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              min(0x10000000 / PAGE_SIZE, totalram_pages / 16)
  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         sector_t data_device_sectors;
 146         void *block_start;
 147         struct wc_entry *entries;
 148         unsigned block_size;
 149         unsigned char block_size_bits;
 150
 151         bool pmem_mode:1;
 152         bool writeback_fua:1;
 153
 154         bool overwrote_committed:1;
 155         bool memory_vmapped:1;
 156
 157         bool start_sector_set:1;
 158         bool high_wm_percent_set:1;
 159         bool low_wm_percent_set:1;
 160         bool max_writeback_jobs_set:1;
 161         bool autocommit_blocks_set:1;
 162         bool autocommit_time_set:1;
 163         bool writeback_fua_set:1;
 164         bool flush_on_suspend:1;
 165
 166         unsigned high_wm_percent_value;
 167         unsigned low_wm_percent_value;
 168         unsigned autocommit_time_value;
 169
 170         unsigned writeback_all;
 171         struct workqueue_struct *writeback_wq;
 172         struct work_struct writeback_work;
 173         struct work_struct flush_work;
 174
 175         struct dm_io_client *dm_io;
 176
 177         raw_spinlock_t endio_list_lock;
 178         struct list_head endio_list;
 179         struct task_struct *endio_thread;
 180
 181         struct task_struct *flush_thread;
 182         struct bio_list flush_list;
 183
 184         struct dm_kcopyd_client *dm_kcopyd;
 185         unsigned long *dirty_bitmap;
 186         unsigned dirty_bitmap_size;
 187
 188         struct bio_set bio_set;
 189         mempool_t copy_pool;
 190 };
 191
 192 #define WB_LIST_INLINE          16
 193
 194 struct writeback_struct {
 195         struct list_head endio_entry;
 196         struct dm_writecache *wc;
 197         struct wc_entry **wc_list;
 198         unsigned wc_list_n;
 199         unsigned page_offset;
 200         struct page *page;
 201         struct wc_entry *wc_list_inline[WB_LIST_INLINE];
 202         struct bio bio;
 203 };
 204
 205 struct copy_struct {
 206         struct list_head endio_entry;
 207         struct dm_writecache *wc;
 208         struct wc_entry *e;
 209         unsigned n_entries;
 210         int error;
 211 };
 212
 213 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(dm_writecache_throttle,
 214                                             "A percentage of time allocated for data copying");
 215
 216 static void wc_lock(struct dm_writecache *wc)
 217 {
 218         mutex_lock(&wc->lock);
 219 }
 220
 221 static void wc_unlock(struct dm_writecache *wc)
 222 {
 223         mutex_unlock(&wc->lock);
 224 }
 225
 226 #ifdef DM_WRITECACHE_HAS_PMEM
 227 static int persistent_memory_claim(struct dm_writecache *wc)
 228 {
 229         int r;
 230         loff_t s;
 231         long p, da;
 232         pfn_t pfn;
 233         int id;
 234         struct page **pages;
 235         sector_t offset;
 236
 237         wc->memory_vmapped = false;
 238
 239         if (!wc->ssd_dev->dax_dev) {
 240                 r = -EOPNOTSUPP;
 241                 goto err1;
 242         }
 243         s = wc->memory_map_size;
 244         p = s >> PAGE_SHIFT;
 245         if (!p) {
 246                 r = -EINVAL;
 247                 goto err1;
 248         }
 249         if (p != s >> PAGE_SHIFT) {
 250                 r = -EOVERFLOW;
 251                 goto err1;
 252         }
 253
 254         offset = get_start_sect(wc->ssd_dev->bdev);
 255         if (offset & (PAGE_SIZE / 512 - 1)) {
 256                 r = -EINVAL;
 257                 goto err1;
 258         }
 259         offset >>= PAGE_SHIFT - 9;
 260
 261         id = dax_read_lock();
 262
 263         da = dax_direct_access(wc->ssd_dev->dax_dev, offset, p, &wc->memory_map, &pfn);
 264         if (da < 0) {
 265                 wc->memory_map = NULL;
 266                 r = da;
 267                 goto err2;
 268         }
 269         if (!pfn_t_has_page(pfn)) {
 270                 wc->memory_map = NULL;
 271                 r = -EOPNOTSUPP;
 272                 goto err2;
 273         }
 274         if (da != p) {
