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