, void *data)
1239 {
1240         void *buf;
1241         unsigned size;
1242         int rw = bio_data_dir(bio);
1243         unsigned remaining_size = wc->block_size;
1244
1245         do {
1246                 struct bio_vec bv = bio_iter_iovec(bio, bio->bi_iter);
1247                 buf = bvec_kmap_local(&bv);
1248                 size = bv.bv_len;
1249                 if (unlikely(size > remaining_size))
1250                         size = remaining_size;
1251
1252                 if (rw == READ) {
1253                         int r;
1254                         r = copy_mc_to_kernel(buf, data, size);
1255                         flush_dcache_page(bio_page(bio));
1256                         if (unlikely(r)) {
1257                                 writecache_error(wc, r, "hardware memory error when reading data: %d", r);
1258                                 bio->bi_status = BLK_STS_IOERR;
1259                         }
1260                 } else {
1261                         flush_dcache_page(bio_page(bio));
1262                         memcpy_flushcache_optimized(data, buf, size);
1263                 }
1264
1265                 kunmap_local(buf);
1266
1267                 data = (char *)data + size;
1268                 remaining_size -= size;
1269                 bio_advance(bio, size);
1270         } while (unlikely(remaining_size));
1271 }
1272
1273 static int writecache_flush_thread(void *data)
1274 {
1275         struct dm_writecache *wc = data;
1276
1277         while (1) {
1278                 struct bio *bio;
1279
1280                 wc_lock(wc);
1281                 bio = bio_list_pop(&wc->flush_list);
1282                 if (!bio) {
1283                         set_current_state(TASK_INTERRUPTIBLE);
1284                         wc_unlock(wc);
1285
1286                         if (unlikely(kthread_should_stop())) {
1287                                 set_current_state(TASK_RUNNING);
1288                                 break;
1289                         }
1290
1291                         schedule();
1292                         continue;
1293                 }
1294
1295                 if (bio_op(bio) == REQ_OP_DISCARD) {
1296                         writecache_discard(wc, bio->bi_iter.bi_sector,
1297                                            bio_end_sector(bio));
1298                         wc_unlock(wc);
1299                         bio_set_dev(bio, wc->dev->bdev);
1300                         submit_bio_noacct(bio);
1301                 } else {
1302                         writecache_flush(wc);
1303                         wc_unlock(wc);
1304                         if (writecache_has_error(wc))
1305                                 bio->bi_status = BLK_STS_IOERR;
1306                         bio_endio(bio);
1307                 }
1308         }
1309
1310         return 0;
1311 }
1312
1313 static void writecache_offload_bio(struct dm_writecache *wc, struct bio *bio)
1314 {
1315         if (bio_list_empty(&wc->flush_list))
1316                 wake_up_process(wc->flush_thread);
1317         bio_list_add(&wc->flush_list, bio);
1318 }
1319
1320 enum wc_map_op {
1321         WC_MAP_SUBMIT,
1322         WC_MAP_REMAP,
1323         WC_MAP_REMAP_ORIGIN,
1324         WC_MAP_RETURN,
1325         WC_MAP_ERROR,
1326 };
1327
1328 static void writecache_map_remap_origin(struct dm_writecache *wc, struct bio *bio,
1329                                         struct wc_entry *e)
1330 {
1331         if (e) {
1332                 sector_t next_boundary =
1333                         read_original_sector(wc, e) - bio->bi_iter.bi_sector;
1334                 if (next_boundary < bio->bi_iter.bi_size >> SECTOR_SHIFT)
1335                         dm_accept_partial_bio(bio, next_boundary);
1336         }
1337 }
1338
1339 static enum wc_map_op writecache_map_read(struct dm_writecache *wc, struct bio *bio)
1340 {
1341         enum wc_map_op map_op;
1342         struct wc_entry *e;
1343
1344 read_next_block:
1345         wc->stats.reads++;
1346         e = writecache_find_entry(wc, bio->bi_iter.bi_sector, WFE_RETURN_FOLLOWING);
1347         if (e && read_original_sector(wc, e) == bio->bi_iter.bi_sector) {
1348                 wc->stats.read_hits++;
1349                 if (WC_MODE_PMEM(wc)) {
1350                         bio_copy_block(wc, bio, memory_data(wc, e));
1351                         if (bio->bi_iter.bi_size)
1352                                 goto read_next_block;
1353                         map_op = WC_MAP_SUBMIT;
1354                 } else {
1355                         dm_accept_partial_bio(bio, wc->block_size >> SECTOR_SHIFT);
1356                         bio_set_dev(bio, wc->ssd_dev->bdev);
1357                         bio->bi_iter.bi_sector = cache_sector(wc, e);
1358                         if (!writecache_entry_is_committed(wc, e))
1359                                 writecache_wait_for_ios(wc, WRITE);
1360                         map_op = WC_MAP_REMAP;
1361                 }
1362         } else {
1363                 writecache_map_remap_origin(wc, bio, e);
1364                 wc->stats.reads += (bio->bi_iter.bi_size - wc->block_size) >> wc->block_size_bits;
1365                 map_op = WC_MAP_REMAP_ORIGIN;
1366         }
1367
1368         return map_op;
1369 }
1370
1371 static void writecache_bio_copy_ssd(struct dm_writecache *wc, struct bio *bio,
1372                                     struct wc_entry *e, bool search_used)
1373 {
1374         unsigned bio_size = wc->block_size;
1375         sector_t start_cache_sec = cache_sector(wc, e);
1376         sector_t current_cache_sec = start_cache_sec + (bio_size >> SECTOR_SHIFT);
1377
1378         while (bio_size < bio->bi_iter.bi_size) {
1379                 if (!search_used) {
1380                         struct wc_entry *f = writecache_pop_from_freelist(wc, current_cache_sec);
1381                         if (!f)
1382                                 break;
1383                         write_original_sector_seq_count(wc, f, bio->bi_iter.bi_sector +
1384                                                         (bio_size >> SECTOR_SHIFT), wc->seq_count);
1385                         writecache_insert_entry(wc, f);
1386                         wc->uncommitted_blocks++;
1387                 } else {
1388                         struct wc_entry *f;
1389                         struct rb_node *next = rb_next(&e->rb_node);
1390                         if (!next)
1391                                 break;
1392                         f = container_of(next, struct wc_entry, rb_node);
1393                         if (f != e + 1)
1394                                 break;
1395                         if (read_original_sector(wc, f) !=
1396                             read_original_sector(wc, e) + (wc->block_size >> SECTOR_SHIFT))
1397                                 break;
1398                         if (unlikely(f->write_in_progress))
1399                                 break;
1400                         if (writecache_entry_is_committed(wc, f))
1401                                 wc->overwrote_committed = true;
1402                         e = f;
1403                 }
1404                 bio_size += wc->block_size;
1405                 current_cache_sec += wc->block_size >> SECTOR_SHIFT;
1406         }
1407
1408         bio_set_dev(bio, wc->ssd_dev->bdev);
1409         bio->bi_iter.bi_sector = start_cache_sec;
1410         dm_accept_partial_bio(bio, bio_size >> SECTOR_SHIFT);
1411
1412         wc->stats.writes += bio->bi_iter.bi_size >> wc->block_size_bits;
1413         wc->stats.writes_allocate += (bio->bi_iter.bi_size - wc->block_size) >> wc->block_size_bits;
1414
1415         if (unlikely(wc->uncommitted_blocks >= wc->autocommit_blocks)) {
1416                 wc->uncommitted_blocks = 0;
1417                 queue_work(wc->writeback_wq, &wc->flush_work);
1418         } else {
1419                 writecache_schedule_autocommit(wc);
1420         }
1421 }
1422
1423 static enum wc_map_op writecache_map_write(struct dm_writecache *wc, struct bio *bio)
1424 {
1425         struct wc_entry *e;
1426
1427         do {
1428                 bool found_entry = false;
1429                 bool search_used = false;
1430                 if (writecache_has_error(wc)) {