 275                 long i;
 276                 wc->memory_map = NULL;
 277                 pages = kvmalloc_array(p, sizeof(struct page *), GFP_KERNEL);
 278                 if (!pages) {
 279                         r = -ENOMEM;
 280                         goto err2;
 281                 }
 282                 i = 0;
 283                 do {
 284                         long daa;
 285                         daa = dax_direct_access(wc->ssd_dev->dax_dev, offset + i, p - i,
 286                                                 NULL, &pfn);
 287                         if (daa <= 0) {
 288                                 r = daa ? daa : -EINVAL;
 289                                 goto err3;
 290                         }
 291                         if (!pfn_t_has_page(pfn)) {
 292                                 r = -EOPNOTSUPP;
 293                                 goto err3;
 294                         }
 295                         while (daa-- && i < p) {
 296                                 pages[i++] = pfn_t_to_page(pfn);
 297                                 pfn.val++;
 298                                 if (!(i & 15))
 299                                         cond_resched();
 300                         }
 301                 } while (i < p);
 302                 wc->memory_map = vmap(pages, p, VM_MAP, PAGE_KERNEL);
 303                 if (!wc->memory_map) {
 304                         r = -ENOMEM;
 305                         goto err3;
 306                 }
 307                 kvfree(pages);
 308                 wc->memory_vmapped = true;
 309         }
 310
 311         dax_read_unlock(id);
 312
 313         wc->memory_map += (size_t)wc->start_sector << SECTOR_SHIFT;
 314         wc->memory_map_size -= (size_t)wc->start_sector << SECTOR_SHIFT;
 315
 316         return 0;
 317 err3:
 318         kvfree(pages);
 319 err2:
 320         dax_read_unlock(id);
 321 err1:
 322         return r;
 323 }
 324 #else
 325 static int persistent_memory_claim(struct dm_writecache *wc)
 326 {
 327         return -EOPNOTSUPP;
 328 }
 329 #endif
 330
 331 static void persistent_memory_release(struct dm_writecache *wc)
 332 {
 333         if (wc->memory_vmapped)
 334                 vunmap(wc->memory_map - ((size_t)wc->start_sector << SECTOR_SHIFT));
 335 }
 336
 337 static struct page *persistent_memory_page(void *addr)
 338 {
 339         if (is_vmalloc_addr(addr))
 340                 return vmalloc_to_page(addr);
 341         else
 342                 return virt_to_page(addr);
 343 }
 344
 345 static unsigned persistent_memory_page_offset(void *addr)
 346 {
 347         return (unsigned long)addr & (PAGE_SIZE - 1);
 348 }
 349
 350 static void persistent_memory_flush_cache(void *ptr, size_t size)
 351 {
 352         if (is_vmalloc_addr(ptr))
 353                 flush_kernel_vmap_range(ptr, size);
 354 }
 355
 356 static void persistent_memory_invalidate_cache(void *ptr, size_t size)
 357 {
 358         if (is_vmalloc_addr(ptr))
 359                 invalidate_kernel_vmap_range(ptr, size);
 360 }
 361
 362 static struct wc_memory_superblock *sb(struct dm_writecache *wc)
 363 {
 364         return wc->memory_map;
 365 }
 366
 367 static struct wc_memory_entry *memory_entry(struct dm_writecache *wc, struct wc_entry *e)
 368 {
 369         if (is_power_of_2(sizeof(struct wc_entry)) && 0)
 370                 return &sb(wc)->entries[e - wc->entries];
 371         else
 372                 return &sb(wc)->entries[e->index];
 373 }
 374
 375 static void *memory_data(struct dm_writecache *wc, struct wc_entry *e)
 376 {
 377         return (char *)wc->block_start + (e->index << wc->block_size_bits);
 378 }
 379
 380 static sector_t cache_sector(struct dm_writecache *wc, struct wc_entry *e)
 381 {
 382         return wc->start_sector + wc->metadata_sectors +
 383                 ((sector_t)e->index << (wc->block_size_bits - SECTOR_SHIFT));
 384 }
 385
 386 static uint64_t read_original_sector(struct dm_writecache *wc, struct wc_entry *e)
 387 {
 388 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 389         return e->original_sector;
 390 #else
 391         return le64_to_cpu(memory_entry(wc, e)->original_sector);
