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                 cond_resched();
 876         }
 877
 878         return 0;
 879 }
 880
 881 static void writecache_resume(struct dm_target *ti)
 882 {
 883         struct dm_writecache *wc = ti->private;
 884         size_t b;
 885         bool need_flush = false;
 886         __le64 sb_seq_count;
 887         int r;
 888
 889         wc_lock(wc);
 890
 891         if (WC_MODE_PMEM(wc))
 892                 persistent_memory_invalidate_cache(wc->memory_map, wc->memory_map_size);
 893
 894         wc->tree = RB_ROOT;
 895         INIT_LIST_HEAD(&wc->lru);
 896         if (WC_MODE_SORT_FREELIST(wc)) {
 897                 wc->freetree = RB_ROOT;
 898                 wc->current_free = NULL;
 899         } else {
 900                 INIT_LIST_HEAD(&wc->freelist);
 901         }
 902         wc->freelist_size = 0;
 903
 904         r = memcpy_mcsafe(&sb_seq_count, &sb(wc)->seq_count, sizeof(uint64_t));
 905         if (r) {
 906                 writecache_error(wc, r, "hardware memory error when reading superblock: %d", r);
 907                 sb_seq_count = cpu_to_le64(0);
 908         }
 909         wc->seq_count = le64_to_cpu(sb_seq_count);
 910
 911 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 912         for (b = 0; b < wc->n_blocks; b++) {
 913                 struct wc_entry *e = &wc->entries[b];
 914                 struct wc_memory_entry wme;
 915                 if (writecache_has_error(wc)) {
 916                         e->original_sector = -1;
 917                         e->seq_count = -1;
 918                         continue;
 919                 }
 920                 r = memcpy_mcsafe(&wme, memory_entry(wc, e), sizeof(struct wc_memory_entry));
 921                 if (r) {
 922                         writecache_error(wc, r, "hardware memory error when reading metadata entry %lu: %d",
 923                                          (unsigned long)b, r);
 924                         e->original_sector = -1;
 925                         e->seq_count = -1;
 926                 } else {
 927                         e->original_sector = le64_to_cpu(wme.original_sector);
 928                         e->seq_count = le64_to_cpu(wme.seq_count);
 929                 }
 930                 cond_resched();
 931         }
 932 #endif
 933         for (b = 0; b < wc->n_blocks; b++) {
 934                 struct wc_entry *e = &wc->entries[b];
 935                 if (!writecache_entry_is_committed(wc, e)) {
 936                         if (read_seq_count(wc, e) != -1) {
 937 erase_this:
 938                                 clear_seq_count(wc, e);
 939                                 need_flush = true;
 940                         }
 941                         writecache_add_to_freelist(wc, e);
 942                 } else {
 943                         struct wc_entry *old;
 944
 945                         old = writecache_find_entry(wc, read_original_sector(wc, e), 0);
 946                         if (!old) {
 947                                 writecache_insert_entry(wc, e);
 948                         } else {
 949                                 if (read_seq_count(wc, old) == read_seq_count(wc, e)) {
 950                                         writecache_error(wc, -EINVAL,
 951                                                  "two identical entries, position %llu, sector %llu, sequence %llu",
 952                                                  (unsigned long long)b, (unsigned long long)read_original_sector(wc, e),
 953                                                  (unsigned long long)read_seq_count(wc, e));
 954                                 }
 955                                 if (read_seq_count(wc, old) > read_seq_count(wc, e)) {
 956                                         goto erase_this;
 957                                 } else {
 958                                         writecache_free_entry(wc, old);
 959                                         writecache_insert_entry(wc, e);
 960                                         need_flush = true;
 961                                 }
 962                         }
 963                 }
 964                 cond_resched();
 965         }
 966
 967         if (need_flush) {
 968                 writecache_flush_all_metadata(wc);
 969                 writecache_commit_flushed(wc, false);
 970         }
 971
 972         writecache_verify_watermark(wc);
 973
 974         wc_unlock(wc);
 975 }
 976
 977 static int process_flush_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
 978 {
 979         if (argc != 1)
 980                 return -EINVAL;
 981
 982         wc_lock(wc);
 983         if (dm_suspended(wc->ti)) {
 984                 wc_unlock(wc);
 985                 return -EBUSY;
 986         }
 987         if (writecache_has_error(wc)) {
 988                 wc_unlock(wc);
 989                 return -EIO;
 990         }
 991
 992         writecache_flush(wc);
 993         wc->writeback_all++;
 994         queue_work(wc->writeback_wq, &wc->writeback_work);
 995         wc_unlock(wc);
 996
 997         flush_workqueue(wc->writeback_wq);
 998
 999         wc_lock(wc);
1000         wc->writeback_all--;
1001         if (writecache_has_error(wc)) {
1002                 wc_unlock(wc);
1003                 return -EIO;
1004         }
1005         wc_unlock(wc);
1006
1007         return 0;
1008 }
1009
1010 static int process_flush_on_suspend_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1011 {
1012         if (argc != 1)
1013                 return -EINVAL;
1014
1015         wc_lock(wc);
1016         wc->flush_on_suspend = true;
1017         wc_unlock(wc);
1018
1019         return 0;
1020 }
1021
1022 static int writecache_message(struct dm_target *ti, unsigned argc, char **argv,
1023                               char *result, unsigned maxlen)
1024 {
1025         int r = -EINVAL;
1026         struct dm_writecache *wc = ti->private;
1027
1028         if (!strcasecmp(argv[0], "flush"))
1029                 r = process_flush_mesg(argc, argv, wc);
1030         else if (!strcasecmp(argv[0], "flush_on_suspend"))
1031                 r = process_flush_on_suspend_mesg(argc, argv, wc);
1032         else
1033                 DMERR("unrecognised message received: %s", argv[0]);
1034
1035         return r;
1036 }
1037
1038 static void bio_copy_block(struct dm_writecache *wc, struct bio *bio, void *data)
1039 {
1040         void *buf;
1041         unsigned long flags;
1042         unsigned size;
1043         int rw = bio_data_dir(bio);
1044         unsigned remaining_size = wc->block_size;
1045
1046         do {
1047                 struct bio_vec bv = bio_iter_iovec(bio, bio->bi_iter);
1048                 buf = bvec_kmap_irq(&bv, &flags);
1049                 size = bv.bv_len;
1050                 if (unlikely(size > remaining_size))
1051                         size = remaining_size;
1052
1053                 if (rw == READ) {
1054                         int r;
1055                         r = memcpy_mcsafe(buf, data, size);
1056                         flush_dcache_page(bio_page(bio));
1057                         if (unlikely(r)) {
1058                                 writecache_error(wc, r, "hardware memory error when reading data: %d", r);
1059                                 bio->bi_status = BLK_STS_IOERR;
1060                         }
1061                 } else {
1062                         flush_dcache_page(bio_page(bio));
1063                         memcpy_flushcache(data, buf, size);
1064                 }
1065
1066                 bvec_kunmap_irq(buf, &flags);
1067
1068                 data = (char *)data + size;
1069                 remaining_size -= size;
