drivers/md/dm-writecache.c

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