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