1431                         wc->stats.writes += bio->bi_iter.bi_size >> wc->block_size_bits;
1432                         return WC_MAP_ERROR;
1433                 }
1434                 e = writecache_find_entry(wc, bio->bi_iter.bi_sector, 0);
1435                 if (e) {
1436                         if (!writecache_entry_is_committed(wc, e)) {
1437                                 wc->stats.write_hits_uncommitted++;
1438                                 search_used = true;
1439                                 goto bio_copy;
1440                         }
1441                         wc->stats.write_hits_committed++;
1442                         if (!WC_MODE_PMEM(wc) && !e->write_in_progress) {
1443                                 wc->overwrote_committed = true;
1444                                 search_used = true;
1445                                 goto bio_copy;
1446                         }
1447                         found_entry = true;
1448                 } else {
1449                         if (unlikely(wc->cleaner) ||
1450                             (wc->metadata_only && !(bio->bi_opf & REQ_META)))
1451                                 goto direct_write;
1452                 }
1453                 e = writecache_pop_from_freelist(wc, (sector_t)-1);
1454                 if (unlikely(!e)) {
1455                         if (!WC_MODE_PMEM(wc) && !found_entry) {
1456 direct_write:
1457                                 e = writecache_find_entry(wc, bio->bi_iter.bi_sector, WFE_RETURN_FOLLOWING);
1458                                 writecache_map_remap_origin(wc, bio, e);
1459                                 wc->stats.writes_around += bio->bi_iter.bi_size >> wc->block_size_bits;
1460                                 wc->stats.writes += bio->bi_iter.bi_size >> wc->block_size_bits;
1461                                 return WC_MAP_REMAP_ORIGIN;
1462                         }
1463                         wc->stats.writes_blocked_on_freelist++;
1464                         writecache_wait_on_freelist(wc);
1465                         continue;
1466                 }
1467                 write_original_sector_seq_count(wc, e, bio->bi_iter.bi_sector, wc->seq_count);
1468                 writecache_insert_entry(wc, e);
1469                 wc->uncommitted_blocks++;
1470                 wc->stats.writes_allocate++;
1471 bio_copy:
1472                 if (WC_MODE_PMEM(wc)) {
1473                         bio_copy_block(wc, bio, memory_data(wc, e));
1474                         wc->stats.writes++;
1475                 } else {
1476                         writecache_bio_copy_ssd(wc, bio, e, search_used);
1477                         return WC_MAP_REMAP;
1478                 }
1479         } while (bio->bi_iter.bi_size);
1480
1481         if (unlikely(bio->bi_opf & REQ_FUA || wc->uncommitted_blocks >= wc->autocommit_blocks))
1482                 writecache_flush(wc);
1483         else
1484                 writecache_schedule_autocommit(wc);
1485
1486         return WC_MAP_SUBMIT;
1487 }
1488
1489 static enum wc_map_op writecache_map_flush(struct dm_writecache *wc, struct bio *bio)
1490 {
1491         if (writecache_has_error(wc))
1492                 return WC_MAP_ERROR;
1493
1494         if (WC_MODE_PMEM(wc)) {
1495                 wc->stats.flushes++;
1496                 writecache_flush(wc);
1497                 if (writecache_has_error(wc))
1498                         return WC_MAP_ERROR;
1499                 else if (unlikely(wc->cleaner) || unlikely(wc->metadata_only))
1500                         return WC_MAP_REMAP_ORIGIN;
1501                 return WC_MAP_SUBMIT;
1502         }
1503         /* SSD: */
1504         if (dm_bio_get_target_bio_nr(bio))
1505                 return WC_MAP_REMAP_ORIGIN;
1506         wc->stats.flushes++;
1507         writecache_offload_bio(wc, bio);
1508         return WC_MAP_RETURN;
1509 }
1510
1511 static enum wc_map_op writecache_map_discard(struct dm_writecache *wc, struct bio *bio)
1512 {
1513         wc->stats.discards += bio->bi_iter.bi_size >> wc->block_size_bits;
1514
1515         if (writecache_has_error(wc))
1516                 return WC_MAP_ERROR;
1517
1518         if (WC_MODE_PMEM(wc)) {
1519                 writecache_discard(wc, bio->bi_iter.bi_sector, bio_end_sector(bio));
1520                 return WC_MAP_REMAP_ORIGIN;
1521         }
1522         /* SSD: */
1523         writecache_offload_bio(wc, bio);
1524         return WC_MAP_RETURN;
1525 }
1526
1527 static int writecache_map(struct dm_target *ti, struct bio *bio)
1528 {
1529         struct dm_writecache *wc = ti->private;
1530         enum wc_map_op map_op;
1531
1532         bio->bi_private = NULL;
1533
1534         wc_lock(wc);
1535
1536         if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1537                 map_op = writecache_map_flush(wc, bio);
1538                 goto done;
1539         }
1540
1541         bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
1542
1543         if (unlikely((((unsigned)bio->bi_iter.bi_sector | bio_sectors(bio)) &
1544                                 (wc->block_size / 512 - 1)) != 0)) {
1545                 DMERR("I/O is not aligned, sector %llu, size %u, block size %u",
1546                       (unsigned long long)bio->bi_iter.bi_sector,
1547                       bio->bi_iter.bi_size, wc->block_size);
1548                 map_op = WC_MAP_ERROR;
1549                 goto done;
1550         }
1551
1552         if (unlikely(bio_op(bio) == REQ_OP_DISCARD)) {
1553                 map_op = writecache_map_discard(wc, bio);
1554                 goto done;
1555         }
1556
1557         if (bio_data_dir(bio) == READ)
1558                 map_op = writecache_map_read(wc, bio);
1559         else
1560                 map_op = writecache_map_write(wc, bio);
1561 done:
1562         switch (map_op) {
1563         case WC_MAP_REMAP_ORIGIN:
1564                 if (likely(wc->pause != 0)) {
1565                         if (bio_op(bio) == REQ_OP_WRITE) {
1566                                 dm_iot_io_begin(&wc->iot, 1);
1567                                 bio->bi_private = (void *)2;
1568                         }
1569                 }
1570                 bio_set_dev(bio, wc->dev->bdev);
1571                 wc_unlock(wc);
1572                 return DM_MAPIO_REMAPPED;
1573
1574         case WC_MAP_REMAP:
1575                 /* make sure that writecache_end_io decrements bio_in_progress: */
1576                 bio->bi_private = (void *)1;
1577                 atomic_inc(&wc->bio_in_progress[bio_data_dir(bio)]);
1578                 wc_unlock(wc);
1579                 return DM_MAPIO_REMAPPED;
1580
1581         case WC_MAP_SUBMIT:
1582                 wc_unlock(wc);
1583                 bio_endio(bio);
1584                 return DM_MAPIO_SUBMITTED;
1585
1586         case WC_MAP_RETURN:
1587                 wc_unlock(wc);
1588                 return DM_MAPIO_SUBMITTED;
1589
1590         case WC_MAP_ERROR:
1591                 wc_unlock(wc);
1592                 bio_io_error(bio);
1593                 return DM_MAPIO_SUBMITTED;
1594
1595         default:
1596                 BUG();
1597                 wc_unlock(wc);
1598                 return DM_MAPIO_KILL;
1599         }
1600 }
1601
1602 static int writecache_end_io(struct dm_target *ti, struct bio *bio, blk_status_t *status)
1603 {
1604         struct dm_writecache *wc = ti->private;
1605
1606         if (bio->bi_private == (void *)1) {
1607                 int dir = bio_data_dir(bio);
1608                 if (atomic_dec_and_test(&wc->bio_in_progress[dir]))
1609                         if (unlikely(waitqueue_active(&wc->bio_in_progress_wait[dir])))
1610                                 wake_up(&wc->bio_in_progress_wait[dir]);
1611         } else if (bio->bi_private == (void *)2) {
1612                 dm_iot_io_end(&wc->iot, 1);
1613         }
1614         return 0;
1615 }
1616
1617 static int writecache_iterate_devices(struct dm_target *ti,
1618                                       iterate_devices_callout_fn fn, void *data)
1619 {
1620         struct dm_writecache *wc = ti->private;
1621
1622         return fn(ti, wc->dev, 0, ti->len, data);
1623 }
1624
1625 static void writecache_io_hints(struct dm_target *ti, struct queue_limits *limits)
