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         if (!WC_MODE_PMEM(wc))
 730                 writecache_wait_for_ios(wc, WRITE);
 731
 732         wc->seq_count++;
 733         pmem_assign(sb(wc)->seq_count, cpu_to_le64(wc->seq_count));
 734         writecache_flush_region(wc, &sb(wc)->seq_count, sizeof sb(wc)->seq_count);
 735         writecache_commit_flushed(wc);
 736
 737         wc->overwrote_committed = false;
 738
 739         need_flush_after_free = false;
 740         while (1) {
 741                 /* Free another committed entry with lower seq-count */
 742                 struct rb_node *rb_node = rb_prev(&e->rb_node);
 743
 744                 if (rb_node) {
 745                         e2 = container_of(rb_node, struct wc_entry, rb_node);
 746                         if (read_original_sector(wc, e2) == read_original_sector(wc, e) &&
 747                             likely(!e2->write_in_progress)) {
 748                                 writecache_free_entry(wc, e2);
 749                                 need_flush_after_free = true;
 750                         }
 751                 }
 752                 if (unlikely(e->lru.prev == &wc->lru))
 753                         break;
 754                 e = container_of(e->lru.prev, struct wc_entry, lru);
 755                 cond_resched();
 756         }
 757
 758         if (need_flush_after_free)
 759                 writecache_commit_flushed(wc);
 760 }
 761
 762 static void writecache_flush_work(struct work_struct *work)
 763 {
 764         struct dm_writecache *wc = container_of(work, struct dm_writecache, flush_work);
 765
 766         wc_lock(wc);
 767         writecache_flush(wc);
 768         wc_unlock(wc);
 769 }
 770
 771 static void writecache_autocommit_timer(struct timer_list *t)
 772 {
 773         struct dm_writecache *wc = from_timer(wc, t, autocommit_timer);
 774         if (!writecache_has_error(wc))
 775                 queue_work(wc->writeback_wq, &wc->flush_work);
 776 }
 777
 778 static void writecache_schedule_autocommit(struct dm_writecache *wc)
 779 {
 780         if (!timer_pending(&wc->autocommit_timer))
 781                 mod_timer(&wc->autocommit_timer, jiffies + wc->autocommit_jiffies);
 782 }
 783
 784 static void writecache_discard(struct dm_writecache *wc, sector_t start, sector_t end)
 785 {
 786         struct wc_entry *e;
 787         bool discarded_something = false;
 788
 789         e = writecache_find_entry(wc, start, WFE_RETURN_FOLLOWING | WFE_LOWEST_SEQ);
 790         if (unlikely(!e))
 791                 return;
 792
 793         while (read_original_sector(wc, e) < end) {
 794                 struct rb_node *node = rb_next(&e->rb_node);
 795
 796                 if (likely(!e->write_in_progress)) {
 797                         if (!discarded_something) {
 798                                 writecache_wait_for_ios(wc, READ);
 799                                 writecache_wait_for_ios(wc, WRITE);
 800                                 discarded_something = true;
 801                         }
 802                         writecache_free_entry(wc, e);
 803                 }
 804
 805                 if (unlikely(!node))
 806                         break;
 807
 808                 e = container_of(node, struct wc_entry, rb_node);
 809         }
 810
 811         if (discarded_something)
 812                 writecache_commit_flushed(wc);
 813 }
 814
 815 static bool writecache_wait_for_writeback(struct dm_writecache *wc)
 816 {
 817         if (wc->writeback_size) {
 818                 writecache_wait_on_freelist(wc);
 819                 return true;
 820         }
 821         return false;
 822 }
 823
 824 static void writecache_suspend(struct dm_target *ti)
 825 {
 826         struct dm_writecache *wc = ti->private;
 827         bool flush_on_suspend;
 828
 829         del_timer_sync(&wc->autocommit_timer);
 830
 831         wc_lock(wc);
 832         writecache_flush(wc);
 833         flush_on_suspend = wc->flush_on_suspend;
 834         if (flush_on_suspend) {
 835                 wc->flush_on_suspend = false;
 836                 wc->writeback_all++;
 837                 queue_work(wc->writeback_wq, &wc->writeback_work);
 838         }
 839         wc_unlock(wc);
 840
 841         flush_workqueue(wc->writeback_wq);
 842
 843         wc_lock(wc);
 844         if (flush_on_suspend)
 845                 wc->writeback_all--;
 846         while (writecache_wait_for_writeback(wc));
 847
 848         if (WC_MODE_PMEM(wc))
 849                 persistent_memory_flush_cache(wc->memory_map, wc->memory_map_size);
 850
 851         writecache_poison_lists(wc);
 852
 853         wc_unlock(wc);
 854 }
 855
 856 static int writecache_alloc_entries(struct dm_writecache *wc)
 857 {
 858         size_t b;
 859
 860         if (wc->entries)
 861                 return 0;
 862         wc->entries = vmalloc(array_size(sizeof(struct wc_entry), wc->n_blocks));
 863         if (!wc->entries)
 864                 return -ENOMEM;
 865         for (b = 0; b < wc->n_blocks; b++) {
 866                 struct wc_entry *e = &wc->entries[b];
 867                 e->index = b;
 868                 e->write_in_progress = false;
 869         }
 870
 871         return 0;
 872 }
 873
 874 static void writecache_resume(struct dm_target *ti)
 875 {
 876         struct dm_writecache *wc = ti->private;
 877         size_t b;
 878         bool need_flush = false;
 879         __le64 sb_seq_count;
 880         int r;
 881
 882         wc_lock(wc);
 883
 884         if (WC_MODE_PMEM(wc))
 885                 persistent_memory_invalidate_cache(wc->memory_map, wc->memory_map_size);
 886
 887         wc->tree = RB_ROOT;
 888         INIT_LIST_HEAD(&wc->lru);
 889         if (WC_MODE_SORT_FREELIST(wc)) {
 890                 wc->freetree = RB_ROOT;
 891                 wc->current_free = NULL;
 892         } else {
 893                 INIT_LIST_HEAD(&wc->freelist);
 894         }
 895         wc->freelist_size = 0;
 896
 897         r = memcpy_mcsafe(&sb_seq_count, &sb(wc)->seq_count, sizeof(uint64_t));
 898         if (r) {
 899                 writecache_error(wc, r, "hardware memory error when reading superblock: %d", r);
 900                 sb_seq_count = cpu_to_le64(0);
 901         }
 902         wc->seq_count = le64_to_cpu(sb_seq_count);
 903
 904 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 905         for (b = 0; b < wc->n_blocks; b++) {
 906                 struct wc_entry *e = &wc->entries[b];
 907                 struct wc_memory_entry wme;
 908                 if (writecache_has_error(wc)) {
 909                         e->original_sector = -1;
 910                         e->seq_count = -1;
 911                         continue;
 912                 }
 913                 r = memcpy_mcsafe(&wme, memory_entry(wc, e), sizeof(struct wc_memory_entry));
 914                 if (r) {
 915                         writecache_error(wc, r, "hardware memory error when reading metadata entry %lu: %d",
 916                                          (unsigned long)b, r);
 917                         e->original_sector = -1;
 918                         e->seq_count = -1;
 919                 } else {
 920                         e->original_sector = le64_to_cpu(wme.original_sector);
 921                         e->seq_count = le64_to_cpu(wme.seq_count);
 922                 }
 923         }
 924 #endif
 925         for (b = 0; b < wc->n_blocks; b++) {
 926                 struct wc_entry *e = &wc->entries[b];
 927                 if (!writecache_entry_is_committed(wc, e)) {
 928                         if (read_seq_count(wc, e) != -1) {
 929 erase_this:
 930                                 clear_seq_count(wc, e);
 931                                 need_flush = true;
 932                         }
 933                         writecache_add_to_freelist(wc, e);
 934                 } else {
 935                         struct wc_entry *old;