1070                 bio_advance(bio, size);
1071         } while (unlikely(remaining_size));
1072 }
1073
1074 static int writecache_flush_thread(void *data)
1075 {
1076         struct dm_writecache *wc = data;
1077
1078         while (1) {
1079                 struct bio *bio;
1080
1081                 wc_lock(wc);
1082                 bio = bio_list_pop(&wc->flush_list);
1083                 if (!bio) {
1084                         set_current_state(TASK_INTERRUPTIBLE);
1085                         wc_unlock(wc);
1086
1087                         if (unlikely(kthread_should_stop())) {
1088                                 set_current_state(TASK_RUNNING);
1089                                 break;
1090                         }
1091
1092                         schedule();
1093                         continue;
1094                 }
1095
1096                 if (bio_op(bio) == REQ_OP_DISCARD) {
1097                         writecache_discard(wc, bio->bi_iter.bi_sector,
1098                                            bio_end_sector(bio));
1099                         wc_unlock(wc);
1100                         bio_set_dev(bio, wc->dev->bdev);
1101                         generic_make_request(bio);
1102                 } else {
1103                         writecache_flush(wc);
1104                         wc_unlock(wc);
1105                         if (writecache_has_error(wc))
1106                                 bio->bi_status = BLK_STS_IOERR;
1107                         bio_endio(bio);
1108                 }
1109         }
1110
1111         return 0;
1112 }
1113
1114 static void writecache_offload_bio(struct dm_writecache *wc, struct bio *bio)
1115 {
1116         if (bio_list_empty(&wc->flush_list))
1117                 wake_up_process(wc->flush_thread);
1118         bio_list_add(&wc->flush_list, bio);
1119 }
1120
1121 static int writecache_map(struct dm_target *ti, struct bio *bio)
1122 {
1123         struct wc_entry *e;
1124         struct dm_writecache *wc = ti->private;
1125
1126         bio->bi_private = NULL;
1127
1128         wc_lock(wc);
1129
1130         if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1131                 if (writecache_has_error(wc))
1132                         goto unlock_error;
1133                 if (WC_MODE_PMEM(wc)) {
1134                         writecache_flush(wc);
1135                         if (writecache_has_error(wc))
1136                                 goto unlock_error;
1137                         goto unlock_submit;
1138                 } else {
1139                         writecache_offload_bio(wc, bio);
1140                         goto unlock_return;
1141                 }
1142         }
1143
1144         bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
1145
1146         if (unlikely((((unsigned)bio->bi_iter.bi_sector | bio_sectors(bio)) &
1147                                 (wc->block_size / 512 - 1)) != 0)) {
1148                 DMERR("I/O is not aligned, sector %llu, size %u, block size %u",
1149                       (unsigned long long)bio->bi_iter.bi_sector,
1150                       bio->bi_iter.bi_size, wc->block_size);
1151                 goto unlock_error;
1152         }
1153
1154         if (unlikely(bio_op(bio) == REQ_OP_DISCARD)) {
1155                 if (writecache_has_error(wc))
1156                         goto unlock_error;
1157                 if (WC_MODE_PMEM(wc)) {
1158                         writecache_discard(wc, bio->bi_iter.bi_sector, bio_end_sector(bio));
1159                         goto unlock_remap_origin;
1160                 } else {
1161                         writecache_offload_bio(wc, bio);
1162                         goto unlock_return;
1163                 }
1164         }
1165
1166         if (bio_data_dir(bio) == READ) {
1167 read_next_block:
1168                 e = writecache_find_entry(wc, bio->bi_iter.bi_sector, WFE_RETURN_FOLLOWING);
1169                 if (e && read_original_sector(wc, e) == bio->bi_iter.bi_sector) {
1170                         if (WC_MODE_PMEM(wc)) {
1171                                 bio_copy_block(wc, bio, memory_data(wc, e));
1172                                 if (bio->bi_iter.bi_size)
1173                                         goto read_next_block;
1174                                 goto unlock_submit;
1175                         } else {
1176                                 dm_accept_partial_bio(bio, wc->block_size >> SECTOR_SHIFT);
1177                                 bio_set_dev(bio, wc->ssd_dev->bdev);
1178                                 bio->bi_iter.bi_sector = cache_sector(wc, e);
1179                                 if (!writecache_entry_is_committed(wc, e))
1180                                         writecache_wait_for_ios(wc, WRITE);
1181                                 goto unlock_remap;
1182                         }
1183                 } else {
1184                         if (e) {
1185                                 sector_t next_boundary =
1186                                         read_original_sector(wc, e) - bio->bi_iter.bi_sector;
1187                                 if (next_boundary < bio->bi_iter.bi_size >> SECTOR_SHIFT) {
1188                                         dm_accept_partial_bio(bio, next_boundary);
1189                                 }
1190                         }
1191                         goto unlock_remap_origin;
1192                 }
1193         } else {
1194                 do {
1195                         if (writecache_has_error(wc))
1196                                 goto unlock_error;
1197                         e = writecache_find_entry(wc, bio->bi_iter.bi_sector, 0);
1198                         if (e) {
1199                                 if (!writecache_entry_is_committed(wc, e))
1200                                         goto bio_copy;
1201                                 if (!WC_MODE_PMEM(wc) && !e->write_in_progress) {
1202                                         wc->overwrote_committed = true;
1203                                         goto bio_copy;
1204                                 }
1205                         }
1206                         e = writecache_pop_from_freelist(wc);
1207                         if (unlikely(!e)) {
1208                                 writecache_wait_on_freelist(wc);
1209                                 continue;
1210                         }
1211                         write_original_sector_seq_count(wc, e, bio->bi_iter.bi_sector, wc->seq_count);
1212                         writecache_insert_entry(wc, e);
1213                         wc->uncommitted_blocks++;
1214 bio_copy:
1215                         if (WC_MODE_PMEM(wc)) {
1216                                 bio_copy_block(wc, bio, memory_data(wc, e));
1217                         } else {
1218                                 dm_accept_partial_bio(bio, wc->block_size >> SECTOR_SHIFT);
1219                                 bio_set_dev(bio, wc->ssd_dev->bdev);
1220                                 bio->bi_iter.bi_sector = cache_sector(wc, e);
1221                                 if (unlikely(wc->uncommitted_blocks >= wc->autocommit_blocks)) {
1222                                         wc->uncommitted_blocks = 0;
1223                                         queue_work(wc->writeback_wq, &wc->flush_work);
1224                                 } else {
1225                                         writecache_schedule_autocommit(wc);
1226                                 }
1227                                 goto unlock_remap;
1228                         }
1229                 } while (bio->bi_iter.bi_size);
1230
1231                 if (unlikely(bio->bi_opf & REQ_FUA ||
1232                              wc->uncommitted_blocks >= wc->autocommit_blocks))