 392 #endif
 393 }
 394
 395 static uint64_t read_seq_count(struct dm_writecache *wc, struct wc_entry *e)
 396 {
 397 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 398         return e->seq_count;
 399 #else
 400         return le64_to_cpu(memory_entry(wc, e)->seq_count);
 401 #endif
 402 }
 403
 404 static void clear_seq_count(struct dm_writecache *wc, struct wc_entry *e)
 405 {
 406 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 407         e->seq_count = -1;
 408 #endif
 409         pmem_assign(memory_entry(wc, e)->seq_count, cpu_to_le64(-1));
 410 }
 411
 412 static void write_original_sector_seq_count(struct dm_writecache *wc, struct wc_entry *e,
 413                                             uint64_t original_sector, uint64_t seq_count)
 414 {
 415         struct wc_memory_entry me;
 416 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 417         e->original_sector = original_sector;
 418         e->seq_count = seq_count;
 419 #endif
 420         me.original_sector = cpu_to_le64(original_sector);
 421         me.seq_count = cpu_to_le64(seq_count);
 422         pmem_assign(*memory_entry(wc, e), me);
 423 }
 424
 425 #define writecache_error(wc, err, msg, arg...)                          \
 426 do {                                                                    \
 427         if (!cmpxchg(&(wc)->error, 0, err))                             \
 428                 DMERR(msg, ##arg);                                      \
 429         wake_up(&(wc)->freelist_wait);                                  \
 430 } while (0)
 431
 432 #define writecache_has_error(wc)        (unlikely(READ_ONCE((wc)->error)))
 433
 434 static void writecache_flush_all_metadata(struct dm_writecache *wc)
 435 {
 436         if (!WC_MODE_PMEM(wc))
 437                 memset(wc->dirty_bitmap, -1, wc->dirty_bitmap_size);
 438 }
 439
 440 static void writecache_flush_region(struct dm_writecache *wc, void *ptr, size_t size)
 441 {
 442         if (!WC_MODE_PMEM(wc))
 443                 __set_bit(((char *)ptr - (char *)wc->memory_map) / BITMAP_GRANULARITY,
 444                           wc->dirty_bitmap);
 445 }
 446
 447 static void writecache_disk_flush(struct dm_writecache *wc, struct dm_dev *dev);
 448
 449 struct io_notify {
 450         struct dm_writecache *wc;
 451         struct completion c;
 452         atomic_t count;
 453 };
 454
 455 static void writecache_notify_io(unsigned long error, void *context)
 456 {
 457         struct io_notify *endio = context;
 458
 459         if (unlikely(error != 0))
 460                 writecache_error(endio->wc, -EIO, "error writing metadata");
 461         BUG_ON(atomic_read(&endio->count) <= 0);
 462         if (atomic_dec_and_test(&endio->count))
 463                 complete(&endio->c);
 464 }
 465
 466 static void writecache_wait_for_ios(struct dm_writecache *wc, int direction)
 467 {
 468         wait_event(wc->bio_in_progress_wait[direction],
 469                    !atomic_read(&wc->bio_in_progress[direction]));
 470 }
 471
 472 static void ssd_commit_flushed(struct dm_writecache *wc, bool wait_for_ios)
 473 {
 474         struct dm_io_region region;
 475         struct dm_io_request req;
 476         struct io_notify endio = {
 477                 wc,
 478                 COMPLETION_INITIALIZER_ONSTACK(endio.c),
 479                 ATOMIC_INIT(1),
 480         };
 481         unsigned bitmap_bits = wc->dirty_bitmap_size * 8;
 482         unsigned i = 0;
 483
 484         while (1) {
 485                 unsigned j;
 486                 i = find_next_bit(wc->dirty_bitmap, bitmap_bits, i);
 487                 if (unlikely(i == bitmap_bits))
 488                         break;
 489                 j = find_next_zero_bit(wc->dirty_bitmap, bitmap_bits, i);
 490
 491                 region.bdev = wc->ssd_dev->bdev;
 492                 region.sector = (sector_t)i * (BITMAP_GRANULARITY >> SECTOR_SHIFT);
 493                 region.count = (sector_t)(j - i) * (BITMAP_GRANULARITY >> SECTOR_SHIFT);
 494
 495                 if (unlikely(region.sector >= wc->metadata_sectors))
 496                         break;