1626 {
1627         struct dm_writecache *wc = ti->private;
1628
1629         if (limits->logical_block_size < wc->block_size)
1630                 limits->logical_block_size = wc->block_size;
1631
1632         if (limits->physical_block_size < wc->block_size)
1633                 limits->physical_block_size = wc->block_size;
1634
1635         if (limits->io_min < wc->block_size)
1636                 limits->io_min = wc->block_size;
1637 }
1638
1639
1640 static void writecache_writeback_endio(struct bio *bio)
1641 {
1642         struct writeback_struct *wb = container_of(bio, struct writeback_struct, bio);
1643         struct dm_writecache *wc = wb->wc;
1644         unsigned long flags;
1645
1646         raw_spin_lock_irqsave(&wc->endio_list_lock, flags);
1647         if (unlikely(list_empty(&wc->endio_list)))
1648                 wake_up_process(wc->endio_thread);
1649         list_add_tail(&wb->endio_entry, &wc->endio_list);
1650         raw_spin_unlock_irqrestore(&wc->endio_list_lock, flags);
1651 }
1652
1653 static void writecache_copy_endio(int read_err, unsigned long write_err, void *ptr)
1654 {
1655         struct copy_struct *c = ptr;
1656         struct dm_writecache *wc = c->wc;
1657
1658         c->error = likely(!(read_err | write_err)) ? 0 : -EIO;
1659
1660         raw_spin_lock_irq(&wc->endio_list_lock);
1661         if (unlikely(list_empty(&wc->endio_list)))
1662                 wake_up_process(wc->endio_thread);
1663         list_add_tail(&c->endio_entry, &wc->endio_list);
1664         raw_spin_unlock_irq(&wc->endio_list_lock);
1665 }
1666
1667 static void __writecache_endio_pmem(struct dm_writecache *wc, struct list_head *list)
1668 {
1669         unsigned i;
1670         struct writeback_struct *wb;
1671         struct wc_entry *e;
1672         unsigned long n_walked = 0;
1673
1674         do {
1675                 wb = list_entry(list->next, struct writeback_struct, endio_entry);
1676                 list_del(&wb->endio_entry);
1677
1678                 if (unlikely(wb->bio.bi_status != BLK_STS_OK))
1679                         writecache_error(wc, blk_status_to_errno(wb->bio.bi_status),
1680                                         "write error %d", wb->bio.bi_status);
1681                 i = 0;
1682                 do {
1683                         e = wb->wc_list[i];
1684                         BUG_ON(!e->write_in_progress);
1685                         e->write_in_progress = false;
1686                         INIT_LIST_HEAD(&e->lru);
1687                         if (!writecache_has_error(wc))
1688                                 writecache_free_entry(wc, e);
1689                         BUG_ON(!wc->writeback_size);
1690                         wc->writeback_size--;
1691                         n_walked++;
1692                         if (unlikely(n_walked >= ENDIO_LATENCY)) {
1693                                 writecache_commit_flushed(wc, false);
1694                                 wc_unlock(wc);
1695                                 wc_lock(wc);
1696                                 n_walked = 0;
1697                         }
1698                 } while (++i < wb->wc_list_n);
1699
1700                 if (wb->wc_list != wb->wc_list_inline)
1701                         kfree(wb->wc_list);
1702                 bio_put(&wb->bio);
1703         } while (!list_empty(list));
1704 }
1705
1706 static void __writecache_endio_ssd(struct dm_writecache *wc, struct list_head *list)
1707 {
1708         struct copy_struct *c;
1709         struct wc_entry *e;
1710
1711         do {
1712                 c = list_entry(list->next, struct copy_struct, endio_entry);
1713                 list_del(&c->endio_entry);
1714
1715                 if (unlikely(c->error))
1716                         writecache_error(wc, c->error, "copy error");
1717
1718                 e = c->e;
1719                 do {
1720                         BUG_ON(!e->write_in_progress);
1721                         e->write_in_progress = false;
1722                         INIT_LIST_HEAD(&e->lru);
1723                         if (!writecache_has_error(wc))
1724                                 writecache_free_entry(wc, e);
1725
1726                         BUG_ON(!wc->writeback_size);
1727                         wc->writeback_size--;
1728                         e++;
1729                 } while (--c->n_entries);
1730                 mempool_free(c, &wc->copy_pool);
1731         } while (!list_empty(list));
1732 }
1733
1734 static int writecache_endio_thread(void *data)
1735 {
1736         struct dm_writecache *wc = data;
1737
1738         while (1) {
1739                 struct list_head list;
1740
1741                 raw_spin_lock_irq(&wc->endio_list_lock);
1742                 if (!list_empty(&wc->endio_list))
1743                         goto pop_from_list;
1744                 set_current_state(TASK_INTERRUPTIBLE);
1745                 raw_spin_unlock_irq(&wc->endio_list_lock);
1746
1747                 if (unlikely(kthread_should_stop())) {
1748                         set_current_state(TASK_RUNNING);
1749                         break;
1750                 }
1751
1752                 schedule();
1753
1754                 continue;
1755
1756 pop_from_list:
1757                 list = wc->endio_list;
1758                 list.next->prev = list.prev->next = &list;
1759                 INIT_LIST_HEAD(&wc->endio_list);
1760                 raw_spin_unlock_irq(&wc->endio_list_lock);
1761
1762                 if (!WC_MODE_FUA(wc))
1763                         writecache_disk_flush(wc, wc->dev);
1764
1765                 wc_lock(wc);
1766
1767                 if (WC_MODE_PMEM(wc)) {
1768                         __writecache_endio_pmem(wc, &list);
1769                 } else {
1770                         __writecache_endio_ssd(wc, &list);
1771                         writecache_wait_for_ios(wc, READ);
1772                 }
1773
1774                 writecache_commit_flushed(wc, false);
1775
1776                 wc_unlock(wc);
1777         }
1778
1779         return 0;
1780 }
1781
1782 static bool wc_add_block(struct writeback_struct *wb, struct wc_entry *e)
1783 {
1784         struct dm_writecache *wc = wb->wc;
1785         unsigned block_size = wc->block_size;
1786         void *address = memory_data(wc, e);
1787
1788         persistent_memory_flush_cache(address, block_size);
1789
1790         if (unlikely(bio_end_sector(&wb->bio) >= wc->data_device_sectors))
1791                 return true;
1792
1793         return bio_add_page(&wb->bio, persistent_memory_page(address),
1794                             block_size, persistent_memory_page_offset(address)) != 0;
1795 }
1796
1797 struct writeback_list {
1798         struct list_head list;
1799         size_t size;
1800 };
1801
1802 static void __writeback_throttle(struct dm_writecache *wc, struct writeback_list *wbl)
1803 {
1804         if (unlikely(wc->max_writeback_jobs)) {
1805                 if (READ_ONCE(wc->writeback_size) - wbl->size >= wc->max_writeback_jobs) {
1806                         wc_lock(wc);
1807                         while (wc->writeback_size - wbl->size >= wc->max_writeback_jobs)
1808                                 writecache_wait_on_freelist(wc);
1809                         wc_unlock(wc);
1810                 }
1811         }
1812         cond_resched();
1813 }
1814
1815 static void __writecache_writeback_pmem(struct dm_writecache *wc, struct writeback_list *wbl)
1816 {
1817         struct wc_entry *e, *f;
1818         struct bio *bio;
1819         struct writeback_struct *wb;
1820         unsigned max_pages;
1821
1822         while (wbl->size) {
1823                 wbl->size--;
1824                 e = container_of(wbl->list.prev, struct wc_entry, lru);
1825                 list_del(&e->lru);
1826
1827                 max_pages = e->wc_list_contiguous;
1828
1829                 bio = bio_alloc_bioset(wc->dev->bdev, max_pages, REQ_OP_WRITE,
1830                                        GFP_NOIO, &wc->bio_set);
1831                 wb = container_of(bio, struct writeback_struct, bio);
1832                 wb->wc = wc;
1833                 bio->bi_end_io = writecache_writeback_endio;
1834                 bio->bi_iter.bi_sector = read_original_sector(wc, e);