 936
 937                         old = writecache_find_entry(wc, read_original_sector(wc, e), 0);
 938                         if (!old) {
 939                                 writecache_insert_entry(wc, e);
 940                         } else {
 941                                 if (read_seq_count(wc, old) == read_seq_count(wc, e)) {
 942                                         writecache_error(wc, -EINVAL,
 943                                                  "two identical entries, position %llu, sector %llu, sequence %llu",
 944                                                  (unsigned long long)b, (unsigned long long)read_original_sector(wc, e),
 945                                                  (unsigned long long)read_seq_count(wc, e));
 946                                 }
 947                                 if (read_seq_count(wc, old) > read_seq_count(wc, e)) {
 948                                         goto erase_this;
 949                                 } else {
 950                                         writecache_free_entry(wc, old);
 951                                         writecache_insert_entry(wc, e);
 952                                         need_flush = true;
 953                                 }
 954                         }
 955                 }
 956                 cond_resched();
 957         }
 958
 959         if (need_flush) {
 960                 writecache_flush_all_metadata(wc);
 961                 writecache_commit_flushed(wc);
 962         }
 963
 964         wc_unlock(wc);
 965 }
 966
 967 static int process_flush_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
 968 {
 969         if (argc != 1)
 970                 return -EINVAL;
 971
 972         wc_lock(wc);
 973         if (dm_suspended(wc->ti)) {
 974                 wc_unlock(wc);
 975                 return -EBUSY;
 976         }
 977         if (writecache_has_error(wc)) {
 978                 wc_unlock(wc);
 979                 return -EIO;
 980         }
 981
 982         writecache_flush(wc);
 983         wc->writeback_all++;
 984         queue_work(wc->writeback_wq, &wc->writeback_work);
 985         wc_unlock(wc);
 986
 987         flush_workqueue(wc->writeback_wq);
 988
 989         wc_lock(wc);
 990         wc->writeback_all--;
 991         if (writecache_has_error(wc)) {
 992                 wc_unlock(wc);
 993                 return -EIO;
 994         }
 995         wc_unlock(wc);
 996
 997         return 0;
 998 }
 999
1000 static int process_flush_on_suspend_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1001 {
1002         if (argc != 1)
1003                 return -EINVAL;
1004
1005         wc_lock(wc);
1006         wc->flush_on_suspend = true;
1007         wc_unlock(wc);
1008
1009         return 0;
1010 }
1011
1012 static int writecache_message(struct dm_target *ti, unsigned argc, char **argv,
1013                               char *result, unsigned maxlen)
1014 {
1015         int r = -EINVAL;
1016         struct dm_writecache *wc = ti->private;
1017
1018         if (!strcasecmp(argv[0], "flush"))
1019                 r = process_flush_mesg(argc, argv, wc);
1020         else if (!strcasecmp(argv[0], "flush_on_suspend"))
1021                 r = process_flush_on_suspend_mesg(argc, argv, wc);
1022         else
1023                 DMERR("unrecognised message received: %s", argv[0]);
1024
1025         return r;
1026 }
1027
1028 static void bio_copy_block(struct dm_writecache *wc, struct bio *bio, void *data)
1029 {
1030         void *buf;
1031         unsigned long flags;
1032         unsigned size;
1033         int rw = bio_data_dir(bio);
1034         unsigned remaining_size = wc->block_size;
1035
1036         do {
1037                 struct bio_vec bv = bio_iter_iovec(bio, bio->bi_iter);
1038                 buf = bvec_kmap_irq(&bv, &flags);
1039                 size = bv.bv_len;
1040                 if (unlikely(size > remaining_size))
1041                         size = remaining_size;
1042
1043                 if (rw == READ) {
1044                         int r;
1045                         r = memcpy_mcsafe(buf, data, size);
1046                         flush_dcache_page(bio_page(bio));
1047                         if (unlikely(r)) {
1048                                 writecache_error(wc, r, "hardware memory error when reading data: %d", r);
1049                                 bio->bi_status = BLK_STS_IOERR;
1050                         }
1051                 } else {
1052                         flush_dcache_page(bio_page(bio));
1053                         memcpy_flushcache(data, buf, size);
1054                 }
1055
1056                 bvec_kunmap_irq(buf, &flags);
1057
1058                 data = (char *)data + size;
1059                 remaining_size -= size;
1060                 bio_advance(bio, size);
1061         } while (unlikely(remaining_size));
1062 }
1063
1064 static int writecache_flush_thread(void *data)
1065 {
1066         struct dm_writecache *wc = data;
1067
1068         while (1) {
1069                 struct bio *bio;
1070
1071                 wc_lock(wc);
1072                 bio = bio_list_pop(&wc->flush_list);
1073                 if (!bio) {
1074                         set_current_state(TASK_INTERRUPTIBLE);
1075                         wc_unlock(wc);
1076
1077                         if (unlikely(kthread_should_stop())) {
1078                                 set_current_state(TASK_RUNNING);
1079                                 break;
1080                         }
1081
1082                         schedule();
1083                         continue;
1084                 }
1085
1086                 if (bio_op(bio) == REQ_OP_DISCARD) {
1087                         writecache_discard(wc, bio->bi_iter.bi_sector,
1088                                            bio_end_sector(bio));
1089                         wc_unlock(wc);
1090                         bio_set_dev(bio, wc->dev->bdev);
1091                         generic_make_request(bio);
1092                 } else {
1093                         writecache_flush(wc);
1094                         wc_unlock(wc);
1095                         if (writecache_has_error(wc))
1096                                 bio->bi_status = BLK_STS_IOERR;
1097                         bio_endio(bio);
1098                 }
1099         }
1100
1101         return 0;
1102 }
1103
1104 static void writecache_offload_bio(struct dm_writecache *wc, struct bio *bio)
1105 {
1106         if (bio_list_empty(&wc->flush_list))
1107                 wake_up_process(wc->flush_thread);
1108         bio_list_add(&wc->flush_list, bio);
1109 }
1110
1111 static int writecache_map(struct dm_target *ti, struct bio *bio)
1112 {
1113         struct wc_entry *e;
1114         struct dm_writecache *wc = ti->private;
1115
1116         bio->bi_private = NULL;
1117
1118         wc_lock(wc);
1119
1120         if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1121                 if (writecache_has_error(wc))
1122                         goto unlock_error;
1123                 if (WC_MODE_PMEM(wc)) {
1124                         writecache_flush(wc);
1125                         if (writecache_has_error(wc))
1126                                 goto unlock_error;
1127                         goto unlock_submit;
1128                 } else {
1129                         writecache_offload_bio(wc, bio);
1130                         goto unlock_return;
1131                 }
1132         }
1133
1134         bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
1135
1136         if (unlikely((((unsigned)bio->bi_iter.bi_sector | bio_sectors(bio)) &
1137                                 (wc->block_size / 512 - 1)) != 0)) {
1138                 DMERR("I/O is not aligned, sector %llu, size %u, block size %u",
1139                       (unsigned long long)bio->bi_iter.bi_sector,
1140                       bio->bi_iter.bi_size, wc->block_size);
1141                 goto unlock_error;