 497                 if (unlikely(region.sector + region.count > wc->metadata_sectors))
 498                         region.count = wc->metadata_sectors - region.sector;
 499
 500                 region.sector += wc->start_sector;
 501                 atomic_inc(&endio.count);
 502                 req.bi_op = REQ_OP_WRITE;
 503                 req.bi_op_flags = REQ_SYNC;
 504                 req.mem.type = DM_IO_VMA;
 505                 req.mem.ptr.vma = (char *)wc->memory_map + (size_t)i * BITMAP_GRANULARITY;
 506                 req.client = wc->dm_io;
 507                 req.notify.fn = writecache_notify_io;
 508                 req.notify.context = &endio;
 509
 510                 /* writing via async dm-io (implied by notify.fn above) won't return an error */
 511                 (void) dm_io(&req, 1, &region, NULL);
 512                 i = j;
 513         }
 514
 515         writecache_notify_io(0, &endio);
 516         wait_for_completion_io(&endio.c);
 517
 518         if (wait_for_ios)
 519                 writecache_wait_for_ios(wc, WRITE);
 520
 521         writecache_disk_flush(wc, wc->ssd_dev);
 522
 523         memset(wc->dirty_bitmap, 0, wc->dirty_bitmap_size);
 524 }
 525
 526 static void writecache_commit_flushed(struct dm_writecache *wc, bool wait_for_ios)
 527 {
 528         if (WC_MODE_PMEM(wc))
 529                 wmb();
 530         else
 531                 ssd_commit_flushed(wc, wait_for_ios);
 532 }
 533
 534 static void writecache_disk_flush(struct dm_writecache *wc, struct dm_dev *dev)
 535 {
 536         int r;
 537         struct dm_io_region region;
 538         struct dm_io_request req;
 539
 540         region.bdev = dev->bdev;
 541         region.sector = 0;
 542         region.count = 0;
 543         req.bi_op = REQ_OP_WRITE;
 544         req.bi_op_flags = REQ_PREFLUSH;
 545         req.mem.type = DM_IO_KMEM;
 546         req.mem.ptr.addr = NULL;
 547         req.client = wc->dm_io;
 548         req.notify.fn = NULL;
 549
 550         r = dm_io(&req, 1, &region, NULL);
 551         if (unlikely(r))
 552                 writecache_error(wc, r, "error flushing metadata: %d", r);
 553 }
 554
 555 #define WFE_RETURN_FOLLOWING    1
 556 #define WFE_LOWEST_SEQ          2
 557
 558 static struct wc_entry *writecache_find_entry(struct dm_writecache *wc,
 559                                               uint64_t block, int flags)
 560 {
 561         struct wc_entry *e;
 562         struct rb_node *node = wc->tree.rb_node;
 563
 564         if (unlikely(!node))
 565                 return NULL;
 566
 567         while (1) {
 568                 e = container_of(node, struct wc_entry, rb_node);
 569                 if (read_original_sector(wc, e) == block)
 570                         break;
 571                 node = (read_original_sector(wc, e) >= block ?
 572                         e->rb_node.rb_left : e->rb_node.rb_right);
 573                 if (unlikely(!node)) {
 574                         if (!(flags & WFE_RETURN_FOLLOWING)) {
 575                                 return NULL;
 576                         }
 577                         if (read_original_sector(wc, e) >= block) {
 578                                 break;
 579                         } else {
 580                                 node = rb_next(&e->rb_node);
 581                                 if (unlikely(!node)) {
 582                                         return NULL;
 583                                 }
 584                                 e = container_of(node, struct wc_entry, rb_node);
 585                                 break;
 586                         }
 587                 }
 588         }
 589
 590         while (1) {
 591                 struct wc_entry *e2;
 592                 if (flags & WFE_LOWEST_SEQ)
 593                         node = rb_prev(&e->rb_node);
 594                 else
 595                         node = rb_next(&e->rb_node);
 596                 if (!node)
 597                         return e;
 598                 e2 = container_of(node, struct wc_entry, rb_node);
 599                 if (read_original_sector(wc, e2) != block)
 600                         return e;