1835                 if (max_pages <= WB_LIST_INLINE ||
1836                     unlikely(!(wb->wc_list = kmalloc_array(max_pages, sizeof(struct wc_entry *),
1837                                                            GFP_NOIO | __GFP_NORETRY |
1838                                                            __GFP_NOMEMALLOC | __GFP_NOWARN)))) {
1839                         wb->wc_list = wb->wc_list_inline;
1840                         max_pages = WB_LIST_INLINE;
1841                 }
1842
1843                 BUG_ON(!wc_add_block(wb, e));
1844
1845                 wb->wc_list[0] = e;
1846                 wb->wc_list_n = 1;
1847
1848                 while (wbl->size && wb->wc_list_n < max_pages) {
1849                         f = container_of(wbl->list.prev, struct wc_entry, lru);
1850                         if (read_original_sector(wc, f) !=
1851                             read_original_sector(wc, e) + (wc->block_size >> SECTOR_SHIFT))
1852                                 break;
1853                         if (!wc_add_block(wb, f))
1854                                 break;
1855                         wbl->size--;
1856                         list_del(&f->lru);
1857                         wb->wc_list[wb->wc_list_n++] = f;
1858                         e = f;
1859                 }
1860                 if (WC_MODE_FUA(wc))
1861                         bio->bi_opf |= REQ_FUA;
1862                 if (writecache_has_error(wc)) {
1863                         bio->bi_status = BLK_STS_IOERR;
1864                         bio_endio(bio);
1865                 } else if (unlikely(!bio_sectors(bio))) {
1866                         bio->bi_status = BLK_STS_OK;
1867                         bio_endio(bio);
1868                 } else {
1869                         submit_bio(bio);
1870                 }
1871
1872                 __writeback_throttle(wc, wbl);
1873         }
1874 }
1875
1876 static void __writecache_writeback_ssd(struct dm_writecache *wc, struct writeback_list *wbl)
1877 {
1878         struct wc_entry *e, *f;
1879         struct dm_io_region from, to;
1880         struct copy_struct *c;
1881
1882         while (wbl->size) {
1883                 unsigned n_sectors;
1884
1885                 wbl->size--;
1886                 e = container_of(wbl->list.prev, struct wc_entry, lru);
1887                 list_del(&e->lru);
1888
1889                 n_sectors = e->wc_list_contiguous << (wc->block_size_bits - SECTOR_SHIFT);
1890
1891                 from.bdev = wc->ssd_dev->bdev;
1892                 from.sector = cache_sector(wc, e);
1893                 from.count = n_sectors;
1894                 to.bdev = wc->dev->bdev;
1895                 to.sector = read_original_sector(wc, e);
1896                 to.count = n_sectors;
1897
1898                 c = mempool_alloc(&wc->copy_pool, GFP_NOIO);
1899                 c->wc = wc;
1900                 c->e = e;
1901                 c->n_entries = e->wc_list_contiguous;
1902
1903                 while ((n_sectors -= wc->block_size >> SECTOR_SHIFT)) {
1904                         wbl->size--;
1905                         f = container_of(wbl->list.prev, struct wc_entry, lru);
1906                         BUG_ON(f != e + 1);
1907                         list_del(&f->lru);
1908                         e = f;
1909                 }
1910
1911                 if (unlikely(to.sector + to.count > wc->data_device_sectors)) {
1912                         if (to.sector >= wc->data_device_sectors) {
1913                                 writecache_copy_endio(0, 0, c);
1914                                 continue;
1915                         }
1916                         from.count = to.count = wc->data_device_sectors - to.sector;
1917                 }
1918
1919                 dm_kcopyd_copy(wc->dm_kcopyd, &from, 1, &to, 0, writecache_copy_endio, c);
1920
1921                 __writeback_throttle(wc, wbl);
1922         }
1923 }
1924
1925 static void writecache_writeback(struct work_struct *work)
1926 {
1927         struct dm_writecache *wc = container_of(work, struct dm_writecache, writeback_work);
1928         struct blk_plug plug;
1929         struct wc_entry *f, *g, *e = NULL;
1930         struct rb_node *node, *next_node;
1931         struct list_head skipped;
1932         struct writeback_list wbl;
1933         unsigned long n_walked;
1934
1935         if (!WC_MODE_PMEM(wc)) {
1936                 /* Wait for any active kcopyd work on behalf of ssd writeback */
1937                 dm_kcopyd_client_flush(wc->dm_kcopyd);
1938         }
1939
1940         if (likely(wc->pause != 0)) {
1941                 while (1) {
1942                         unsigned long idle;
1943                         if (unlikely(wc->cleaner) || unlikely(wc->writeback_all) ||
1944                             unlikely(dm_suspended(wc->ti)))
1945                                 break;
1946                         idle = dm_iot_idle_time(&wc->iot);
1947                         if (idle >= wc->pause)
1948                                 break;
1949                         idle = wc->pause - idle;
1950                         if (idle > HZ)
1951                                 idle = HZ;
1952                         schedule_timeout_idle(idle);
1953                 }
1954         }
1955
1956         wc_lock(wc);
1957 restart:
1958         if (writecache_has_error(wc)) {
1959                 wc_unlock(wc);
1960                 return;
1961         }
1962
1963         if (unlikely(wc->writeback_all)) {
1964                 if (writecache_wait_for_writeback(wc))
1965                         goto restart;
1966         }
1967
1968         if (wc->overwrote_committed) {
1969                 writecache_wait_for_ios(wc, WRITE);
1970         }
1971
1972         n_walked = 0;
1973         INIT_LIST_HEAD(&skipped);
1974         INIT_LIST_HEAD(&wbl.list);
1975         wbl.size = 0;
1976         while (!list_empty(&wc->lru) &&
1977                (wc->writeback_all ||
1978                 wc->freelist_size + wc->writeback_size <= wc->freelist_low_watermark ||
1979                 (jiffies - container_of(wc->lru.prev, struct wc_entry, lru)->age >=
1980                  wc->max_age - wc->max_age / MAX_AGE_DIV))) {
1981
1982                 n_walked++;
1983                 if (unlikely(n_walked > WRITEBACK_LATENCY) &&
1984                     likely(!wc->writeback_all)) {
1985                         if (likely(!dm_suspended(wc->ti)))
1986                                 queue_work(wc->writeback_wq, &wc->writeback_work);
1987                         break;
1988                 }
1989
1990                 if (unlikely(wc->writeback_all)) {
1991                         if (unlikely(!e)) {
1992                                 writecache_flush(wc);
1993                                 e = container_of(rb_first(&wc->tree), struct wc_entry, rb_node);
1994                         } else
1995                                 e = g;
1996                 } else
1997                         e = container_of(wc->lru.prev, struct wc_entry, lru);
1998                 BUG_ON(e->write_in_progress);
1999                 if (unlikely(!writecache_entry_is_committed(wc, e))) {
2000                         writecache_flush(wc);
2001                 }
2002                 node = rb_prev(&e->rb_node);
2003                 if (node) {
2004                         f = container_of(node, struct wc_entry, rb_node);
2005                         if (unlikely(read_original_sector(wc, f) ==
2006                                      read_original_sector(wc, e))) {
2007                                 BUG_ON(!f->write_in_progress);
2008                                 list_move(&e->lru, &skipped);
2009                                 cond_resched();
2010                                 continue;
2011                         }
2012                 }
2013                 wc->writeback_size++;
2014                 list_move(&e->lru, &wbl.list);
2015                 wbl.size++;
2016                 e->write_in_progress = true;
2017                 e->wc_list_contiguous = 1;
2018
2019                 f = e;
2020
2021                 while (1) {
2022                         next_node = rb_next(&f->rb_node);