1142         }
1143
1144         if (unlikely(bio_op(bio) == REQ_OP_DISCARD)) {
1145                 if (writecache_has_error(wc))
1146                         goto unlock_error;
1147                 if (WC_MODE_PMEM(wc)) {
1148                         writecache_discard(wc, bio->bi_iter.bi_sector, bio_end_sector(bio));
1149                         goto unlock_remap_origin;
1150                 } else {
1151                         writecache_offload_bio(wc, bio);
1152                         goto unlock_return;
1153                 }
1154         }
1155
1156         if (bio_data_dir(bio) == READ) {
1157 read_next_block:
1158                 e = writecache_find_entry(wc, bio->bi_iter.bi_sector, WFE_RETURN_FOLLOWING);
1159                 if (e && read_original_sector(wc, e) == bio->bi_iter.bi_sector) {
1160                         if (WC_MODE_PMEM(wc)) {
1161                                 bio_copy_block(wc, bio, memory_data(wc, e));
1162                                 if (bio->bi_iter.bi_size)
1163                                         goto read_next_block;
1164                                 goto unlock_submit;
1165                         } else {
1166                                 dm_accept_partial_bio(bio, wc->block_size >> SECTOR_SHIFT);
1167                                 bio_set_dev(bio, wc->ssd_dev->bdev);
1168                                 bio->bi_iter.bi_sector = cache_sector(wc, e);
1169                                 if (!writecache_entry_is_committed(wc, e))
1170                                         writecache_wait_for_ios(wc, WRITE);
1171                                 goto unlock_remap;
1172                         }
1173                 } else {
1174                         if (e) {
1175                                 sector_t next_boundary =
1176                                         read_original_sector(wc, e) - bio->bi_iter.bi_sector;
1177                                 if (next_boundary < bio->bi_iter.bi_size >> SECTOR_SHIFT) {
1178                                         dm_accept_partial_bio(bio, next_boundary);
1179                                 }
1180                         }
1181                         goto unlock_remap_origin;
1182                 }
1183         } else {
1184                 do {
1185                         if (writecache_has_error(wc))
1186                                 goto unlock_error;
1187                         e = writecache_find_entry(wc, bio->bi_iter.bi_sector, 0);
1188                         if (e) {
1189                                 if (!writecache_entry_is_committed(wc, e))
1190                                         goto bio_copy;
1191                                 if (!WC_MODE_PMEM(wc) && !e->write_in_progress) {
1192                                         wc->overwrote_committed = true;
1193                                         goto bio_copy;
1194                                 }
1195                         }
1196                         e = writecache_pop_from_freelist(wc);
1197                         if (unlikely(!e)) {
1198                                 writecache_wait_on_freelist(wc);
1199                                 continue;
1200                         }
1201                         write_original_sector_seq_count(wc, e, bio->bi_iter.bi_sector, wc->seq_count);
1202                         writecache_insert_entry(wc, e);
1203                         wc->uncommitted_blocks++;
1204 bio_copy:
1205                         if (WC_MODE_PMEM(wc)) {
1206                                 bio_copy_block(wc, bio, memory_data(wc, e));
1207                         } else {
1208                                 dm_accept_partial_bio(bio, wc->block_size >> SECTOR_SHIFT);
1209                                 bio_set_dev(bio, wc->ssd_dev->bdev);
1210                                 bio->bi_iter.bi_sector = cache_sector(wc, e);
1211                                 if (unlikely(wc->uncommitted_blocks >= wc->autocommit_blocks)) {
1212                                         wc->uncommitted_blocks = 0;
1213                                         queue_work(wc->writeback_wq, &wc->flush_work);
1214                                 } else {
1215                                         writecache_schedule_autocommit(wc);
1216                                 }
1217                                 goto unlock_remap;
1218                         }
1219                 } while (bio->bi_iter.bi_size);
1220
1221                 if (unlikely(bio->bi_opf & REQ_FUA ||
1222                              wc->uncommitted_blocks >= wc->autocommit_blocks))
1223                         writecache_flush(wc);
1224                 else
1225                         writecache_schedule_autocommit(wc);
1226                 goto unlock_submit;
1227         }
1228
1229 unlock_remap_origin:
1230         bio_set_dev(bio, wc->dev->bdev);
1231         wc_unlock(wc);
1232         return DM_MAPIO_REMAPPED;
1233
1234 unlock_remap:
1235         /* make sure that writecache_end_io decrements bio_in_progress: */
1236         bio->bi_private = (void *)1;
1237         atomic_inc(&wc->bio_in_progress[bio_data_dir(bio)]);
1238         wc_unlock(wc);
1239         return DM_MAPIO_REMAPPED;
1240
1241 unlock_submit:
1242         wc_unlock(wc);
1243         bio_endio(bio);
1244         return DM_MAPIO_SUBMITTED;
1245
1246 unlock_return:
1247         wc_unlock(wc);
1248         return DM_MAPIO_SUBMITTED;
1249
1250 unlock_error:
1251         wc_unlock(wc);
1252         bio_io_error(bio);
1253         return DM_MAPIO_SUBMITTED;
1254 }
1255
1256 static int writecache_end_io(struct dm_target *ti, struct bio *bio, blk_status_t *status)
1257 {
1258         struct dm_writecache *wc = ti->private;
1259
1260         if (bio->bi_private != NULL) {
1261                 int dir = bio_data_dir(bio);
1262                 if (atomic_dec_and_test(&wc->bio_in_progress[dir]))
1263                         if (unlikely(waitqueue_active(&wc->bio_in_progress_wait[dir])))
1264                                 wake_up(&wc->bio_in_progress_wait[dir]);
1265         }
1266         return 0;
1267 }
1268
1269 static int writecache_iterate_devices(struct dm_target *ti,
1270                                       iterate_devices_callout_fn fn, void *data)
1271 {
1272         struct dm_writecache *wc = ti->private;
1273
1274         return fn(ti, wc->dev, 0, ti->len, data);
1275 }
1276
1277 static void writecache_io_hints(struct dm_target *ti, struct queue_limits *limits)
1278 {
1279         struct dm_writecache *wc = ti->private;
1280
1281         if (limits->logical_block_size < wc->block_size)
1282                 limits->logical_block_size = wc->block_size;
1283
1284         if (limits->physical_block_size < wc->block_size)
1285                 limits->physical_block_size = wc->block_size;
1286
1287         if (limits->io_min < wc->block_size)
1288                 limits->io_min = wc->block_size;
1289 }
1290
1291
1292 static void writecache_writeback_endio(struct bio *bio)
1293 {
1294         struct writeback_struct *wb = container_of(bio, struct writeback_struct, bio);
1295         struct dm_writecache *wc = wb->wc;
1296         unsigned long flags;
1297
1298         raw_spin_lock_irqsave(&wc->endio_list_lock, flags);
1299         if (unlikely(list_empty(&wc->endio_list)))
1300                 wake_up_process(wc->endio_thread);
1301         list_add_tail(&wb->endio_entry, &wc->endio_list);
1302         raw_spin_unlock_irqrestore(&wc->endio_list_lock, flags);
1303 }
1304
1305 static void writecache_copy_endio(int read_err, unsigned long write_err, void *ptr)
1306 {
1307         struct copy_struct *c = ptr;
1308         struct dm_writecache *wc = c->wc;
1309
1310         c->error = likely(!(read_err | write_err)) ? 0 : -EIO;
1311
1312         raw_spin_lock_irq(&wc->endio_list_lock);
1313         if (unlikely(list_empty(&wc->endio_list)))
1314                 wake_up_process(wc->endio_thread);