1233                         writecache_flush(wc);
1234                 else
1235                         writecache_schedule_autocommit(wc);
1236                 goto unlock_submit;
1237         }
1238
1239 unlock_remap_origin:
1240         bio_set_dev(bio, wc->dev->bdev);
1241         wc_unlock(wc);
1242         return DM_MAPIO_REMAPPED;
1243
1244 unlock_remap:
1245         /* make sure that writecache_end_io decrements bio_in_progress: */
1246         bio->bi_private = (void *)1;
1247         atomic_inc(&wc->bio_in_progress[bio_data_dir(bio)]);
1248         wc_unlock(wc);
1249         return DM_MAPIO_REMAPPED;
1250
1251 unlock_submit:
1252         wc_unlock(wc);
1253         bio_endio(bio);
1254         return DM_MAPIO_SUBMITTED;
1255
1256 unlock_return:
1257         wc_unlock(wc);
1258         return DM_MAPIO_SUBMITTED;
1259
1260 unlock_error:
1261         wc_unlock(wc);
1262         bio_io_error(bio);
1263         return DM_MAPIO_SUBMITTED;
1264 }
1265
1266 static int writecache_end_io(struct dm_target *ti, struct bio *bio, blk_status_t *status)
1267 {
1268         struct dm_writecache *wc = ti->private;
1269
1270         if (bio->bi_private != NULL) {
1271                 int dir = bio_data_dir(bio);
1272                 if (atomic_dec_and_test(&wc->bio_in_progress[dir]))
1273                         if (unlikely(waitqueue_active(&wc->bio_in_progress_wait[dir])))
1274                                 wake_up(&wc->bio_in_progress_wait[dir]);
1275         }
1276         return 0;
1277 }
1278
1279 static int writecache_iterate_devices(struct dm_target *ti,
1280                                       iterate_devices_callout_fn fn, void *data)
1281 {
1282         struct dm_writecache *wc = ti->private;
1283
1284         return fn(ti, wc->dev, 0, ti->len, data);
1285 }
1286
1287 static void writecache_io_hints(struct dm_target *ti, struct queue_limits *limits)
1288 {
1289         struct dm_writecache *wc = ti->private;
1290
1291         if (limits->logical_block_size < wc->block_size)
1292                 limits->logical_block_size = wc->block_size;
1293
1294         if (limits->physical_block_size < wc->block_size)
1295                 limits->physical_block_size = wc->block_size;
1296
1297         if (limits->io_min < wc->block_size)
1298                 limits->io_min = wc->block_size;
1299 }
1300
1301
1302 static void writecache_writeback_endio(struct bio *bio)
1303 {
1304         struct writeback_struct *wb = container_of(bio, struct writeback_struct, bio);
1305         struct dm_writecache *wc = wb->wc;
1306         unsigned long flags;
1307
1308         raw_spin_lock_irqsave(&wc->endio_list_lock, flags);
1309         if (unlikely(list_empty(&wc->endio_list)))
1310                 wake_up_process(wc->endio_thread);
1311         list_add_tail(&wb->endio_entry, &wc->endio_list);
1312         raw_spin_unlock_irqrestore(&wc->endio_list_lock, flags);
1313 }
1314
1315 static void writecache_copy_endio(int read_err, unsigned long write_err, void *ptr)
1316 {
1317         struct copy_struct *c = ptr;
1318         struct dm_writecache *wc = c->wc;
1319
1320         c->error = likely(!(read_err | write_err)) ? 0 : -EIO;
1321
1322         raw_spin_lock_irq(&wc->endio_list_lock);
1323         if (unlikely(list_empty(&wc->endio_list)))
1324                 wake_up_process(wc->endio_thread);
1325         list_add_tail(&c->endio_entry, &wc->endio_list);
1326         raw_spin_unlock_irq(&wc->endio_list_lock);
1327 }
1328
1329 static void __writecache_endio_pmem(struct dm_writecache *wc, struct list_head *list)
1330 {
1331         unsigned i;
1332         struct writeback_struct *wb;
1333         struct wc_entry *e;
1334         unsigned long n_walked = 0;
1335
1336         do {
1337                 wb = list_entry(list->next, struct writeback_struct, endio_entry);
1338                 list_del(&wb->endio_entry);
1339
1340                 if (unlikely(wb->bio.bi_status != BLK_STS_OK))
1341                         writecache_error(wc, blk_status_to_errno(wb->bio.bi_status),
1342                                         "write error %d", wb->bio.bi_status);
1343                 i = 0;
1344                 do {
1345                         e = wb->wc_list[i];
1346                         BUG_ON(!e->write_in_progress);
1347                         e->write_in_progress = false;
1348                         INIT_LIST_HEAD(&e->lru);
1349                         if (!writecache_has_error(wc))
1350                                 writecache_free_entry(wc, e);
1351                         BUG_ON(!wc->writeback_size);
1352                         wc->writeback_size--;
1353                         n_walked++;
1354                         if (unlikely(n_walked >= ENDIO_LATENCY)) {
1355                                 writecache_commit_flushed(wc, false);
1356                                 wc_unlock(wc);
1357                                 wc_lock(wc);
1358                                 n_walked = 0;
1359                         }
1360                 } while (++i < wb->wc_list_n);
1361
1362                 if (wb->wc_list != wb->wc_list_inline)
1363                         kfree(wb->wc_list);
1364                 bio_put(&wb->bio);
1365         } while (!list_empty(list));
1366 }
1367
1368 static void __writecache_endio_ssd(struct dm_writecache *wc, struct list_head *list)
1369 {
1370         struct copy_struct *c;
1371         struct wc_entry *e;
1372
1373         do {
1374                 c = list_entry(list->next, struct copy_struct, endio_entry);
1375                 list_del(&c->endio_entry);
1376
1377                 if (unlikely(c->error))
1378                         writecache_error(wc, c->error, "copy error");
1379
1380                 e = c->e;
1381                 do {
1382                         BUG_ON(!e->write_in_progress);
1383                         e->write_in_progress = false;
1384                         INIT_LIST_HEAD(&e->lru);
1385                         if (!writecache_has_error(wc))
1386                                 writecache_free_entry(wc, e);
1387
1388                         BUG_ON(!wc->writeback_size);
1389                         wc->writeback_size--;
1390                         e++;
1391                 } while (--c->n_entries);
1392                 mempool_free(c, &wc->copy_pool);
1393         } while (!list_empty(list));
1394 }
1395
1396 static int writecache_endio_thread(void *data)
1397 {
1398         struct dm_writecache *wc = data;
1399
1400         while (1) {
1401                 struct list_head list;
1402
1403                 raw_spin_lock_irq(&wc->endio_list_lock);
1404                 if (!list_empty(&wc->endio_list))
1405                         goto pop_from_list;
1406                 set_current_state(TASK_INTERRUPTIBLE);
1407                 raw_spin_unlock_irq(&wc->endio_list_lock);
1408
1409                 if (unlikely(kthread_should_stop())) {
1410                         set_current_state(TASK_RUNNING);
1411                         break;
1412                 }
1413
1414                 schedule();
1415
1416                 continue;
1417
1418 pop_from_list:
1419                 list = wc->endio_list;
1420                 list.next->prev = list.prev->next = &list;
1421                 INIT_LIST_HEAD(&wc->endio_list);
1422                 raw_spin_unlock_irq(&wc->endio_list_lock);
1423
1424                 if (!WC_MODE_FUA(wc))
1425                         writecache_disk_flush(wc, wc->dev);
1426
1427                 wc_lock(wc);
1428
1429                 if (WC_MODE_PMEM(wc)) {
1430                         __writecache_endio_pmem(wc, &list);
1431                 } else {
1432                         __writecache_endio_ssd(wc, &list);