2023                         if (unlikely(!next_node))
2024                                 break;
2025                         g = container_of(next_node, struct wc_entry, rb_node);
2026                         if (unlikely(read_original_sector(wc, g) ==
2027                             read_original_sector(wc, f))) {
2028                                 f = g;
2029                                 continue;
2030                         }
2031                         if (read_original_sector(wc, g) !=
2032                             read_original_sector(wc, f) + (wc->block_size >> SECTOR_SHIFT))
2033                                 break;
2034                         if (unlikely(g->write_in_progress))
2035                                 break;
2036                         if (unlikely(!writecache_entry_is_committed(wc, g)))
2037                                 break;
2038
2039                         if (!WC_MODE_PMEM(wc)) {
2040                                 if (g != f + 1)
2041                                         break;
2042                         }
2043
2044                         n_walked++;
2045                         //if (unlikely(n_walked > WRITEBACK_LATENCY) && likely(!wc->writeback_all))
2046                         //      break;
2047
2048                         wc->writeback_size++;
2049                         list_move(&g->lru, &wbl.list);
2050                         wbl.size++;
2051                         g->write_in_progress = true;
2052                         g->wc_list_contiguous = BIO_MAX_VECS;
2053                         f = g;
2054                         e->wc_list_contiguous++;
2055                         if (unlikely(e->wc_list_contiguous == BIO_MAX_VECS)) {
2056                                 if (unlikely(wc->writeback_all)) {
2057                                         next_node = rb_next(&f->rb_node);
2058                                         if (likely(next_node))
2059                                                 g = container_of(next_node, struct wc_entry, rb_node);
2060                                 }
2061                                 break;
2062                         }
2063                 }
2064                 cond_resched();
2065         }
2066
2067         if (!list_empty(&skipped)) {
2068                 list_splice_tail(&skipped, &wc->lru);
2069                 /*
2070                  * If we didn't do any progress, we must wait until some
2071                  * writeback finishes to avoid burning CPU in a loop
2072                  */
2073                 if (unlikely(!wbl.size))
2074                         writecache_wait_for_writeback(wc);
2075         }
2076
2077         wc_unlock(wc);
2078
2079         blk_start_plug(&plug);
2080
2081         if (WC_MODE_PMEM(wc))
2082                 __writecache_writeback_pmem(wc, &wbl);
2083         else
2084                 __writecache_writeback_ssd(wc, &wbl);
2085
2086         blk_finish_plug(&plug);
2087
2088         if (unlikely(wc->writeback_all)) {
2089                 wc_lock(wc);
2090                 while (writecache_wait_for_writeback(wc));
2091                 wc_unlock(wc);
2092         }
2093 }
2094
2095 static int calculate_memory_size(uint64_t device_size, unsigned block_size,
2096                                  size_t *n_blocks_p, size_t *n_metadata_blocks_p)
2097 {
2098         uint64_t n_blocks, offset;
2099         struct wc_entry e;
2100
2101         n_blocks = device_size;
2102         do_div(n_blocks, block_size + sizeof(struct wc_memory_entry));
2103
2104         while (1) {
2105                 if (!n_blocks)
2106                         return -ENOSPC;
2107                 /* Verify the following entries[n_blocks] won't overflow */
2108                 if (n_blocks >= ((size_t)-sizeof(struct wc_memory_superblock) /
2109                                  sizeof(struct wc_memory_entry)))
2110                         return -EFBIG;
2111                 offset = offsetof(struct wc_memory_superblock, entries[n_blocks]);
2112                 offset = (offset + block_size - 1) & ~(uint64_t)(block_size - 1);
2113                 if (offset + n_blocks * block_size <= device_size)
2114                         break;
2115                 n_blocks--;
2116         }
2117
2118         /* check if the bit field overflows */
2119         e.index = n_blocks;
2120         if (e.index != n_blocks)
2121                 return -EFBIG;
2122
2123         if (n_blocks_p)
2124                 *n_blocks_p = n_blocks;
2125         if (n_metadata_blocks_p)
2126                 *n_metadata_blocks_p = offset >> __ffs(block_size);
2127         return 0;
2128 }
2129
2130 static int init_memory(struct dm_writecache *wc)
2131 {
2132         size_t b;
2133         int r;
2134
2135         r = calculate_memory_size(wc->memory_map_size, wc->block_size, &wc->n_blocks, NULL);
2136         if (r)
2137                 return r;
2138
2139         r = writecache_alloc_entries(wc);
2140         if (r)
2141                 return r;
2142
2143         for (b = 0; b < ARRAY_SIZE(sb(wc)->padding); b++)
2144                 pmem_assign(sb(wc)->padding[b], cpu_to_le64(0));
2145         pmem_assign(sb(wc)->version, cpu_to_le32(MEMORY_SUPERBLOCK_VERSION));
2146         pmem_assign(sb(wc)->block_size, cpu_to_le32(wc->block_size));
2147         pmem_assign(sb(wc)->n_blocks, cpu_to_le64(wc->n_blocks));
2148         pmem_assign(sb(wc)->seq_count, cpu_to_le64(0));
2149
2150         for (b = 0; b < wc->n_blocks; b++) {
2151                 write_original_sector_seq_count(wc, &wc->entries[b], -1, -1);
2152                 cond_resched();
2153         }
2154
2155         writecache_flush_all_metadata(wc);
2156         writecache_commit_flushed(wc, false);
2157         pmem_assign(sb(wc)->magic, cpu_to_le32(MEMORY_SUPERBLOCK_MAGIC));
2158         writecache_flush_region(wc, &sb(wc)->magic, sizeof sb(wc)->magic);
2159         writecache_commit_flushed(wc, false);
2160
2161         return 0;
2162 }
2163
2164 static void writecache_dtr(struct dm_target *ti)
2165 {
2166         struct dm_writecache *wc = ti->private;
2167
2168         if (!wc)
2169                 return;
2170
2171         if (wc->endio_thread)
2172                 kthread_stop(wc->endio_thread);
2173
2174         if (wc->flush_thread)
2175                 kthread_stop(wc->flush_thread);
2176
2177         bioset_exit(&wc->bio_set);
2178
2179         mempool_exit(&wc->copy_pool);
2180
2181         if (wc->writeback_wq)
2182                 destroy_workqueue(wc->writeback_wq);
2183
2184         if (wc->dev)
2185                 dm_put_device(ti, wc->dev);
2186
2187         if (wc->ssd_dev)
2188                 dm_put_device(ti, wc->ssd_dev);
2189
2190         vfree(wc->entries);
2191
2192         if (wc->memory_map) {
2193                 if (WC_MODE_PMEM(wc))
2194                         persistent_memory_release(wc);
2195                 else
2196                         vfree(wc->memory_map);
2197         }
2198
2199         if (wc->dm_kcopyd)
2200                 dm_kcopyd_client_destroy(wc->dm_kcopyd);
2201
2202         if (wc->dm_io)
2203                 dm_io_client_destroy(wc->dm_io);
2204
2205         vfree(wc->dirty_bitmap);
2206
2207         kfree(wc);
2208 }
2209
2210 static int writecache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2211 {
2212         struct dm_writecache *wc;
2213         struct dm_arg_set as;
2214         const char *string;
2215         unsigned opt_params;
2216         size_t offset, data_size;
2217         int i, r;
2218         char dummy;
2219         int high_wm_percent = HIGH_WATERMARK;
2220         int low_wm_percent = LOW_WATERMARK;
2221         uint64_t x;
2222         struct wc_memory_superblock s;
2223
2224         static struct dm_arg _args[] = {
2225                 {0, 18, "Invalid number of feature args"},
2226         };
2227
2228         as.argc = argc;
2229         as.argv = argv;
2230
2231         wc = kzalloc(sizeof(struct dm_writecache), GFP_KERNEL);
2232         if (!wc) {
2233                 ti->error = "Cannot allocate writecache structure";
2234                 r = -ENOMEM;
2235                 goto bad;
2236         }
2237         ti->private = wc;
2238         wc->ti = ti;
2239
2240         mutex_init(&wc->lock);