1315         list_add_tail(&c->endio_entry, &wc->endio_list);
1316         raw_spin_unlock_irq(&wc->endio_list_lock);
1317 }
1318
1319 static void __writecache_endio_pmem(struct dm_writecache *wc, struct list_head *list)
1320 {
1321         unsigned i;
1322         struct writeback_struct *wb;
1323         struct wc_entry *e;
1324         unsigned long n_walked = 0;
1325
1326         do {
1327                 wb = list_entry(list->next, struct writeback_struct, endio_entry);
1328                 list_del(&wb->endio_entry);
1329
1330                 if (unlikely(wb->bio.bi_status != BLK_STS_OK))
1331                         writecache_error(wc, blk_status_to_errno(wb->bio.bi_status),
1332                                         "write error %d", wb->bio.bi_status);
1333                 i = 0;
1334                 do {
1335                         e = wb->wc_list[i];
1336                         BUG_ON(!e->write_in_progress);
1337                         e->write_in_progress = false;
1338                         INIT_LIST_HEAD(&e->lru);
1339                         if (!writecache_has_error(wc))
1340                                 writecache_free_entry(wc, e);
1341                         BUG_ON(!wc->writeback_size);
1342                         wc->writeback_size--;
1343                         n_walked++;
1344                         if (unlikely(n_walked >= ENDIO_LATENCY)) {
1345                                 writecache_commit_flushed(wc);
1346                                 wc_unlock(wc);
1347                                 wc_lock(wc);
1348                                 n_walked = 0;
1349                         }
1350                 } while (++i < wb->wc_list_n);
1351
1352                 if (wb->wc_list != wb->wc_list_inline)
1353                         kfree(wb->wc_list);
1354                 bio_put(&wb->bio);
1355         } while (!list_empty(list));
1356 }
1357
1358 static void __writecache_endio_ssd(struct dm_writecache *wc, struct list_head *list)
1359 {
1360         struct copy_struct *c;
1361         struct wc_entry *e;
1362
1363         do {
1364                 c = list_entry(list->next, struct copy_struct, endio_entry);
1365                 list_del(&c->endio_entry);
1366
1367                 if (unlikely(c->error))
1368                         writecache_error(wc, c->error, "copy error");
1369
1370                 e = c->e;
1371                 do {
1372                         BUG_ON(!e->write_in_progress);
1373                         e->write_in_progress = false;
1374                         INIT_LIST_HEAD(&e->lru);
1375                         if (!writecache_has_error(wc))
1376                                 writecache_free_entry(wc, e);
1377
1378                         BUG_ON(!wc->writeback_size);
1379                         wc->writeback_size--;
1380                         e++;
1381                 } while (--c->n_entries);
1382                 mempool_free(c, &wc->copy_pool);
1383         } while (!list_empty(list));
1384 }
1385
1386 static int writecache_endio_thread(void *data)
1387 {
1388         struct dm_writecache *wc = data;
1389
1390         while (1) {
1391                 struct list_head list;
1392
1393                 raw_spin_lock_irq(&wc->endio_list_lock);
1394                 if (!list_empty(&wc->endio_list))
1395                         goto pop_from_list;
1396                 set_current_state(TASK_INTERRUPTIBLE);
1397                 raw_spin_unlock_irq(&wc->endio_list_lock);
1398
1399                 if (unlikely(kthread_should_stop())) {
1400                         set_current_state(TASK_RUNNING);
1401                         break;
1402                 }
1403
1404                 schedule();
1405
1406                 continue;
1407
1408 pop_from_list:
1409                 list = wc->endio_list;
1410                 list.next->prev = list.prev->next = &list;
1411                 INIT_LIST_HEAD(&wc->endio_list);
1412                 raw_spin_unlock_irq(&wc->endio_list_lock);
1413
1414                 if (!WC_MODE_FUA(wc))
1415                         writecache_disk_flush(wc, wc->dev);
1416
1417                 wc_lock(wc);
1418
1419                 if (WC_MODE_PMEM(wc)) {
1420                         __writecache_endio_pmem(wc, &list);
1421                 } else {
1422                         __writecache_endio_ssd(wc, &list);
1423                         writecache_wait_for_ios(wc, READ);
1424                 }
1425
1426                 writecache_commit_flushed(wc);
1427
1428                 wc_unlock(wc);
1429         }
1430
1431         return 0;
1432 }
1433
1434 static bool wc_add_block(struct writeback_struct *wb, struct wc_entry *e, gfp_t gfp)
1435 {
1436         struct dm_writecache *wc = wb->wc;
1437         unsigned block_size = wc->block_size;
1438         void *address = memory_data(wc, e);
1439
1440         persistent_memory_flush_cache(address, block_size);
1441         return bio_add_page(&wb->bio, persistent_memory_page(address),
1442                             block_size, persistent_memory_page_offset(address)) != 0;
1443 }
1444
1445 struct writeback_list {
1446         struct list_head list;
1447         size_t size;
1448 };
1449
1450 static void __writeback_throttle(struct dm_writecache *wc, struct writeback_list *wbl)
1451 {
1452         if (unlikely(wc->max_writeback_jobs)) {
1453                 if (READ_ONCE(wc->writeback_size) - wbl->size >= wc->max_writeback_jobs) {
1454                         wc_lock(wc);
1455                         while (wc->writeback_size - wbl->size >= wc->max_writeback_jobs)
1456                                 writecache_wait_on_freelist(wc);
1457                         wc_unlock(wc);
1458                 }
1459         }
1460         cond_resched();
1461 }
1462
1463 static void __writecache_writeback_pmem(struct dm_writecache *wc, struct writeback_list *wbl)
1464 {
1465         struct wc_entry *e, *f;
1466         struct bio *bio;
1467         struct writeback_struct *wb;
1468         unsigned max_pages;
1469
1470         while (wbl->size) {
1471                 wbl->size--;
1472                 e = container_of(wbl->list.prev, struct wc_entry, lru);
1473                 list_del(&e->lru);
1474
1475                 max_pages = e->wc_list_contiguous;
1476
1477                 bio = bio_alloc_bioset(GFP_NOIO, max_pages, &wc->bio_set);
1478                 wb = container_of(bio, struct writeback_struct, bio);
1479                 wb->wc = wc;
1480                 bio->bi_end_io = writecache_writeback_endio;
1481                 bio_set_dev(bio, wc->dev->bdev);
1482                 bio->bi_iter.bi_sector = read_original_sector(wc, e);
1483                 if (max_pages <= WB_LIST_INLINE ||
1484                     unlikely(!(wb->wc_list = kmalloc_array(max_pages, sizeof(struct wc_entry *),
1485                                                            GFP_NOIO | __GFP_NORETRY |
1486                                                            __GFP_NOMEMALLOC | __GFP_NOWARN)))) {
1487                         wb->wc_list = wb->wc_list_inline;
1488                         max_pages = WB_LIST_INLINE;
1489                 }
1490
1491                 BUG_ON(!wc_add_block(wb, e, GFP_NOIO));
1492
1493                 wb->wc_list[0] = e;
1494                 wb->wc_list_n = 1;
1495
1496                 while (wbl->size && wb->wc_list_n < max_pages) {
1497                         f = container_of(wbl->list.prev, struct wc_entry, lru);
1498                         if (read_original_sector(wc, f) !=
1499                             read_original_sector(wc, e) + (wc->block_size >> SECTOR_SHIFT))
1500                                 break;
1501                         if (!wc_add_block(wb, f, GFP_NOWAIT | __GFP_NOWARN))
1502                                 break;
1503                         wbl->size--;
1504                         list_del(&f->lru);
1505                         wb->wc_list[wb->wc_list_n++] = f;
1506                         e = f;