 601                 e = e2;
 602         }
 603 }
 604
 605 static void writecache_insert_entry(struct dm_writecache *wc, struct wc_entry *ins)
 606 {
 607         struct wc_entry *e;
 608         struct rb_node **node = &wc->tree.rb_node, *parent = NULL;
 609
 610         while (*node) {
 611                 e = container_of(*node, struct wc_entry, rb_node);
 612                 parent = &e->rb_node;
 613                 if (read_original_sector(wc, e) > read_original_sector(wc, ins))
 614                         node = &parent->rb_left;
 615                 else
 616                         node = &parent->rb_right;
 617         }
 618         rb_link_node(&ins->rb_node, parent, node);
 619         rb_insert_color(&ins->rb_node, &wc->tree);
 620         list_add(&ins->lru, &wc->lru);
 621 }
 622
 623 static void writecache_unlink(struct dm_writecache *wc, struct wc_entry *e)
 624 {
 625         list_del(&e->lru);
 626         rb_erase(&e->rb_node, &wc->tree);
 627 }
 628
 629 static void writecache_add_to_freelist(struct dm_writecache *wc, struct wc_entry *e)
 630 {
 631         if (WC_MODE_SORT_FREELIST(wc)) {
 632                 struct rb_node **node = &wc->freetree.rb_node, *parent = NULL;
 633                 if (unlikely(!*node))
 634                         wc->current_free = e;
 635                 while (*node) {
 636                         parent = *node;
 637                         if (&e->rb_node < *node)
 638                                 node = &parent->rb_left;
 639                         else
 640                                 node = &parent->rb_right;
 641                 }
 642                 rb_link_node(&e->rb_node, parent, node);
 643                 rb_insert_color(&e->rb_node, &wc->freetree);
 644         } else {
 645                 list_add_tail(&e->lru, &wc->freelist);
 646         }
 647         wc->freelist_size++;
 648 }
 649
 650 static inline void writecache_verify_watermark(struct dm_writecache *wc)
 651 {
 652         if (unlikely(wc->freelist_size + wc->writeback_size <= wc->freelist_high_watermark))
 653                 queue_work(wc->writeback_wq, &wc->writeback_work);
 654 }
 655
 656 static struct wc_entry *writecache_pop_from_freelist(struct dm_writecache *wc)
 657 {
 658         struct wc_entry *e;
 659
 660         if (WC_MODE_SORT_FREELIST(wc)) {
 661                 struct rb_node *next;
 662                 if (unlikely(!wc->current_free))
 663                         return NULL;
 664                 e = wc->current_free;
 665                 next = rb_next(&e->rb_node);
 666                 rb_erase(&e->rb_node, &wc->freetree);
 667                 if (unlikely(!next))
 668                         next = rb_first(&wc->freetree);
 669                 wc->current_free = next ? container_of(next, struct wc_entry, rb_node) : NULL;
 670         } else {
 671                 if (unlikely(list_empty(&wc->freelist)))
 672                         return NULL;
 673                 e = container_of(wc->freelist.next, struct wc_entry, lru);
 674                 list_del(&e->lru);
 675         }
 676         wc->freelist_size--;
 677
 678         writecache_verify_watermark(wc);
 679
 680         return e;
 681 }
 682
 683 static void writecache_free_entry(struct dm_writecache *wc, struct wc_entry *e)
 684 {
 685         writecache_unlink(wc, e);
 686         writecache_add_to_freelist(wc, e);
 687         clear_seq_count(wc, e);
 688         writecache_flush_region(wc, memory_entry(wc, e), sizeof(struct wc_memory_entry));
 689         if (unlikely(waitqueue_active(&wc->freelist_wait)))
 690                 wake_up(&wc->freelist_wait);
 691 }
 692
 693 static void writecache_wait_on_freelist(struct dm_writecache *wc)
 694 {
 695         DEFINE_WAIT(wait);
 696
 697         prepare_to_wait(&wc->freelist_wait, &wait, TASK_UNINTERRUPTIBLE);
 698         wc_unlock(wc);
 699         io_schedule();
 700         finish_wait(&wc->freelist_wait, &wait);
 701         wc_lock(wc);
 702 }
 703
 704 static void writecache_poison_lists(struct dm_writecache *wc)
 705 {
 706         /*
 707          * Catch incorrect access to these values while the device is suspended.