2241         wc->max_age = MAX_AGE_UNSPECIFIED;
2242         writecache_poison_lists(wc);
2243         init_waitqueue_head(&wc->freelist_wait);
2244         timer_setup(&wc->autocommit_timer, writecache_autocommit_timer, 0);
2245         timer_setup(&wc->max_age_timer, writecache_max_age_timer, 0);
2246
2247         for (i = 0; i < 2; i++) {
2248                 atomic_set(&wc->bio_in_progress[i], 0);
2249                 init_waitqueue_head(&wc->bio_in_progress_wait[i]);
2250         }
2251
2252         wc->dm_io = dm_io_client_create();
2253         if (IS_ERR(wc->dm_io)) {
2254                 r = PTR_ERR(wc->dm_io);
2255                 ti->error = "Unable to allocate dm-io client";
2256                 wc->dm_io = NULL;
2257                 goto bad;
2258         }
2259
2260         wc->writeback_wq = alloc_workqueue("writecache-writeback", WQ_MEM_RECLAIM, 1);
2261         if (!wc->writeback_wq) {
2262                 r = -ENOMEM;
2263                 ti->error = "Could not allocate writeback workqueue";
2264                 goto bad;
2265         }
2266         INIT_WORK(&wc->writeback_work, writecache_writeback);
2267         INIT_WORK(&wc->flush_work, writecache_flush_work);
2268
2269         dm_iot_init(&wc->iot);
2270
2271         raw_spin_lock_init(&wc->endio_list_lock);
2272         INIT_LIST_HEAD(&wc->endio_list);
2273         wc->endio_thread = kthread_run(writecache_endio_thread, wc, "writecache_endio");
2274         if (IS_ERR(wc->endio_thread)) {
2275                 r = PTR_ERR(wc->endio_thread);
2276                 wc->endio_thread = NULL;
2277                 ti->error = "Couldn't spawn endio thread";
2278                 goto bad;
2279         }
2280
2281         /*
2282          * Parse the mode (pmem or ssd)
2283          */
2284         string = dm_shift_arg(&as);
2285         if (!string)
2286                 goto bad_arguments;
2287
2288         if (!strcasecmp(string, "s")) {
2289                 wc->pmem_mode = false;
2290         } else if (!strcasecmp(string, "p")) {
2291 #ifdef DM_WRITECACHE_HAS_PMEM
2292                 wc->pmem_mode = true;
2293                 wc->writeback_fua = true;
2294 #else
2295                 /*
2296                  * If the architecture doesn't support persistent memory or
2297                  * the kernel doesn't support any DAX drivers, this driver can
2298                  * only be used in SSD-only mode.
2299                  */
2300                 r = -EOPNOTSUPP;
2301                 ti->error = "Persistent memory or DAX not supported on this system";
2302                 goto bad;
2303 #endif
2304         } else {
2305                 goto bad_arguments;
2306         }
2307
2308         if (WC_MODE_PMEM(wc)) {
2309                 r = bioset_init(&wc->bio_set, BIO_POOL_SIZE,
2310                                 offsetof(struct writeback_struct, bio),
2311                                 BIOSET_NEED_BVECS);
2312                 if (r) {
2313                         ti->error = "Could not allocate bio set";
2314                         goto bad;
2315                 }
2316         } else {
2317                 wc->pause = PAUSE_WRITEBACK;
2318                 r = mempool_init_kmalloc_pool(&wc->copy_pool, 1, sizeof(struct copy_struct));
2319                 if (r) {
2320                         ti->error = "Could not allocate mempool";
2321                         goto bad;
2322                 }
2323         }
2324
2325         /*
2326          * Parse the origin data device
2327          */
2328         string = dm_shift_arg(&as);
2329         if (!string)
2330                 goto bad_arguments;
2331         r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->dev);
2332         if (r) {
2333                 ti->error = "Origin data device lookup failed";
2334                 goto bad;
2335         }
2336
2337         /*
2338          * Parse cache data device (be it pmem or ssd)
2339          */
2340         string = dm_shift_arg(&as);
2341         if (!string)
2342                 goto bad_arguments;
2343
2344         r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->ssd_dev);
2345         if (r) {
2346                 ti->error = "Cache data device lookup failed";
2347                 goto bad;
2348         }
2349         wc->memory_map_size = bdev_nr_bytes(wc->ssd_dev->bdev);
2350
2351         /*
2352          * Parse the cache block size
2353          */
2354         string = dm_shift_arg(&as);
2355         if (!string)
2356                 goto bad_arguments;
2357         if (sscanf(string, "%u%c", &wc->block_size, &dummy) != 1 ||
2358             wc->block_size < 512 || wc->block_size > PAGE_SIZE ||
2359             (wc->block_size & (wc->block_size - 1))) {
2360                 r = -EINVAL;
2361                 ti->error = "Invalid block size";
2362                 goto bad;
2363         }
2364         if (wc->block_size < bdev_logical_block_size(wc->dev->bdev) ||
2365             wc->block_size < bdev_logical_block_size(wc->ssd_dev->bdev)) {
2366                 r = -EINVAL;
2367                 ti->error = "Block size is smaller than device logical block size";
2368                 goto bad;
2369         }
2370         wc->block_size_bits = __ffs(wc->block_size);
2371
2372         wc->max_writeback_jobs = MAX_WRITEBACK_JOBS;
2373         wc->autocommit_blocks = !WC_MODE_PMEM(wc) ? AUTOCOMMIT_BLOCKS_SSD : AUTOCOMMIT_BLOCKS_PMEM;
2374         wc->autocommit_jiffies = msecs_to_jiffies(AUTOCOMMIT_MSEC);
2375
2376         /*
2377          * Parse optional arguments
2378          */
2379         r = dm_read_arg_group(_args, &as, &opt_params, &ti->error);
2380         if (r)
2381                 goto bad;
2382
2383         while (opt_params) {
2384                 string = dm_shift_arg(&as), opt_params--;
2385                 if (!strcasecmp(string, "start_sector") && opt_params >= 1) {
2386                         unsigned long long start_sector;
2387                         string = dm_shift_arg(&as), opt_params--;
2388                         if (sscanf(string, "%llu%c", &start_sector, &dummy) != 1)
2389                                 goto invalid_optional;
2390                         wc->start_sector = start_sector;
2391                         wc->start_sector_set = true;
2392                         if (wc->start_sector != start_sector ||
2393                             wc->start_sector >= wc->memory_map_size >> SECTOR_SHIFT)
2394                                 goto invalid_optional;
2395                 } else if (!strcasecmp(string, "high_watermark") && opt_params >= 1) {
2396                         string = dm_shift_arg(&as), opt_params--;
2397                         if (sscanf(string, "%d%c", &high_wm_percent, &dummy) != 1)
2398                                 goto invalid_optional;
2399                         if (high_wm_percent < 0 || high_wm_percent > 100)
2400                                 goto invalid_optional;
2401                         wc->high_wm_percent_value = high_wm_percent;
2402                         wc->high_wm_percent_set = true;
2403                 } else if (!strcasecmp(string, "low_watermark") && opt_params >= 1) {
2404                         string = dm_shift_arg(&as), opt_params--;
2405                         if (sscanf(string, "%d%c", &low_wm_percent, &dummy) != 1)
2406                                 goto invalid_optional;
2407                         if (low_wm_percent < 0 || low_wm_percent > 100)
2408                                 goto invalid_optional;
2409                         wc->low_wm_percent_value = low_wm_percent;
2410                         wc->low_wm_percent_set = true;
2411                 } else if (!strcasecmp(string, "writeback_jobs") && opt_params >= 1) {
2412                         string = dm_shift_arg(&as), opt_params--;
2413                         if (sscanf(string, "%u%c", &wc->max_writeback_jobs, &dummy) != 1)
2414                                 goto invalid_optional;
2415                         wc->max_writeback_jobs_set = true;
2416                 } else if (!strcasecmp(string, "autocommit_blocks") && opt_params >= 1) {
2417                         string = dm_shift_arg(&as), opt_params--;
2418                         if (sscanf(string, "%u%c", &wc->autocommit_blocks, &dummy) != 1)