1433                         writecache_wait_for_ios(wc, READ);
1434                 }
1435
1436                 writecache_commit_flushed(wc, false);
1437
1438                 wc_unlock(wc);
1439         }
1440
1441         return 0;
1442 }
1443
1444 static bool wc_add_block(struct writeback_struct *wb, struct wc_entry *e, gfp_t gfp)
1445 {
1446         struct dm_writecache *wc = wb->wc;
1447         unsigned block_size = wc->block_size;
1448         void *address = memory_data(wc, e);
1449
1450         persistent_memory_flush_cache(address, block_size);
1451         return bio_add_page(&wb->bio, persistent_memory_page(address),
1452                             block_size, persistent_memory_page_offset(address)) != 0;
1453 }
1454
1455 struct writeback_list {
1456         struct list_head list;
1457         size_t size;
1458 };
1459
1460 static void __writeback_throttle(struct dm_writecache *wc, struct writeback_list *wbl)
1461 {
1462         if (unlikely(wc->max_writeback_jobs)) {
1463                 if (READ_ONCE(wc->writeback_size) - wbl->size >= wc->max_writeback_jobs) {
1464                         wc_lock(wc);
1465                         while (wc->writeback_size - wbl->size >= wc->max_writeback_jobs)
1466                                 writecache_wait_on_freelist(wc);
1467                         wc_unlock(wc);
1468                 }
1469         }
1470         cond_resched();
1471 }
1472
1473 static void __writecache_writeback_pmem(struct dm_writecache *wc, struct writeback_list *wbl)
1474 {
1475         struct wc_entry *e, *f;
1476         struct bio *bio;
1477         struct writeback_struct *wb;
1478         unsigned max_pages;
1479
1480         while (wbl->size) {
1481                 wbl->size--;
1482                 e = container_of(wbl->list.prev, struct wc_entry, lru);
1483                 list_del(&e->lru);
1484
1485                 max_pages = e->wc_list_contiguous;
1486
1487                 bio = bio_alloc_bioset(GFP_NOIO, max_pages, &wc->bio_set);
1488                 wb = container_of(bio, struct writeback_struct, bio);
1489                 wb->wc = wc;
1490                 bio->bi_end_io = writecache_writeback_endio;
1491                 bio_set_dev(bio, wc->dev->bdev);
1492                 bio->bi_iter.bi_sector = read_original_sector(wc, e);
1493                 if (max_pages <= WB_LIST_INLINE ||
1494                     unlikely(!(wb->wc_list = kmalloc_array(max_pages, sizeof(struct wc_entry *),
1495                                                            GFP_NOIO | __GFP_NORETRY |
1496                                                            __GFP_NOMEMALLOC | __GFP_NOWARN)))) {
1497                         wb->wc_list = wb->wc_list_inline;
1498                         max_pages = WB_LIST_INLINE;
1499                 }
1500
1501                 BUG_ON(!wc_add_block(wb, e, GFP_NOIO));
1502
1503                 wb->wc_list[0] = e;
1504                 wb->wc_list_n = 1;
1505
1506                 while (wbl->size && wb->wc_list_n < max_pages) {
1507                         f = container_of(wbl->list.prev, struct wc_entry, lru);
1508                         if (read_original_sector(wc, f) !=
1509                             read_original_sector(wc, e) + (wc->block_size >> SECTOR_SHIFT))
1510                                 break;
1511                         if (!wc_add_block(wb, f, GFP_NOWAIT | __GFP_NOWARN))
1512                                 break;
1513                         wbl->size--;
1514                         list_del(&f->lru);
1515                         wb->wc_list[wb->wc_list_n++] = f;
1516                         e = f;
1517                 }
1518                 bio_set_op_attrs(bio, REQ_OP_WRITE, WC_MODE_FUA(wc) * REQ_FUA);
1519                 if (writecache_has_error(wc)) {
1520                         bio->bi_status = BLK_STS_IOERR;
1521                         bio_endio(bio);
1522                 } else {
1523                         submit_bio(bio);
1524                 }
1525
1526                 __writeback_throttle(wc, wbl);
1527         }
1528 }
1529
1530 static void __writecache_writeback_ssd(struct dm_writecache *wc, struct writeback_list *wbl)
1531 {
1532         struct wc_entry *e, *f;
1533         struct dm_io_region from, to;
1534         struct copy_struct *c;
1535
1536         while (wbl->size) {
1537                 unsigned n_sectors;
1538
1539                 wbl->size--;
1540                 e = container_of(wbl->list.prev, struct wc_entry, lru);
1541                 list_del(&e->lru);
1542
1543                 n_sectors = e->wc_list_contiguous << (wc->block_size_bits - SECTOR_SHIFT);
1544
1545                 from.bdev = wc->ssd_dev->bdev;
1546                 from.sector = cache_sector(wc, e);
1547                 from.count = n_sectors;
1548                 to.bdev = wc->dev->bdev;
1549                 to.sector = read_original_sector(wc, e);
1550                 to.count = n_sectors;
1551
1552                 c = mempool_alloc(&wc->copy_pool, GFP_NOIO);
1553                 c->wc = wc;
1554                 c->e = e;
1555                 c->n_entries = e->wc_list_contiguous;
1556
1557                 while ((n_sectors -= wc->block_size >> SECTOR_SHIFT)) {
1558                         wbl->size--;
1559                         f = container_of(wbl->list.prev, struct wc_entry, lru);
1560                         BUG_ON(f != e + 1);
1561                         list_del(&f->lru);
1562                         e = f;
1563                 }
1564
1565                 dm_kcopyd_copy(wc->dm_kcopyd, &from, 1, &to, 0, writecache_copy_endio, c);
1566
1567                 __writeback_throttle(wc, wbl);
1568         }
1569 }
1570
1571 static void writecache_writeback(struct work_struct *work)
1572 {
1573         struct dm_writecache *wc = container_of(work, struct dm_writecache, writeback_work);
1574         struct blk_plug plug;
1575         struct wc_entry *f, *g, *e = NULL;
1576         struct rb_node *node, *next_node;
1577         struct list_head skipped;
1578         struct writeback_list wbl;
1579         unsigned long n_walked;
1580
1581         wc_lock(wc);
1582 restart:
1583         if (writecache_has_error(wc)) {
1584                 wc_unlock(wc);
1585                 return;
1586         }
1587
1588         if (unlikely(wc->writeback_all)) {
1589                 if (writecache_wait_for_writeback(wc))
1590                         goto restart;
1591         }
1592
1593         if (wc->overwrote_committed) {
1594                 writecache_wait_for_ios(wc, WRITE);
1595         }
1596
1597         n_walked = 0;
1598         INIT_LIST_HEAD(&skipped);
1599         INIT_LIST_HEAD(&wbl.list);
1600         wbl.size = 0;
1601         while (!list_empty(&wc->lru) &&
1602                (wc->writeback_all ||
1603                 wc->freelist_size + wc->writeback_size <= wc->freelist_low_watermark)) {
1604
1605                 n_walked++;
1606                 if (unlikely(n_walked > WRITEBACK_LATENCY) &&
1607                     likely(!wc->writeback_all) && likely(!dm_suspended(wc->ti))) {
1608                         queue_work(wc->writeback_wq, &wc->writeback_work);
1609                         break;
1610                 }
1611
1612                 if (unlikely(wc->writeback_all)) {
1613                         if (unlikely(!e)) {
1614                                 writecache_flush(wc);
1615                                 e = container_of(rb_first(&wc->tree), struct wc_entry, rb_node);
1616                         } else
1617                                 e = g;
1618                 } else
1619                         e = container_of(wc->lru.prev, struct wc_entry, lru);
1620                 BUG_ON(e->write_in_progress);