1507                 }
1508                 bio_set_op_attrs(bio, REQ_OP_WRITE, WC_MODE_FUA(wc) * REQ_FUA);
1509                 if (writecache_has_error(wc)) {
1510                         bio->bi_status = BLK_STS_IOERR;
1511                         bio_endio(bio);
1512                 } else {
1513                         submit_bio(bio);
1514                 }
1515
1516                 __writeback_throttle(wc, wbl);
1517         }
1518 }
1519
1520 static void __writecache_writeback_ssd(struct dm_writecache *wc, struct writeback_list *wbl)
1521 {
1522         struct wc_entry *e, *f;
1523         struct dm_io_region from, to;
1524         struct copy_struct *c;
1525
1526         while (wbl->size) {
1527                 unsigned n_sectors;
1528
1529                 wbl->size--;
1530                 e = container_of(wbl->list.prev, struct wc_entry, lru);
1531                 list_del(&e->lru);
1532
1533                 n_sectors = e->wc_list_contiguous << (wc->block_size_bits - SECTOR_SHIFT);
1534
1535                 from.bdev = wc->ssd_dev->bdev;
1536                 from.sector = cache_sector(wc, e);
1537                 from.count = n_sectors;
1538                 to.bdev = wc->dev->bdev;
1539                 to.sector = read_original_sector(wc, e);
1540                 to.count = n_sectors;
1541
1542                 c = mempool_alloc(&wc->copy_pool, GFP_NOIO);
1543                 c->wc = wc;
1544                 c->e = e;
1545                 c->n_entries = e->wc_list_contiguous;
1546
1547                 while ((n_sectors -= wc->block_size >> SECTOR_SHIFT)) {
1548                         wbl->size--;
1549                         f = container_of(wbl->list.prev, struct wc_entry, lru);
1550                         BUG_ON(f != e + 1);
1551                         list_del(&f->lru);
1552                         e = f;
1553                 }
1554
1555                 dm_kcopyd_copy(wc->dm_kcopyd, &from, 1, &to, 0, writecache_copy_endio, c);
1556
1557                 __writeback_throttle(wc, wbl);
1558         }
1559 }
1560
1561 static void writecache_writeback(struct work_struct *work)
1562 {
1563         struct dm_writecache *wc = container_of(work, struct dm_writecache, writeback_work);
1564         struct blk_plug plug;
1565         struct wc_entry *f, *uninitialized_var(g), *e = NULL;
1566         struct rb_node *node, *next_node;
1567         struct list_head skipped;
1568         struct writeback_list wbl;
1569         unsigned long n_walked;
1570
1571         wc_lock(wc);
1572 restart:
1573         if (writecache_has_error(wc)) {
1574                 wc_unlock(wc);
1575                 return;
1576         }
1577
1578         if (unlikely(wc->writeback_all)) {
1579                 if (writecache_wait_for_writeback(wc))
1580                         goto restart;
1581         }
1582
1583         if (wc->overwrote_committed) {
1584                 writecache_wait_for_ios(wc, WRITE);
1585         }
1586
1587         n_walked = 0;
1588         INIT_LIST_HEAD(&skipped);
1589         INIT_LIST_HEAD(&wbl.list);
1590         wbl.size = 0;
1591         while (!list_empty(&wc->lru) &&
1592                (wc->writeback_all ||
1593                 wc->freelist_size + wc->writeback_size <= wc->freelist_low_watermark)) {
1594
1595                 n_walked++;
1596                 if (unlikely(n_walked > WRITEBACK_LATENCY) &&
1597                     likely(!wc->writeback_all) && likely(!dm_suspended(wc->ti))) {
1598                         queue_work(wc->writeback_wq, &wc->writeback_work);
1599                         break;
1600                 }
1601
1602                 if (unlikely(wc->writeback_all)) {
1603                         if (unlikely(!e)) {
1604                                 writecache_flush(wc);
1605                                 e = container_of(rb_first(&wc->tree), struct wc_entry, rb_node);
1606                         } else
1607                                 e = g;
1608                 } else
1609                         e = container_of(wc->lru.prev, struct wc_entry, lru);
1610                 BUG_ON(e->write_in_progress);
1611                 if (unlikely(!writecache_entry_is_committed(wc, e))) {
1612                         writecache_flush(wc);
1613                 }
1614                 node = rb_prev(&e->rb_node);
1615                 if (node) {
1616                         f = container_of(node, struct wc_entry, rb_node);
1617                         if (unlikely(read_original_sector(wc, f) ==
1618                                      read_original_sector(wc, e))) {
1619                                 BUG_ON(!f->write_in_progress);
1620                                 list_del(&e->lru);
1621                                 list_add(&e->lru, &skipped);
1622                                 cond_resched();
1623                                 continue;
1624                         }
1625                 }
1626                 wc->writeback_size++;
1627                 list_del(&e->lru);
1628                 list_add(&e->lru, &wbl.list);
1629                 wbl.size++;
1630                 e->write_in_progress = true;
1631                 e->wc_list_contiguous = 1;
1632
1633                 f = e;
1634
1635                 while (1) {
1636                         next_node = rb_next(&f->rb_node);
1637                         if (unlikely(!next_node))
1638                                 break;
1639                         g = container_of(next_node, struct wc_entry, rb_node);
1640                         if (unlikely(read_original_sector(wc, g) ==
1641                             read_original_sector(wc, f))) {
1642                                 f = g;
1643                                 continue;
1644                         }
1645                         if (read_original_sector(wc, g) !=
1646                             read_original_sector(wc, f) + (wc->block_size >> SECTOR_SHIFT))
1647                                 break;
1648                         if (unlikely(g->write_in_progress))
1649                                 break;
1650                         if (unlikely(!writecache_entry_is_committed(wc, g)))
1651                                 break;
1652
1653                         if (!WC_MODE_PMEM(wc)) {
1654                                 if (g != f + 1)
1655                                         break;
1656                         }
1657
1658                         n_walked++;
1659                         //if (unlikely(n_walked > WRITEBACK_LATENCY) && likely(!wc->writeback_all))
1660                         //      break;
1661
1662                         wc->writeback_size++;
1663                         list_del(&g->lru);
1664                         list_add(&g->lru, &wbl.list);
1665                         wbl.size++;
1666                         g->write_in_progress = true;
1667                         g->wc_list_contiguous = BIO_MAX_PAGES;
1668                         f = g;
1669                         e->wc_list_contiguous++;
1670                         if (unlikely(e->wc_list_contiguous == BIO_MAX_PAGES)) {
1671                                 if (unlikely(wc->writeback_all)) {
1672                                         next_node = rb_next(&f->rb_node);
1673                                         if (likely(next_node))
1674                                                 g = container_of(next_node, struct wc_entry, rb_node);
1675                                 }
1676                                 break;
1677                         }
1678                 }
1679                 cond_resched();
1680         }
1681
1682         if (!list_empty(&skipped)) {
1683                 list_splice_tail(&skipped, &wc->lru);
1684                 /*
1685                  * If we didn't do any progress, we must wait until some
1686                  * writeback finishes to avoid burning CPU in a loop
1687                  */
1688                 if (unlikely(!wbl.size))