2419                                 goto invalid_optional;
2420                         wc->autocommit_blocks_set = true;
2421                 } else if (!strcasecmp(string, "autocommit_time") && opt_params >= 1) {
2422                         unsigned autocommit_msecs;
2423                         string = dm_shift_arg(&as), opt_params--;
2424                         if (sscanf(string, "%u%c", &autocommit_msecs, &dummy) != 1)
2425                                 goto invalid_optional;
2426                         if (autocommit_msecs > 3600000)
2427                                 goto invalid_optional;
2428                         wc->autocommit_jiffies = msecs_to_jiffies(autocommit_msecs);
2429                         wc->autocommit_time_value = autocommit_msecs;
2430                         wc->autocommit_time_set = true;
2431                 } else if (!strcasecmp(string, "max_age") && opt_params >= 1) {
2432                         unsigned max_age_msecs;
2433                         string = dm_shift_arg(&as), opt_params--;
2434                         if (sscanf(string, "%u%c", &max_age_msecs, &dummy) != 1)
2435                                 goto invalid_optional;
2436                         if (max_age_msecs > 86400000)
2437                                 goto invalid_optional;
2438                         wc->max_age = msecs_to_jiffies(max_age_msecs);
2439                         wc->max_age_set = true;
2440                         wc->max_age_value = max_age_msecs;
2441                 } else if (!strcasecmp(string, "cleaner")) {
2442                         wc->cleaner_set = true;
2443                         wc->cleaner = true;
2444                 } else if (!strcasecmp(string, "fua")) {
2445                         if (WC_MODE_PMEM(wc)) {
2446                                 wc->writeback_fua = true;
2447                                 wc->writeback_fua_set = true;
2448                         } else goto invalid_optional;
2449                 } else if (!strcasecmp(string, "nofua")) {
2450                         if (WC_MODE_PMEM(wc)) {
2451                                 wc->writeback_fua = false;
2452                                 wc->writeback_fua_set = true;
2453                         } else goto invalid_optional;
2454                 } else if (!strcasecmp(string, "metadata_only")) {
2455                         wc->metadata_only = true;
2456                 } else if (!strcasecmp(string, "pause_writeback") && opt_params >= 1) {
2457                         unsigned pause_msecs;
2458                         if (WC_MODE_PMEM(wc))
2459                                 goto invalid_optional;
2460                         string = dm_shift_arg(&as), opt_params--;
2461                         if (sscanf(string, "%u%c", &pause_msecs, &dummy) != 1)
2462                                 goto invalid_optional;
2463                         if (pause_msecs > 60000)
2464                                 goto invalid_optional;
2465                         wc->pause = msecs_to_jiffies(pause_msecs);
2466                         wc->pause_set = true;
2467                         wc->pause_value = pause_msecs;
2468                 } else {
2469 invalid_optional:
2470                         r = -EINVAL;
2471                         ti->error = "Invalid optional argument";
2472                         goto bad;
2473                 }
2474         }
2475
2476         if (high_wm_percent < low_wm_percent) {
2477                 r = -EINVAL;
2478                 ti->error = "High watermark must be greater than or equal to low watermark";
2479                 goto bad;
2480         }
2481
2482         if (WC_MODE_PMEM(wc)) {
2483                 if (!dax_synchronous(wc->ssd_dev->dax_dev)) {
2484                         r = -EOPNOTSUPP;
2485                         ti->error = "Asynchronous persistent memory not supported as pmem cache";
2486                         goto bad;
2487                 }
2488
2489                 r = persistent_memory_claim(wc);
2490                 if (r) {
2491                         ti->error = "Unable to map persistent memory for cache";
2492                         goto bad;
2493                 }
2494         } else {
2495                 size_t n_blocks, n_metadata_blocks;
2496                 uint64_t n_bitmap_bits;
2497
2498                 wc->memory_map_size -= (uint64_t)wc->start_sector << SECTOR_SHIFT;
2499
2500                 bio_list_init(&wc->flush_list);
2501                 wc->flush_thread = kthread_run(writecache_flush_thread, wc, "dm_writecache_flush");
2502                 if (IS_ERR(wc->flush_thread)) {
2503                         r = PTR_ERR(wc->flush_thread);
2504                         wc->flush_thread = NULL;
2505                         ti->error = "Couldn't spawn flush thread";
2506                         goto bad;
2507                 }
2508
2509                 r = calculate_memory_size(wc->memory_map_size, wc->block_size,
2510                                           &n_blocks, &n_metadata_blocks);
2511                 if (r) {
2512                         ti->error = "Invalid device size";
2513                         goto bad;
2514                 }
2515
2516                 n_bitmap_bits = (((uint64_t)n_metadata_blocks << wc->block_size_bits) +
2517                                  BITMAP_GRANULARITY - 1) / BITMAP_GRANULARITY;
2518                 /* this is limitation of test_bit functions */
2519                 if (n_bitmap_bits > 1U << 31) {
2520                         r = -EFBIG;
2521                         ti->error = "Invalid device size";
2522                         goto bad;
2523                 }
2524
2525                 wc->memory_map = vmalloc(n_metadata_blocks << wc->block_size_bits);
2526                 if (!wc->memory_map) {
2527                         r = -ENOMEM;
2528                         ti->error = "Unable to allocate memory for metadata";
2529                         goto bad;
2530                 }
2531
2532                 wc->dm_kcopyd = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2533                 if (IS_ERR(wc->dm_kcopyd)) {
2534                         r = PTR_ERR(wc->dm_kcopyd);
2535                         ti->error = "Unable to allocate dm-kcopyd client";
2536                         wc->dm_kcopyd = NULL;
2537                         goto bad;
2538                 }
2539
2540                 wc->metadata_sectors = n_metadata_blocks << (wc->block_size_bits - SECTOR_SHIFT);
2541                 wc->dirty_bitmap_size = (n_bitmap_bits + BITS_PER_LONG - 1) /
2542                         BITS_PER_LONG * sizeof(unsigned long);
2543                 wc->dirty_bitmap = vzalloc(wc->dirty_bitmap_size);
2544                 if (!wc->dirty_bitmap) {
2545                         r = -ENOMEM;
2546                         ti->error = "Unable to allocate dirty bitmap";
2547                         goto bad;
2548                 }
2549
2550                 r = writecache_read_metadata(wc, wc->block_size >> SECTOR_SHIFT);
2551                 if (r) {
2552                         ti->error = "Unable to read first block of metadata";
2553                         goto bad;
2554                 }
2555         }
2556
2557         r = copy_mc_to_kernel(&s, sb(wc), sizeof(struct wc_memory_superblock));
2558         if (r) {
2559                 ti->error = "Hardware memory error when reading superblock";
2560                 goto bad;
2561         }
2562         if (!le32_to_cpu(s.magic) && !le32_to_cpu(s.version)) {
2563                 r = init_memory(wc);
2564                 if (r) {
2565                         ti->error = "Unable to initialize device";
2566                         goto bad;
2567                 }
2568                 r = copy_mc_to_kernel(&s, sb(wc),
2569                                       sizeof(struct wc_memory_superblock));
2570                 if (r) {
2571                         ti->error = "Hardware memory error when reading superblock";
2572                         goto bad;
2573                 }
2574         }
2575
2576         if (le32_to_cpu(s.magic) != MEMORY_SUPERBLOCK_MAGIC) {
2577                 ti->error = "Invalid magic in the superblock";