1621                 if (unlikely(!writecache_entry_is_committed(wc, e))) {
1622                         writecache_flush(wc);
1623                 }
1624                 node = rb_prev(&e->rb_node);
1625                 if (node) {
1626                         f = container_of(node, struct wc_entry, rb_node);
1627                         if (unlikely(read_original_sector(wc, f) ==
1628                                      read_original_sector(wc, e))) {
1629                                 BUG_ON(!f->write_in_progress);
1630                                 list_del(&e->lru);
1631                                 list_add(&e->lru, &skipped);
1632                                 cond_resched();
1633                                 continue;
1634                         }
1635                 }
1636                 wc->writeback_size++;
1637                 list_del(&e->lru);
1638                 list_add(&e->lru, &wbl.list);
1639                 wbl.size++;
1640                 e->write_in_progress = true;
1641                 e->wc_list_contiguous = 1;
1642
1643                 f = e;
1644
1645                 while (1) {
1646                         next_node = rb_next(&f->rb_node);
1647                         if (unlikely(!next_node))
1648                                 break;
1649                         g = container_of(next_node, struct wc_entry, rb_node);
1650                         if (unlikely(read_original_sector(wc, g) ==
1651                             read_original_sector(wc, f))) {
1652                                 f = g;
1653                                 continue;
1654                         }
1655                         if (read_original_sector(wc, g) !=
1656                             read_original_sector(wc, f) + (wc->block_size >> SECTOR_SHIFT))
1657                                 break;
1658                         if (unlikely(g->write_in_progress))
1659                                 break;
1660                         if (unlikely(!writecache_entry_is_committed(wc, g)))
1661                                 break;
1662
1663                         if (!WC_MODE_PMEM(wc)) {
1664                                 if (g != f + 1)
1665                                         break;
1666                         }
1667
1668                         n_walked++;
1669                         //if (unlikely(n_walked > WRITEBACK_LATENCY) && likely(!wc->writeback_all))
1670                         //      break;
1671
1672                         wc->writeback_size++;
1673                         list_del(&g->lru);
1674                         list_add(&g->lru, &wbl.list);
1675                         wbl.size++;
1676                         g->write_in_progress = true;
1677                         g->wc_list_contiguous = BIO_MAX_PAGES;
1678                         f = g;
1679                         e->wc_list_contiguous++;
1680                         if (unlikely(e->wc_list_contiguous == BIO_MAX_PAGES)) {
1681                                 if (unlikely(wc->writeback_all)) {
1682                                         next_node = rb_next(&f->rb_node);
1683                                         if (likely(next_node))
1684                                                 g = container_of(next_node, struct wc_entry, rb_node);
1685                                 }
1686                                 break;
1687                         }
1688                 }
1689                 cond_resched();
1690         }
1691
1692         if (!list_empty(&skipped)) {
1693                 list_splice_tail(&skipped, &wc->lru);
1694                 /*
1695                  * If we didn't do any progress, we must wait until some
1696                  * writeback finishes to avoid burning CPU in a loop
1697                  */
1698                 if (unlikely(!wbl.size))
1699                         writecache_wait_for_writeback(wc);
1700         }
1701
1702         wc_unlock(wc);
1703
1704         blk_start_plug(&plug);
1705
1706         if (WC_MODE_PMEM(wc))
1707                 __writecache_writeback_pmem(wc, &wbl);
1708         else
1709                 __writecache_writeback_ssd(wc, &wbl);
1710
1711         blk_finish_plug(&plug);
1712
1713         if (unlikely(wc->writeback_all)) {
1714                 wc_lock(wc);
1715                 while (writecache_wait_for_writeback(wc));
1716                 wc_unlock(wc);
1717         }
1718 }
1719
1720 static int calculate_memory_size(uint64_t device_size, unsigned block_size,
1721                                  size_t *n_blocks_p, size_t *n_metadata_blocks_p)
1722 {
1723         uint64_t n_blocks, offset;
1724         struct wc_entry e;
1725
1726         n_blocks = device_size;
1727         do_div(n_blocks, block_size + sizeof(struct wc_memory_entry));
1728
1729         while (1) {
1730                 if (!n_blocks)
1731                         return -ENOSPC;
1732                 /* Verify the following entries[n_blocks] won't overflow */
1733                 if (n_blocks >= ((size_t)-sizeof(struct wc_memory_superblock) /
1734                                  sizeof(struct wc_memory_entry)))
1735                         return -EFBIG;
1736                 offset = offsetof(struct wc_memory_superblock, entries[n_blocks]);
1737                 offset = (offset + block_size - 1) & ~(uint64_t)(block_size - 1);
1738                 if (offset + n_blocks * block_size <= device_size)
1739                         break;
1740                 n_blocks--;
1741         }
1742
1743         /* check if the bit field overflows */
1744         e.index = n_blocks;
1745         if (e.index != n_blocks)
1746                 return -EFBIG;
1747
1748         if (n_blocks_p)
1749                 *n_blocks_p = n_blocks;
1750         if (n_metadata_blocks_p)
1751                 *n_metadata_blocks_p = offset >> __ffs(block_size);
1752         return 0;
1753 }
1754
1755 static int init_memory(struct dm_writecache *wc)
1756 {
1757         size_t b;
1758         int r;
1759
1760         r = calculate_memory_size(wc->memory_map_size, wc->block_size, &wc->n_blocks, NULL);
1761         if (r)
1762                 return r;
1763
1764         r = writecache_alloc_entries(wc);
1765         if (r)
1766                 return r;
1767
1768         for (b = 0; b < ARRAY_SIZE(sb(wc)->padding); b++)
1769                 pmem_assign(sb(wc)->padding[b], cpu_to_le64(0));
1770         pmem_assign(sb(wc)->version, cpu_to_le32(MEMORY_SUPERBLOCK_VERSION));
1771         pmem_assign(sb(wc)->block_size, cpu_to_le32(wc->block_size));
1772         pmem_assign(sb(wc)->n_blocks, cpu_to_le64(wc->n_blocks));
1773         pmem_assign(sb(wc)->seq_count, cpu_to_le64(0));
1774
1775         for (b = 0; b < wc->n_blocks; b++) {
1776                 write_original_sector_seq_count(wc, &wc->entries[b], -1, -1);
1777                 cond_resched();
1778         }
1779
1780         writecache_flush_all_metadata(wc);
1781         writecache_commit_flushed(wc, false);
1782         pmem_assign(sb(wc)->magic, cpu_to_le32(MEMORY_SUPERBLOCK_MAGIC));
1783         writecache_flush_region(wc, &sb(wc)->magic, sizeof sb(wc)->magic);
1784         writecache_commit_flushed(wc, false);
1785
1786         return 0;
1787 }
1788
1789 static void writecache_dtr(struct dm_target *ti)
1790 {
1791         struct dm_writecache *wc = ti->private;
1792
1793         if (!wc)
1794                 return;
1795
1796         if (wc->endio_thread)
1797                 kthread_stop(wc->endio_thread);
1798
1799         if (wc->flush_thread)
1800                 kthread_stop(wc->flush_thread);
1801
1802         bioset_exit(&wc->bio_set);
1803
1804         mempool_exit(&wc->copy_pool);
1805
1806         if (wc->writeback_wq)
1807                 destroy_workqueue(wc->writeback_wq);