1689                         writecache_wait_for_writeback(wc);
1690         }
1691
1692         wc_unlock(wc);
1693
1694         blk_start_plug(&plug);
1695
1696         if (WC_MODE_PMEM(wc))
1697                 __writecache_writeback_pmem(wc, &wbl);
1698         else
1699                 __writecache_writeback_ssd(wc, &wbl);
1700
1701         blk_finish_plug(&plug);
1702
1703         if (unlikely(wc->writeback_all)) {
1704                 wc_lock(wc);
1705                 while (writecache_wait_for_writeback(wc));
1706                 wc_unlock(wc);
1707         }
1708 }
1709
1710 static int calculate_memory_size(uint64_t device_size, unsigned block_size,
1711                                  size_t *n_blocks_p, size_t *n_metadata_blocks_p)
1712 {
1713         uint64_t n_blocks, offset;
1714         struct wc_entry e;
1715
1716         n_blocks = device_size;
1717         do_div(n_blocks, block_size + sizeof(struct wc_memory_entry));
1718
1719         while (1) {
1720                 if (!n_blocks)
1721                         return -ENOSPC;
1722                 /* Verify the following entries[n_blocks] won't overflow */
1723                 if (n_blocks >= ((size_t)-sizeof(struct wc_memory_superblock) /
1724                                  sizeof(struct wc_memory_entry)))
1725                         return -EFBIG;
1726                 offset = offsetof(struct wc_memory_superblock, entries[n_blocks]);
1727                 offset = (offset + block_size - 1) & ~(uint64_t)(block_size - 1);
1728                 if (offset + n_blocks * block_size <= device_size)
1729                         break;
1730                 n_blocks--;
1731         }
1732
1733         /* check if the bit field overflows */
1734         e.index = n_blocks;
1735         if (e.index != n_blocks)
1736                 return -EFBIG;
1737
1738         if (n_blocks_p)
1739                 *n_blocks_p = n_blocks;
1740         if (n_metadata_blocks_p)
1741                 *n_metadata_blocks_p = offset >> __ffs(block_size);
1742         return 0;
1743 }
1744
1745 static int init_memory(struct dm_writecache *wc)
1746 {
1747         size_t b;
1748         int r;
1749
1750         r = calculate_memory_size(wc->memory_map_size, wc->block_size, &wc->n_blocks, NULL);
1751         if (r)
1752                 return r;
1753
1754         r = writecache_alloc_entries(wc);
1755         if (r)
1756                 return r;
1757
1758         for (b = 0; b < ARRAY_SIZE(sb(wc)->padding); b++)
1759                 pmem_assign(sb(wc)->padding[b], cpu_to_le64(0));
1760         pmem_assign(sb(wc)->version, cpu_to_le32(MEMORY_SUPERBLOCK_VERSION));
1761         pmem_assign(sb(wc)->block_size, cpu_to_le32(wc->block_size));
1762         pmem_assign(sb(wc)->n_blocks, cpu_to_le64(wc->n_blocks));
1763         pmem_assign(sb(wc)->seq_count, cpu_to_le64(0));
1764
1765         for (b = 0; b < wc->n_blocks; b++)
1766                 write_original_sector_seq_count(wc, &wc->entries[b], -1, -1);
1767
1768         writecache_flush_all_metadata(wc);
1769         writecache_commit_flushed(wc);
1770         pmem_assign(sb(wc)->magic, cpu_to_le32(MEMORY_SUPERBLOCK_MAGIC));
1771         writecache_flush_region(wc, &sb(wc)->magic, sizeof sb(wc)->magic);
1772         writecache_commit_flushed(wc);
1773
1774         return 0;
1775 }
1776
1777 static void writecache_dtr(struct dm_target *ti)
1778 {
1779         struct dm_writecache *wc = ti->private;
1780
1781         if (!wc)
1782                 return;
1783
1784         if (wc->endio_thread)
1785                 kthread_stop(wc->endio_thread);
1786
1787         if (wc->flush_thread)
1788                 kthread_stop(wc->flush_thread);
1789
1790         bioset_exit(&wc->bio_set);
1791
1792         mempool_exit(&wc->copy_pool);
1793
1794         if (wc->writeback_wq)
1795                 destroy_workqueue(wc->writeback_wq);
1796
1797         if (wc->dev)
1798                 dm_put_device(ti, wc->dev);
1799
1800         if (wc->ssd_dev)
1801                 dm_put_device(ti, wc->ssd_dev);
1802
1803         if (wc->entries)
1804                 vfree(wc->entries);
1805
1806         if (wc->memory_map) {
1807                 if (WC_MODE_PMEM(wc))
1808                         persistent_memory_release(wc);
1809                 else
1810                         vfree(wc->memory_map);
1811         }
1812
1813         if (wc->dm_kcopyd)
1814                 dm_kcopyd_client_destroy(wc->dm_kcopyd);
1815
1816         if (wc->dm_io)
1817                 dm_io_client_destroy(wc->dm_io);
1818
1819         if (wc->dirty_bitmap)
1820                 vfree(wc->dirty_bitmap);
1821
1822         kfree(wc);
1823 }
1824
1825 static int writecache_ctr(struct dm_target *ti, unsigned argc, char **argv)
1826 {
1827         struct dm_writecache *wc;
1828         struct dm_arg_set as;
1829         const char *string;
1830         unsigned opt_params;
1831         size_t offset, data_size;
1832         int i, r;
1833         char dummy;
1834         int high_wm_percent = HIGH_WATERMARK;
1835         int low_wm_percent = LOW_WATERMARK;
1836         uint64_t x;
1837         struct wc_memory_superblock s;
1838
1839         static struct dm_arg _args[] = {
1840                 {0, 10, "Invalid number of feature args"},
1841         };
1842
1843         as.argc = argc;
1844         as.argv = argv;
1845
1846         wc = kzalloc(sizeof(struct dm_writecache), GFP_KERNEL);
1847         if (!wc) {
1848                 ti->error = "Cannot allocate writecache structure";
1849                 r = -ENOMEM;
1850                 goto bad;
1851         }
1852         ti->private = wc;
1853         wc->ti = ti;
1854
1855         mutex_init(&wc->lock);
1856         writecache_poison_lists(wc);
1857         init_waitqueue_head(&wc->freelist_wait);
1858         timer_setup(&wc->autocommit_timer, writecache_autocommit_timer, 0);
1859
1860         for (i = 0; i < 2; i++) {
1861                 atomic_set(&wc->bio_in_progress[i], 0);
1862                 init_waitqueue_head(&wc->bio_in_progress_wait[i]);
1863         }
1864
1865         wc->dm_io = dm_io_client_create();
1866         if (IS_ERR(wc->dm_io)) {
1867                 r = PTR_ERR(wc->dm_io);
1868                 ti->error = "Unable to allocate dm-io client";
1869                 wc->dm_io = NULL;
1870                 goto bad;
1871         }
1872
1873         wc->writeback_wq = alloc_workqueue("writecache-writeback", WQ_MEM_RECLAIM, 1);
1874         if (!wc->writeback_wq) {
1875                 r = -ENOMEM;
1876                 ti->error = "Could not allocate writeback workqueue";
1877                 goto bad;
1878         }
1879         INIT_WORK(&wc->writeback_work, writecache_writeback);
1880         INIT_WORK(&wc->flush_work, writecache_flush_work);
1881
1882         raw_spin_lock_init(&wc->endio_list_lock);
1883         INIT_LIST_HEAD(&wc->endio_list);
1884         wc->endio_thread = kthread_create(writecache_endio_thread, wc, "writecache_endio");
1885         if (IS_ERR(wc->endio_thread)) {
1886                 r = PTR_ERR(wc->endio_thread);
1887                 wc->endio_thread = NULL;
1888                 ti->error = "Couldn't spawn endio thread";
1889                 goto bad;
1890         }
1891         wake_up_process(wc->endio_thread);
1892
1893         /*
1894          * Parse the mode (pmem or ssd)
1895          */
1896         string = dm_shift_arg(&as);
1897         if (!string)
1898                 goto bad_arguments;
1899
1900         if (!strcasecmp(string, "s")) {
1901                 wc->pmem_mode = false;
1902         } else if (!strcasecmp(string, "p")) {
1903 #ifdef DM_WRITECACHE_HAS_PMEM
1904                 wc->pmem_mode = true;
1905                 wc->writeback_fua = true;
1906 #else
1907                 /*
1908                  * If the architecture doesn't support persistent memory or
1909                  * the kernel doesn't support any DAX drivers, this driver can
1910                  * only be used in SSD-only mode.