2578                 r = -EINVAL;
2579                 goto bad;
2580         }
2581
2582         if (le32_to_cpu(s.version) != MEMORY_SUPERBLOCK_VERSION) {
2583                 ti->error = "Invalid version in the superblock";
2584                 r = -EINVAL;
2585                 goto bad;
2586         }
2587
2588         if (le32_to_cpu(s.block_size) != wc->block_size) {
2589                 ti->error = "Block size does not match superblock";
2590                 r = -EINVAL;
2591                 goto bad;
2592         }
2593
2594         wc->n_blocks = le64_to_cpu(s.n_blocks);
2595
2596         offset = wc->n_blocks * sizeof(struct wc_memory_entry);
2597         if (offset / sizeof(struct wc_memory_entry) != le64_to_cpu(sb(wc)->n_blocks)) {
2598 overflow:
2599                 ti->error = "Overflow in size calculation";
2600                 r = -EINVAL;
2601                 goto bad;
2602         }
2603         offset += sizeof(struct wc_memory_superblock);
2604         if (offset < sizeof(struct wc_memory_superblock))
2605                 goto overflow;
2606         offset = (offset + wc->block_size - 1) & ~(size_t)(wc->block_size - 1);
2607         data_size = wc->n_blocks * (size_t)wc->block_size;
2608         if (!offset || (data_size / wc->block_size != wc->n_blocks) ||
2609             (offset + data_size < offset))
2610                 goto overflow;
2611         if (offset + data_size > wc->memory_map_size) {
2612                 ti->error = "Memory area is too small";
2613                 r = -EINVAL;
2614                 goto bad;
2615         }
2616
2617         wc->metadata_sectors = offset >> SECTOR_SHIFT;
2618         wc->block_start = (char *)sb(wc) + offset;
2619
2620         x = (uint64_t)wc->n_blocks * (100 - high_wm_percent);
2621         x += 50;
2622         do_div(x, 100);
2623         wc->freelist_high_watermark = x;
2624         x = (uint64_t)wc->n_blocks * (100 - low_wm_percent);
2625         x += 50;
2626         do_div(x, 100);
2627         wc->freelist_low_watermark = x;
2628
2629         if (wc->cleaner)
2630                 activate_cleaner(wc);
2631
2632         r = writecache_alloc_entries(wc);
2633         if (r) {
2634                 ti->error = "Cannot allocate memory";
2635                 goto bad;
2636         }
2637
2638         ti->num_flush_bios = WC_MODE_PMEM(wc) ? 1 : 2;
2639         ti->flush_supported = true;
2640         ti->num_discard_bios = 1;
2641
2642         if (WC_MODE_PMEM(wc))
2643                 persistent_memory_flush_cache(wc->memory_map, wc->memory_map_size);
2644
2645         return 0;
2646
2647 bad_arguments:
2648         r = -EINVAL;
2649         ti->error = "Bad arguments";
2650 bad:
2651         writecache_dtr(ti);
2652         return r;
2653 }
2654
2655 static void writecache_status(struct dm_target *ti, status_type_t type,
2656                               unsigned status_flags, char *result, unsigned maxlen)
2657 {
2658         struct dm_writecache *wc = ti->private;
2659         unsigned extra_args;
2660         unsigned sz = 0;
2661
2662         switch (type) {
2663         case STATUSTYPE_INFO:
2664                 DMEMIT("%ld %llu %llu %llu %llu %llu %llu %llu %llu %llu %llu %llu %llu %llu",
2665                        writecache_has_error(wc),
2666                        (unsigned long long)wc->n_blocks, (unsigned long long)wc->freelist_size,
2667                        (unsigned long long)wc->writeback_size,
2668                        wc->stats.reads,
2669                        wc->stats.read_hits,
2670                        wc->stats.writes,
2671                        wc->stats.write_hits_uncommitted,
2672                        wc->stats.write_hits_committed,
2673                        wc->stats.writes_around,
2674                        wc->stats.writes_allocate,
2675                        wc->stats.writes_blocked_on_freelist,
2676                        wc->stats.flushes,
2677                        wc->stats.discards);
2678                 break;
2679         case STATUSTYPE_TABLE:
2680                 DMEMIT("%c %s %s %u ", WC_MODE_PMEM(wc) ? 'p' : 's',
2681                                 wc->dev->name, wc->ssd_dev->name, wc->block_size);
2682                 extra_args = 0;
2683                 if (wc->start_sector_set)
2684                         extra_args += 2;
2685                 if (wc->high_wm_percent_set)
2686                         extra_args += 2;
2687                 if (wc->low_wm_percent_set)
2688                         extra_args += 2;
2689                 if (wc->max_writeback_jobs_set)
2690                         extra_args += 2;
2691                 if (wc->autocommit_blocks_set)
2692                         extra_args += 2;
2693                 if (wc->autocommit_time_set)
2694                         extra_args += 2;
2695                 if (wc->max_age_set)
2696                         extra_args += 2;
2697                 if (wc->cleaner_set)
2698                         extra_args++;
2699                 if (wc->writeback_fua_set)
2700                         extra_args++;
2701                 if (wc->metadata_only)
2702                         extra_args++;
2703                 if (wc->pause_set)
2704                         extra_args += 2;
2705
2706                 DMEMIT("%u", extra_args);
2707                 if (wc->start_sector_set)
2708                         DMEMIT(" start_sector %llu", (unsigned long long)wc->start_sector);
2709                 if (wc->high_wm_percent_set)
2710                         DMEMIT(" high_watermark %u", wc->high_wm_percent_value);
2711                 if (wc->low_wm_percent_set)
2712                         DMEMIT(" low_watermark %u", wc->low_wm_percent_value);
2713                 if (wc->max_writeback_jobs_set)
2714                         DMEMIT(" writeback_jobs %u", wc->max_writeback_jobs);
2715                 if (wc->autocommit_blocks_set)
2716                         DMEMIT(" autocommit_blocks %u", wc->autocommit_blocks);
2717                 if (wc->autocommit_time_set)
2718                         DMEMIT(" autocommit_time %u", wc->autocommit_time_value);
2719                 if (wc->max_age_set)
2720                         DMEMIT(" max_age %u", wc->max_age_value);
2721                 if (wc->cleaner_set)
2722                         DMEMIT(" cleaner");
2723                 if (wc->writeback_fua_set)
2724                         DMEMIT(" %sfua", wc->writeback_fua ? "" : "no");
2725                 if (wc->metadata_only)
2726                         DMEMIT(" metadata_only");
2727                 if (wc->pause_set)
2728                         DMEMIT(" pause_writeback %u", wc->pause_value);
2729                 break;
2730         case STATUSTYPE_IMA:
2731                 *result = '\0';
2732                 break;
2733         }
2734 }
2735
2736 static struct target_type writecache_target = {
2737         .name                   = "writecache",
2738         .version                = {1, 6, 0},
2739         .module                 = THIS_MODULE,
2740         .ctr                    = writecache_ctr,
2741         .dtr                    = writecache_dtr,
2742         .status                 = writecache_status,
2743         .postsuspend            = writecache_suspend,
2744         .resume                 = writecache_resume,
2745         .message                = writecache_message,
2746         .map                    = writecache_map,
2747         .end_io                 = writecache_end_io,
2748         .iterate_devices        = writecache_iterate_devices,
2749         .io_hints               = writecache_io_hints,
2750 };
2751
2752 static int __init dm_writecache_init(void)
2753 {
2754         int r;
2755
2756         r = dm_register_target(&writecache_target);
2757         if (r < 0) {
2758                 DMERR("register failed %d", r);
2759                 return r;
2760         }
2761
2762         return 0;
2763 }
2764
2765 static void __exit dm_writecache_exit(void)
2766 {
2767         dm_unregister_target(&writecache_target);
2768 }
2769
2770 module_init(dm_writecache_init);
2771 module_exit(dm_writecache_exit);
2772
2773 MODULE_DESCRIPTION(DM_NAME " writecache target");
2774 MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
2775 MODULE_LICENSE("GPL");