1808
1809         if (wc->dev)
1810                 dm_put_device(ti, wc->dev);
1811
1812         if (wc->ssd_dev)
1813                 dm_put_device(ti, wc->ssd_dev);
1814
1815         if (wc->entries)
1816                 vfree(wc->entries);
1817
1818         if (wc->memory_map) {
1819                 if (WC_MODE_PMEM(wc))
1820                         persistent_memory_release(wc);
1821                 else
1822                         vfree(wc->memory_map);
1823         }
1824
1825         if (wc->dm_kcopyd)
1826                 dm_kcopyd_client_destroy(wc->dm_kcopyd);
1827
1828         if (wc->dm_io)
1829                 dm_io_client_destroy(wc->dm_io);
1830
1831         if (wc->dirty_bitmap)
1832                 vfree(wc->dirty_bitmap);
1833
1834         kfree(wc);
1835 }
1836
1837 static int writecache_ctr(struct dm_target *ti, unsigned argc, char **argv)
1838 {
1839         struct dm_writecache *wc;
1840         struct dm_arg_set as;
1841         const char *string;
1842         unsigned opt_params;
1843         size_t offset, data_size;
1844         int i, r;
1845         char dummy;
1846         int high_wm_percent = HIGH_WATERMARK;
1847         int low_wm_percent = LOW_WATERMARK;
1848         uint64_t x;
1849         struct wc_memory_superblock s;
1850
1851         static struct dm_arg _args[] = {
1852                 {0, 10, "Invalid number of feature args"},
1853         };
1854
1855         as.argc = argc;
1856         as.argv = argv;
1857
1858         wc = kzalloc(sizeof(struct dm_writecache), GFP_KERNEL);
1859         if (!wc) {
1860                 ti->error = "Cannot allocate writecache structure";
1861                 r = -ENOMEM;
1862                 goto bad;
1863         }
1864         ti->private = wc;
1865         wc->ti = ti;
1866
1867         mutex_init(&wc->lock);
1868         writecache_poison_lists(wc);
1869         init_waitqueue_head(&wc->freelist_wait);
1870         timer_setup(&wc->autocommit_timer, writecache_autocommit_timer, 0);
1871
1872         for (i = 0; i < 2; i++) {
1873                 atomic_set(&wc->bio_in_progress[i], 0);
1874                 init_waitqueue_head(&wc->bio_in_progress_wait[i]);
1875         }
1876
1877         wc->dm_io = dm_io_client_create();
1878         if (IS_ERR(wc->dm_io)) {
1879                 r = PTR_ERR(wc->dm_io);
1880                 ti->error = "Unable to allocate dm-io client";
1881                 wc->dm_io = NULL;
1882                 goto bad;
1883         }
1884
1885         wc->writeback_wq = alloc_workqueue("writecache-writeback", WQ_MEM_RECLAIM, 1);
1886         if (!wc->writeback_wq) {
1887                 r = -ENOMEM;
1888                 ti->error = "Could not allocate writeback workqueue";
1889                 goto bad;
1890         }
1891         INIT_WORK(&wc->writeback_work, writecache_writeback);
1892         INIT_WORK(&wc->flush_work, writecache_flush_work);
1893
1894         raw_spin_lock_init(&wc->endio_list_lock);
1895         INIT_LIST_HEAD(&wc->endio_list);
1896         wc->endio_thread = kthread_create(writecache_endio_thread, wc, "writecache_endio");
1897         if (IS_ERR(wc->endio_thread)) {
1898                 r = PTR_ERR(wc->endio_thread);
1899                 wc->endio_thread = NULL;
1900                 ti->error = "Couldn't spawn endio thread";
1901                 goto bad;
1902         }
1903         wake_up_process(wc->endio_thread);
1904
1905         /*
1906          * Parse the mode (pmem or ssd)
1907          */
1908         string = dm_shift_arg(&as);
1909         if (!string)
1910                 goto bad_arguments;
1911
1912         if (!strcasecmp(string, "s")) {
1913                 wc->pmem_mode = false;
1914         } else if (!strcasecmp(string, "p")) {
1915 #ifdef DM_WRITECACHE_HAS_PMEM
1916                 wc->pmem_mode = true;
1917                 wc->writeback_fua = true;
1918 #else
1919                 /*
1920                  * If the architecture doesn't support persistent memory or
1921                  * the kernel doesn't support any DAX drivers, this driver can
1922                  * only be used in SSD-only mode.
1923                  */
1924                 r = -EOPNOTSUPP;
1925                 ti->error = "Persistent memory or DAX not supported on this system";
1926                 goto bad;
1927 #endif
1928         } else {
1929                 goto bad_arguments;
1930         }
1931
1932         if (WC_MODE_PMEM(wc)) {
1933                 r = bioset_init(&wc->bio_set, BIO_POOL_SIZE,
1934                                 offsetof(struct writeback_struct, bio),
1935                                 BIOSET_NEED_BVECS);
1936                 if (r) {
1937                         ti->error = "Could not allocate bio set";
1938                         goto bad;
1939                 }
1940         } else {
1941                 r = mempool_init_kmalloc_pool(&wc->copy_pool, 1, sizeof(struct copy_struct));
1942                 if (r) {
1943                         ti->error = "Could not allocate mempool";
1944                         goto bad;
1945                 }
1946         }
1947
1948         /*
1949          * Parse the origin data device
1950          */
1951         string = dm_shift_arg(&as);
1952         if (!string)
1953                 goto bad_arguments;
1954         r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->dev);
1955         if (r) {
1956                 ti->error = "Origin data device lookup failed";
1957                 goto bad;
1958         }
1959
1960         /*
1961          * Parse cache data device (be it pmem or ssd)
1962          */
1963         string = dm_shift_arg(&as);
1964         if (!string)
1965                 goto bad_arguments;
1966
1967         r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->ssd_dev);
1968         if (r) {
1969                 ti->error = "Cache data device lookup failed";
1970                 goto bad;
1971         }
1972         wc->memory_map_size = i_size_read(wc->ssd_dev->bdev->bd_inode);
1973
1974         /*
1975          * Parse the cache block size
1976          */
1977         string = dm_shift_arg(&as);
1978         if (!string)
1979                 goto bad_arguments;
1980         if (sscanf(string, "%u%c", &wc->block_size, &dummy) != 1 ||
1981             wc->block_size < 512 || wc->block_size > PAGE_SIZE ||
1982             (wc->block_size & (wc->block_size - 1))) {
1983                 r = -EINVAL;
1984                 ti->error = "Invalid block size";
1985                 goto bad;
1986         }
1987         wc->block_size_bits = __ffs(wc->block_size);
1988
1989         wc->max_writeback_jobs = MAX_WRITEBACK_JOBS;
1990         wc->autocommit_blocks = !WC_MODE_PMEM(wc) ? AUTOCOMMIT_BLOCKS_SSD : AUTOCOMMIT_BLOCKS_PMEM;
1991         wc->autocommit_jiffies = msecs_to_jiffies(AUTOCOMMIT_MSEC);
1992
1993         /*
1994          * Parse optional arguments
1995          */
1996         r = dm_read_arg_group(_args, &as, &opt_params, &ti->error);
1997         if (r)
1998                 goto bad;
1999
2000         while (opt_params) {
2001                 string = dm_shift_arg(&as), opt_params--;
2002                 if (!strcasecmp(string, "start_sector") && opt_params >= 1) {
2003                         unsigned long long start_sector;
2004                         string = dm_shift_arg(&as), opt_params--;
2005                         if (sscanf(string, "%llu%c", &start_sector, &dummy) != 1)
2006                                 goto invalid_optional;
2007                         wc->start_sector = start_sector;
2008                         if (wc->start_sector != start_sector ||
2009                             wc->start_sector >= wc->memory_map_size >> SECTOR_SHIFT)
2010                                 goto invalid_optional;