1911                  */
1912                 r = -EOPNOTSUPP;
1913                 ti->error = "Persistent memory or DAX not supported on this system";
1914                 goto bad;
1915 #endif
1916         } else {
1917                 goto bad_arguments;
1918         }
1919
1920         if (WC_MODE_PMEM(wc)) {
1921                 r = bioset_init(&wc->bio_set, BIO_POOL_SIZE,
1922                                 offsetof(struct writeback_struct, bio),
1923                                 BIOSET_NEED_BVECS);
1924                 if (r) {
1925                         ti->error = "Could not allocate bio set";
1926                         goto bad;
1927                 }
1928         } else {
1929                 r = mempool_init_kmalloc_pool(&wc->copy_pool, 1, sizeof(struct copy_struct));
1930                 if (r) {
1931                         ti->error = "Could not allocate mempool";
1932                         goto bad;
1933                 }
1934         }
1935
1936         /*
1937          * Parse the origin data device
1938          */
1939         string = dm_shift_arg(&as);
1940         if (!string)
1941                 goto bad_arguments;
1942         r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->dev);
1943         if (r) {
1944                 ti->error = "Origin data device lookup failed";
1945                 goto bad;
1946         }
1947
1948         /*
1949          * Parse cache data device (be it pmem or ssd)
1950          */
1951         string = dm_shift_arg(&as);
1952         if (!string)
1953                 goto bad_arguments;
1954
1955         r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->ssd_dev);
1956         if (r) {
1957                 ti->error = "Cache data device lookup failed";
1958                 goto bad;
1959         }
1960         wc->memory_map_size = i_size_read(wc->ssd_dev->bdev->bd_inode);
1961
1962         /*
1963          * Parse the cache block size
1964          */
1965         string = dm_shift_arg(&as);
1966         if (!string)
1967                 goto bad_arguments;
1968         if (sscanf(string, "%u%c", &wc->block_size, &dummy) != 1 ||
1969             wc->block_size < 512 || wc->block_size > PAGE_SIZE ||
1970             (wc->block_size & (wc->block_size - 1))) {
1971                 r = -EINVAL;
1972                 ti->error = "Invalid block size";
1973                 goto bad;
1974         }
1975         wc->block_size_bits = __ffs(wc->block_size);
1976
1977         wc->max_writeback_jobs = MAX_WRITEBACK_JOBS;
1978         wc->autocommit_blocks = !WC_MODE_PMEM(wc) ? AUTOCOMMIT_BLOCKS_SSD : AUTOCOMMIT_BLOCKS_PMEM;
1979         wc->autocommit_jiffies = msecs_to_jiffies(AUTOCOMMIT_MSEC);
1980
1981         /*
1982          * Parse optional arguments
1983          */
1984         r = dm_read_arg_group(_args, &as, &opt_params, &ti->error);
1985         if (r)
1986                 goto bad;
1987
1988         while (opt_params) {
1989                 string = dm_shift_arg(&as), opt_params--;
1990                 if (!strcasecmp(string, "start_sector") && opt_params >= 1) {
1991                         unsigned long long start_sector;
1992                         string = dm_shift_arg(&as), opt_params--;
1993                         if (sscanf(string, "%llu%c", &start_sector, &dummy) != 1)
1994                                 goto invalid_optional;
1995                         wc->start_sector = start_sector;
1996                         if (wc->start_sector != start_sector ||
1997                             wc->start_sector >= wc->memory_map_size >> SECTOR_SHIFT)
1998                                 goto invalid_optional;
1999                 } else if (!strcasecmp(string, "high_watermark") && opt_params >= 1) {
2000                         string = dm_shift_arg(&as), opt_params--;
2001                         if (sscanf(string, "%d%c", &high_wm_percent, &dummy) != 1)
2002                                 goto invalid_optional;
2003                         if (high_wm_percent < 0 || high_wm_percent > 100)
2004                                 goto invalid_optional;
2005                         wc->high_wm_percent_set = true;
2006                 } else if (!strcasecmp(string, "low_watermark") && opt_params >= 1) {
2007                         string = dm_shift_arg(&as), opt_params--;
2008                         if (sscanf(string, "%d%c", &low_wm_percent, &dummy) != 1)
2009                                 goto invalid_optional;
2010                         if (low_wm_percent < 0 || low_wm_percent > 100)
2011                                 goto invalid_optional;
2012                         wc->low_wm_percent_set = true;
2013                 } else if (!strcasecmp(string, "writeback_jobs") && opt_params >= 1) {
2014                         string = dm_shift_arg(&as), opt_params--;
2015                         if (sscanf(string, "%u%c", &wc->max_writeback_jobs, &dummy) != 1)
2016                                 goto invalid_optional;
2017                         wc->max_writeback_jobs_set = true;
2018                 } else if (!strcasecmp(string, "autocommit_blocks") && opt_params >= 1) {
2019                         string = dm_shift_arg(&as), opt_params--;
2020                         if (sscanf(string, "%u%c", &wc->autocommit_blocks, &dummy) != 1)
2021                                 goto invalid_optional;
2022                         wc->autocommit_blocks_set = true;
2023                 } else if (!strcasecmp(string, "autocommit_time") && opt_params >= 1) {
2024                         unsigned autocommit_msecs;
2025                         string = dm_shift_arg(&as), opt_params--;
2026                         if (sscanf(string, "%u%c", &autocommit_msecs, &dummy) != 1)
2027                                 goto invalid_optional;
2028                         if (autocommit_msecs > 3600000)
2029                                 goto invalid_optional;
2030                         wc->autocommit_jiffies = msecs_to_jiffies(autocommit_msecs);
2031                         wc->autocommit_time_set = true;
2032                 } else if (!strcasecmp(string, "fua")) {
2033                         if (WC_MODE_PMEM(wc)) {
2034                                 wc->writeback_fua = true;
2035                                 wc->writeback_fua_set = true;
2036                         } else goto invalid_optional;
2037                 } else if (!strcasecmp(string, "nofua")) {
2038                         if (WC_MODE_PMEM(wc)) {
2039                                 wc->writeback_fua = false;
2040                                 wc->writeback_fua_set = true;
2041                         } else goto invalid_optional;
2042                 } else {
2043 invalid_optional:
2044                         r = -EINVAL;
2045                         ti->error = "Invalid optional argument";
2046                         goto bad;
2047                 }
2048         }
2049
2050         if (high_wm_percent < low_wm_percent) {
2051                 r = -EINVAL;
2052                 ti->error = "High watermark must be greater than or equal to low watermark";
2053                 goto bad;
2054         }
2055
2056         if (WC_MODE_PMEM(wc)) {
2057                 r = persistent_memory_claim(wc);
2058                 if (r) {
2059                         ti->error = "Unable to map persistent memory for cache";
2060                         goto bad;
2061                 }
2062         } else {
2063                 struct dm_io_region region;
2064                 struct dm_io_request req;
2065                 size_t n_blocks, n_metadata_blocks;
2066                 uint64_t n_bitmap_bits;
2067
2068                 wc->memory_map_size -= (uint64_t)wc->start_sector << SECTOR_SHIFT;
2069
2070                 bio_list_init(&wc->flush_list);
2071                 wc->flush_thread = kthread_create(writecache_flush_thread, wc, "dm_writecache_flush");
2072                 if (IS_ERR(wc->flush_thread)) {
2073                         r = PTR_ERR(wc->flush_thread);
2074                         wc->flush_thread = NULL;
2075                         ti->error = "Couldn't spawn flush thread";
2076                         goto bad;
2077                 }
2078                 wake_up_process(wc->flush_thread);
2079
2080                 r = calculate_memory_size(wc->memory_map_size, wc->block_size,
2081                                           &n_blocks, &n_metadata_blocks);
2082                 if (r) {
2083                         ti->error = "Invalid device size";
2084                         goto bad;
2085                 }
2086
2087                 n_bitmap_bits = (((uint64_t)n_metadata_blocks << wc->block_size_bits) +
2088                                  BITMAP_GRANULARITY - 1) / BITMAP_GRANULARITY;
2089                 /* this is limitation of test_bit functions */
2090                 if (n_bitmap_bits > 1U << 31) {
2091                         r = -EFBIG;
2092                         ti->error = "Invalid device size";
2093                         goto bad;
2094                 }
2095
2096                 wc->memory_map = vmalloc(n_metadata_blocks << wc->block_size_bits);
2097                 if (!wc->memory_map) {
2098                         r = -ENOMEM;
2099                         ti->error = "Unable to allocate memory for metadata";
2100                         goto bad;
2101                 }
2102
2103                 wc->dm_kcopyd = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2104                 if (IS_ERR(wc->dm_kcopyd)) {
2105                         r = PTR_ERR(wc->dm_kcopyd);
2106                         ti->error = "Unable to allocate dm-kcopyd client";
2107                         wc->dm_kcopyd = NULL;
2108                         goto bad;
2109                 }
2110
2111                 wc->metadata_sectors = n_metadata_blocks << (wc->block_size_bits - SECTOR_SHIFT);
2112                 wc->dirty_bitmap_size = (n_bitmap_bits + BITS_PER_LONG - 1) /
2113                         BITS_PER_LONG * sizeof(unsigned long);
2114                 wc->dirty_bitmap = vzalloc(wc->dirty_bitmap_size);
2115                 if (!wc->dirty_bitmap) {
2116                         r = -ENOMEM;
2117                         ti->error = "Unable to allocate dirty bitmap";
2118                         goto bad;
2119                 }
2120
2121                 region.bdev = wc->ssd_dev->bdev;
2122                 region.sector = wc->start_sector;
2123                 region.count = wc->metadata_sectors;
2124                 req.bi_op = REQ_OP_READ;
2125                 req.bi_op_flags = REQ_SYNC;
2126                 req.mem.type = DM_IO_VMA;
2127                 req.mem.ptr.vma = (char *)wc->memory_map;
2128                 req.client = wc->dm_io;
2129                 req.notify.fn = NULL;
2130
2131                 r = dm_io(&req, 1, &region, NULL);
2132                 if (r) {
2133                         ti->error = "Unable to read metadata";
2134                         goto bad;
2135                 }
2136         }
2137
2138         r = memcpy_mcsafe(&s, sb(wc), sizeof(struct wc_memory_superblock));
2139         if (r) {
2140                 ti->error = "Hardware memory error when reading superblock";
2141                 goto bad;
2142         }
2143         if (!le32_to_cpu(s.magic) && !le32_to_cpu(s.version)) {
2144                 r = init_memory(wc);
2145                 if (r) {
2146                         ti->error = "Unable to initialize device";
2147                         goto bad;
2148                 }
2149                 r = memcpy_mcsafe(&s, sb(wc), sizeof(struct wc_memory_superblock));
2150                 if (r) {
2151                         ti->error = "Hardware memory error when reading superblock";
2152                         goto bad;
2153                 }
2154         }
2155
2156         if (le32_to_cpu(s.magic) != MEMORY_SUPERBLOCK_MAGIC) {
2157                 ti->error = "Invalid magic in the superblock";
2158                 r = -EINVAL;
2159                 goto bad;
2160         }
2161
2162         if (le32_to_cpu(s.version) != MEMORY_SUPERBLOCK_VERSION) {
2163                 ti->error = "Invalid version in the superblock";
2164                 r = -EINVAL;
2165                 goto bad;
2166         }
2167
2168         if (le32_to_cpu(s.block_size) != wc->block_size) {
2169                 ti->error = "Block size does not match superblock";
2170                 r = -EINVAL;
2171                 goto bad;
2172         }
2173
2174         wc->n_blocks = le64_to_cpu(s.n_blocks);
2175
2176         offset = wc->n_blocks * sizeof(struct wc_memory_entry);
2177         if (offset / sizeof(struct wc_memory_entry) != le64_to_cpu(sb(wc)->n_blocks)) {
2178 overflow:
2179                 ti->error = "Overflow in size calculation";
2180                 r = -EINVAL;
2181                 goto bad;
2182         }
2183         offset += sizeof(struct wc_memory_superblock);
2184         if (offset < sizeof(struct wc_memory_superblock))
2185                 goto overflow;
2186         offset = (offset + wc->block_size - 1) & ~(size_t)(wc->block_size - 1);
2187         data_size = wc->n_blocks * (size_t)wc->block_size;
2188         if (!offset || (data_size / wc->block_size != wc->n_blocks) ||
2189             (offset + data_size < offset))
2190                 goto overflow;
2191         if (offset + data_size > wc->memory_map_size) {
2192                 ti->error = "Memory area is too small";
2193                 r = -EINVAL;
2194                 goto bad;
2195         }
2196
2197         wc->metadata_sectors = offset >> SECTOR_SHIFT;
2198         wc->block_start = (char *)sb(wc) + offset;
2199
2200         x = (uint64_t)wc->n_blocks * (100 - high_wm_percent);
2201         x += 50;
2202         do_div(x, 100);
2203         wc->freelist_high_watermark = x;
2204         x = (uint64_t)wc->n_blocks * (100 - low_wm_percent);
2205         x += 50;
2206         do_div(x, 100);
2207         wc->freelist_low_watermark = x;
2208
2209         r = writecache_alloc_entries(wc);
2210         if (r) {
2211                 ti->error = "Cannot allocate memory";
2212                 goto bad;
2213         }
2214
2215         ti->num_flush_bios = 1;
2216         ti->flush_supported = true;
2217         ti->num_discard_bios = 1;
2218
2219         if (WC_MODE_PMEM(wc))
2220                 persistent_memory_flush_cache(wc->memory_map, wc->memory_map_size);
2221
2222         return 0;
2223
2224 bad_arguments:
2225         r = -EINVAL;
2226         ti->error = "Bad arguments";
2227 bad:
2228         writecache_dtr(ti);
2229         return r;
2230 }
2231
2232 static void writecache_status(struct dm_target *ti, status_type_t type,
2233                               unsigned status_flags, char *result, unsigned maxlen)
2234 {
2235         struct dm_writecache *wc = ti->private;
2236         unsigned extra_args;
2237         unsigned sz = 0;
2238         uint64_t x;
2239
2240         switch (type) {
2241         case STATUSTYPE_INFO:
2242                 DMEMIT("%ld %llu %llu %llu", writecache_has_error(wc),
2243                        (unsigned long long)wc->n_blocks, (unsigned long long)wc->freelist_size,
2244                        (unsigned long long)wc->writeback_size);
2245                 break;
2246         case STATUSTYPE_TABLE:
2247                 DMEMIT("%c %s %s %u ", WC_MODE_PMEM(wc) ? 'p' : 's',
2248                                 wc->dev->name, wc->ssd_dev->name, wc->block_size);
2249                 extra_args = 0;
2250                 if (wc->start_sector)
2251                         extra_args += 2;
2252                 if (wc->high_wm_percent_set)
2253                         extra_args += 2;
2254                 if (wc->low_wm_percent_set)
2255                         extra_args += 2;
2256                 if (wc->max_writeback_jobs_set)
2257                         extra_args += 2;
2258                 if (wc->autocommit_blocks_set)
2259                         extra_args += 2;
2260                 if (wc->autocommit_time_set)
2261                         extra_args += 2;
2262                 if (wc->writeback_fua_set)
2263                         extra_args++;
2264
2265                 DMEMIT("%u", extra_args);
2266                 if (wc->start_sector)
2267                         DMEMIT(" start_sector %llu", (unsigned long long)wc->start_sector);
2268                 if (wc->high_wm_percent_set) {
2269                         x = (uint64_t)wc->freelist_high_watermark * 100;
2270                         x += wc->n_blocks / 2;
2271                         do_div(x, (size_t)wc->n_blocks);
2272                         DMEMIT(" high_watermark %u", 100 - (unsigned)x);
2273                 }
2274                 if (wc->low_wm_percent_set) {
2275                         x = (uint64_t)wc->freelist_low_watermark * 100;
2276                         x += wc->n_blocks / 2;
2277                         do_div(x, (size_t)wc->n_blocks);
2278                         DMEMIT(" low_watermark %u", 100 - (unsigned)x);
2279                 }
2280                 if (wc->max_writeback_jobs_set)
2281                         DMEMIT(" writeback_jobs %u", wc->max_writeback_jobs);
2282                 if (wc->autocommit_blocks_set)
2283                         DMEMIT(" autocommit_blocks %u", wc->autocommit_blocks);
2284                 if (wc->autocommit_time_set)
2285                         DMEMIT(" autocommit_time %u", jiffies_to_msecs(wc->autocommit_jiffies));
2286                 if (wc->writeback_fua_set)
2287                         DMEMIT(" %sfua", wc->writeback_fua ? "" : "no");
2288                 break;
2289         }
2290 }
2291
2292 static struct target_type writecache_target = {
2293         .name                   = "writecache",
2294         .version                = {1, 1, 1},
2295         .module                 = THIS_MODULE,
2296         .ctr                    = writecache_ctr,
2297         .dtr                    = writecache_dtr,
2298         .status                 = writecache_status,
2299         .postsuspend            = writecache_suspend,
2300         .resume                 = writecache_resume,
2301         .message                = writecache_message,
2302         .map                    = writecache_map,
2303         .end_io                 = writecache_end_io,
2304         .iterate_devices        = writecache_iterate_devices,
2305         .io_hints               = writecache_io_hints,
2306 };
2307
2308 static int __init dm_writecache_init(void)
2309 {
2310         int r;
2311
2312         r = dm_register_target(&writecache_target);
2313         if (r < 0) {
2314                 DMERR("register failed %d", r);
2315                 return r;
2316         }
2317
2318         return 0;
2319 }
2320
2321 static void __exit dm_writecache_exit(void)
2322 {
2323         dm_unregister_target(&writecache_target);
2324 }
2325
2326 module_init(dm_writecache_init);
2327 module_exit(dm_writecache_exit);
2328
2329 MODULE_DESCRIPTION(DM_NAME " writecache target");
2330 MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
2331 MODULE_LICENSE("GPL");