2011                 } else if (!strcasecmp(string, "high_watermark") && opt_params >= 1) {
2012                         string = dm_shift_arg(&as), opt_params--;
2013                         if (sscanf(string, "%d%c", &high_wm_percent, &dummy) != 1)
2014                                 goto invalid_optional;
2015                         if (high_wm_percent < 0 || high_wm_percent > 100)
2016                                 goto invalid_optional;
2017                         wc->high_wm_percent_set = true;
2018                 } else if (!strcasecmp(string, "low_watermark") && opt_params >= 1) {
2019                         string = dm_shift_arg(&as), opt_params--;
2020                         if (sscanf(string, "%d%c", &low_wm_percent, &dummy) != 1)
2021                                 goto invalid_optional;
2022                         if (low_wm_percent < 0 || low_wm_percent > 100)
2023                                 goto invalid_optional;
2024                         wc->low_wm_percent_set = true;
2025                 } else if (!strcasecmp(string, "writeback_jobs") && opt_params >= 1) {
2026                         string = dm_shift_arg(&as), opt_params--;
2027                         if (sscanf(string, "%u%c", &wc->max_writeback_jobs, &dummy) != 1)
2028                                 goto invalid_optional;
2029                         wc->max_writeback_jobs_set = true;
2030                 } else if (!strcasecmp(string, "autocommit_blocks") && opt_params >= 1) {
2031                         string = dm_shift_arg(&as), opt_params--;
2032                         if (sscanf(string, "%u%c", &wc->autocommit_blocks, &dummy) != 1)
2033                                 goto invalid_optional;
2034                         wc->autocommit_blocks_set = true;
2035                 } else if (!strcasecmp(string, "autocommit_time") && opt_params >= 1) {
2036                         unsigned autocommit_msecs;
2037                         string = dm_shift_arg(&as), opt_params--;
2038                         if (sscanf(string, "%u%c", &autocommit_msecs, &dummy) != 1)
2039                                 goto invalid_optional;
2040                         if (autocommit_msecs > 3600000)
2041                                 goto invalid_optional;
2042                         wc->autocommit_jiffies = msecs_to_jiffies(autocommit_msecs);
2043                         wc->autocommit_time_set = true;
2044                 } else if (!strcasecmp(string, "fua")) {
2045                         if (WC_MODE_PMEM(wc)) {
2046                                 wc->writeback_fua = true;
2047                                 wc->writeback_fua_set = true;
2048                         } else goto invalid_optional;
2049                 } else if (!strcasecmp(string, "nofua")) {
2050                         if (WC_MODE_PMEM(wc)) {
2051                                 wc->writeback_fua = false;
2052                                 wc->writeback_fua_set = true;
2053                         } else goto invalid_optional;
2054                 } else {
2055 invalid_optional:
2056                         r = -EINVAL;
2057                         ti->error = "Invalid optional argument";
2058                         goto bad;
2059                 }
2060         }
2061
2062         if (high_wm_percent < low_wm_percent) {
2063                 r = -EINVAL;
2064                 ti->error = "High watermark must be greater than or equal to low watermark";
2065                 goto bad;
2066         }
2067
2068         if (WC_MODE_PMEM(wc)) {
2069                 r = persistent_memory_claim(wc);
2070                 if (r) {
2071                         ti->error = "Unable to map persistent memory for cache";
2072                         goto bad;
2073                 }
2074         } else {
2075                 struct dm_io_region region;
2076                 struct dm_io_request req;
2077                 size_t n_blocks, n_metadata_blocks;
2078                 uint64_t n_bitmap_bits;
2079
2080                 wc->memory_map_size -= (uint64_t)wc->start_sector << SECTOR_SHIFT;
2081
2082                 bio_list_init(&wc->flush_list);
2083                 wc->flush_thread = kthread_create(writecache_flush_thread, wc, "dm_writecache_flush");
2084                 if (IS_ERR(wc->flush_thread)) {
2085                         r = PTR_ERR(wc->flush_thread);
2086                         wc->flush_thread = NULL;
2087                         ti->error = "Couldn't spawn flush thread";
2088                         goto bad;
2089                 }
2090                 wake_up_process(wc->flush_thread);
2091
2092                 r = calculate_memory_size(wc->memory_map_size, wc->block_size,
2093                                           &n_blocks, &n_metadata_blocks);
2094                 if (r) {
2095                         ti->error = "Invalid device size";
2096                         goto bad;
2097                 }
2098
2099                 n_bitmap_bits = (((uint64_t)n_metadata_blocks << wc->block_size_bits) +
2100                                  BITMAP_GRANULARITY - 1) / BITMAP_GRANULARITY;
2101                 /* this is limitation of test_bit functions */
2102                 if (n_bitmap_bits > 1U << 31) {
2103                         r = -EFBIG;
2104                         ti->error = "Invalid device size";
2105                         goto bad;
2106                 }
2107
2108                 wc->memory_map = vmalloc(n_metadata_blocks << wc->block_size_bits);
2109                 if (!wc->memory_map) {
2110                         r = -ENOMEM;
2111                         ti->error = "Unable to allocate memory for metadata";
2112                         goto bad;
2113                 }
2114
2115                 wc->dm_kcopyd = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2116                 if (IS_ERR(wc->dm_kcopyd)) {
2117                         r = PTR_ERR(wc->dm_kcopyd);
2118                         ti->error = "Unable to allocate dm-kcopyd client";
2119                         wc->dm_kcopyd = NULL;
2120                         goto bad;
2121                 }
2122
2123                 wc->metadata_sectors = n_metadata_blocks << (wc->block_size_bits - SECTOR_SHIFT);
2124                 wc->dirty_bitmap_size = (n_bitmap_bits + BITS_PER_LONG - 1) /
2125                         BITS_PER_LONG * sizeof(unsigned long);
2126                 wc->dirty_bitmap = vzalloc(wc->dirty_bitmap_size);
2127                 if (!wc->dirty_bitmap) {
2128                         r = -ENOMEM;
2129                         ti->error = "Unable to allocate dirty bitmap";
2130                         goto bad;
2131                 }
2132
2133                 region.bdev = wc->ssd_dev->bdev;
2134                 region.sector = wc->start_sector;
2135                 region.count = wc->metadata_sectors;
2136                 req.bi_op = REQ_OP_READ;
2137                 req.bi_op_flags = REQ_SYNC;
2138                 req.mem.type = DM_IO_VMA;
2139                 req.mem.ptr.vma = (char *)wc->memory_map;
2140                 req.client = wc->dm_io;
2141                 req.notify.fn = NULL;
2142
2143                 r = dm_io(&req, 1, &region, NULL);
2144                 if (r) {
2145                         ti->error = "Unable to read metadata";
2146                         goto bad;
2147                 }
2148         }
2149
2150         r = memcpy_mcsafe(&s, sb(wc), sizeof(struct wc_memory_superblock));
2151         if (r) {
2152                 ti->error = "Hardware memory error when reading superblock";
2153                 goto bad;
2154         }
2155         if (!le32_to_cpu(s.magic) && !le32_to_cpu(s.version)) {
2156                 r = init_memory(wc);
2157                 if (r) {
2158                         ti->error = "Unable to initialize device";
2159                         goto bad;
2160                 }
2161                 r = memcpy_mcsafe(&s, sb(wc), sizeof(struct wc_memory_superblock));
2162                 if (r) {
2163                         ti->error = "Hardware memory error when reading superblock";
2164                         goto bad;
2165                 }
2166         }
2167
2168         if (le32_to_cpu(s.magic) != MEMORY_SUPERBLOCK_MAGIC) {
2169                 ti->error = "Invalid magic in the superblock";
2170                 r = -EINVAL;
2171                 goto bad;
2172         }
2173
2174         if (le32_to_cpu(s.version) != MEMORY_SUPERBLOCK_VERSION) {
2175                 ti->error = "Invalid version in the superblock";
2176                 r = -EINVAL;
2177                 goto bad;
2178         }
2179
2180         if (le32_to_cpu(s.block_size) != wc->block_size) {
2181                 ti->error = "Block size does not match superblock";
2182                 r = -EINVAL;
2183                 goto bad;
2184         }
2185
2186         wc->n_blocks = le64_to_cpu(s.n_blocks);
2187
2188         offset = wc->n_blocks * sizeof(struct wc_memory_entry);
2189         if (offset / sizeof(struct wc_memory_entry) != le64_to_cpu(sb(wc)->n_blocks)) {
2190 overflow:
2191                 ti->error = "Overflow in size calculation";
2192                 r = -EINVAL;
2193                 goto bad;
2194         }
2195         offset += sizeof(struct wc_memory_superblock);
2196         if (offset < sizeof(struct wc_memory_superblock))
2197                 goto overflow;
2198         offset = (offset + wc->block_size - 1) & ~(size_t)(wc->block_size - 1);
2199         data_size = wc->n_blocks * (size_t)wc->block_size;
2200         if (!offset || (data_size / wc->block_size != wc->n_blocks) ||
2201             (offset + data_size < offset))
2202                 goto overflow;
2203         if (offset + data_size > wc->memory_map_size) {
2204                 ti->error = "Memory area is too small";
2205                 r = -EINVAL;
2206                 goto bad;
2207         }
2208
2209         wc->metadata_sectors = offset >> SECTOR_SHIFT;
2210         wc->block_start = (char *)sb(wc) + offset;
2211
2212         x = (uint64_t)wc->n_blocks * (100 - high_wm_percent);
2213         x += 50;
2214         do_div(x, 100);
2215         wc->freelist_high_watermark = x;
2216         x = (uint64_t)wc->n_blocks * (100 - low_wm_percent);
2217         x += 50;
2218         do_div(x, 100);
2219         wc->freelist_low_watermark = x;
2220
2221         r = writecache_alloc_entries(wc);
2222         if (r) {
2223                 ti->error = "Cannot allocate memory";
2224                 goto bad;
2225         }
2226
2227         ti->num_flush_bios = 1;
2228         ti->flush_supported = true;
2229         ti->num_discard_bios = 1;
2230
2231         if (WC_MODE_PMEM(wc))
2232                 persistent_memory_flush_cache(wc->memory_map, wc->memory_map_size);
2233
2234         return 0;
2235
2236 bad_arguments:
2237         r = -EINVAL;
2238         ti->error = "Bad arguments";
2239 bad:
2240         writecache_dtr(ti);
2241         return r;
2242 }
2243
2244 static void writecache_status(struct dm_target *ti, status_type_t type,
2245                               unsigned status_flags, char *result, unsigned maxlen)
2246 {
2247         struct dm_writecache *wc = ti->private;
2248         unsigned extra_args;
2249         unsigned sz = 0;
2250         uint64_t x;
2251
2252         switch (type) {
2253         case STATUSTYPE_INFO:
2254                 DMEMIT("%ld %llu %llu %llu", writecache_has_error(wc),
2255                        (unsigned long long)wc->n_blocks, (unsigned long long)wc->freelist_size,
2256                        (unsigned long long)wc->writeback_size);
2257                 break;
2258         case STATUSTYPE_TABLE:
2259                 DMEMIT("%c %s %s %u ", WC_MODE_PMEM(wc) ? 'p' : 's',
2260                                 wc->dev->name, wc->ssd_dev->name, wc->block_size);
2261                 extra_args = 0;
2262                 if (wc->start_sector)
2263                         extra_args += 2;
2264                 if (wc->high_wm_percent_set)
2265                         extra_args += 2;
2266                 if (wc->low_wm_percent_set)
2267                         extra_args += 2;
2268                 if (wc->max_writeback_jobs_set)
2269                         extra_args += 2;
2270                 if (wc->autocommit_blocks_set)
2271                         extra_args += 2;
2272                 if (wc->autocommit_time_set)
2273                         extra_args += 2;
2274                 if (wc->writeback_fua_set)
2275                         extra_args++;
2276
2277                 DMEMIT("%u", extra_args);
2278                 if (wc->start_sector)
2279                         DMEMIT(" start_sector %llu", (unsigned long long)wc->start_sector);
2280                 if (wc->high_wm_percent_set) {
2281                         x = (uint64_t)wc->freelist_high_watermark * 100;
2282                         x += wc->n_blocks / 2;
2283                         do_div(x, (size_t)wc->n_blocks);
2284                         DMEMIT(" high_watermark %u", 100 - (unsigned)x);
2285                 }
2286                 if (wc->low_wm_percent_set) {
2287                         x = (uint64_t)wc->freelist_low_watermark * 100;
2288                         x += wc->n_blocks / 2;
2289                         do_div(x, (size_t)wc->n_blocks);
2290                         DMEMIT(" low_watermark %u", 100 - (unsigned)x);
2291                 }
2292                 if (wc->max_writeback_jobs_set)
2293                         DMEMIT(" writeback_jobs %u", wc->max_writeback_jobs);
2294                 if (wc->autocommit_blocks_set)
2295                         DMEMIT(" autocommit_blocks %u", wc->autocommit_blocks);
2296                 if (wc->autocommit_time_set)
2297                         DMEMIT(" autocommit_time %u", jiffies_to_msecs(wc->autocommit_jiffies));
2298                 if (wc->writeback_fua_set)
2299                         DMEMIT(" %sfua", wc->writeback_fua ? "" : "no");
2300                 break;
2301         }
2302 }
2303
2304 static struct target_type writecache_target = {
2305         .name                   = "writecache",
2306         .version                = {1, 1, 1},
2307         .module                 = THIS_MODULE,
2308         .ctr                    = writecache_ctr,
2309         .dtr                    = writecache_dtr,
2310         .status                 = writecache_status,
2311         .postsuspend            = writecache_suspend,
2312         .resume                 = writecache_resume,
2313         .message                = writecache_message,
2314         .map                    = writecache_map,
2315         .end_io                 = writecache_end_io,
2316         .iterate_devices        = writecache_iterate_devices,
2317         .io_hints               = writecache_io_hints,
2318 };
2319
2320 static int __init dm_writecache_init(void)
2321 {
2322         int r;
2323
2324         r = dm_register_target(&writecache_target);
2325         if (r < 0) {
2326                 DMERR("register failed %d", r);
2327                 return r;
2328         }
2329
2330         return 0;
2331 }
2332
2333 static void __exit dm_writecache_exit(void)
2334 {
2335         dm_unregister_target(&writecache_target);
2336 }
2337
2338 module_init(dm_writecache_init);
2339 module_exit(dm_writecache_exit);
2340
2341 MODULE_DESCRIPTION(DM_NAME " writecache target");
2342 MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
2343 MODULE_LICENSE("GPL");