 708          */
 709         memset(&wc->tree, -1, sizeof wc->tree);
 710         wc->lru.next = LIST_POISON1;
 711         wc->lru.prev = LIST_POISON2;
 712         wc->freelist.next = LIST_POISON1;
 713         wc->freelist.prev = LIST_POISON2;
 714 }
 715
 716 static void writecache_flush_entry(struct dm_writecache *wc, struct wc_entry *e)
 717 {
 718         writecache_flush_region(wc, memory_entry(wc, e), sizeof(struct wc_memory_entry));
 719         if (WC_MODE_PMEM(wc))
 720                 writecache_flush_region(wc, memory_data(wc, e), wc->block_size);
 721 }
 722
 723 static bool writecache_entry_is_committed(struct dm_writecache *wc, struct wc_entry *e)
 724 {
 725         return read_seq_count(wc, e) < wc->seq_count;
 726 }
 727
 728 static void writecache_flush(struct dm_writecache *wc)
 729 {
 730         struct wc_entry *e, *e2;
 731         bool need_flush_after_free;
 732
 733         wc->uncommitted_blocks = 0;
 734         del_timer(&wc->autocommit_timer);
 735
 736         if (list_empty(&wc->lru))
 737                 return;
 738
 739         e = container_of(wc->lru.next, struct wc_entry, lru);
 740         if (writecache_entry_is_committed(wc, e)) {
 741                 if (wc->overwrote_committed) {
 742                         writecache_wait_for_ios(wc, WRITE);
 743                         writecache_disk_flush(wc, wc->ssd_dev);
 744                         wc->overwrote_committed = false;
 745                 }
 746                 return;
 747         }
 748         while (1) {
 749                 writecache_flush_entry(wc, e);
 750                 if (unlikely(e->lru.next == &wc->lru))
 751                         break;
 752                 e2 = container_of(e->lru.next, struct wc_entry, lru);
 753                 if (writecache_entry_is_committed(wc, e2))
 754                         break;
 755                 e = e2;
 756                 cond_resched();
 757         }
 758         writecache_commit_flushed(wc, true);
 759
 760         wc->seq_count++;
 761         pmem_assign(sb(wc)->seq_count, cpu_to_le64(wc->seq_count));
 762         writecache_flush_region(wc, &sb(wc)->seq_count, sizeof sb(wc)->seq_count);
 763         writecache_commit_flushed(wc, false);
 764
 765         wc->overwrote_committed = false;
 766
 767         need_flush_after_free = false;
 768         while (1) {
 769                 /* Free another committed entry with lower seq-count */
 770                 struct rb_node *rb_node = rb_prev(&e->rb_node);
 771
 772                 if (rb_node) {
 773                         e2 = container_of(rb_node, struct wc_entry, rb_node);
 774                         if (read_original_sector(wc, e2) == read_original_sector(wc, e) &&
 775                             likely(!e2->write_in_progress)) {
 776                                 writecache_free_entry(wc, e2);
 777                                 need_flush_after_free = true;
 778                         }
 779                 }
 780                 if (unlikely(e->lru.prev == &wc->lru))
 781                         break;
 782                 e = container_of(e->lru.prev, struct wc_entry, lru);
 783                 cond_resched();
 784         }
 785
 786         if (need_flush_after_free)
 787                 writecache_commit_flushed(wc, false);
 788 }
 789
 790 static void writecache_flush_work(struct work_struct *work)
 791 {
 792         struct dm_writecache *wc = container_of(work, struct dm_writecache, flush_work);
 793
 794         wc_lock(wc);
 795         writecache_flush(wc);
 796         wc_unlock(wc);
 797 }
 798
 799 static void writecache_autocommit_timer(struct timer_list *t)
 800 {
 801         struct dm_writecache *wc = from_timer(wc, t, autocommit_timer);
 802         if (!writecache_has_error(wc))
 803                 queue_work(wc->writeback_wq, &wc->flush_work);
 804 }
 805
 806 static void writecache_schedule_autocommit(struct dm_writecache *wc)
 807 {
 808         if (!timer_pending(&wc->autocommit_timer))
 809                 mod_timer(&wc->autocommit_timer, jiffies + wc->autocommit_jiffies);
 810 }
 811
 812 static void writecache_discard(struct dm_writecache *wc, sector_t start, sector_t end)
 813 {
 814         struct wc_entry *e;
 815         bool discarded_something = false;
 816
 817         e = writecache_find_entry(wc, start, WFE_RETURN_FOLLOWING | WFE_LOWEST_SEQ);
 818         if (unlikely(!e))
 819                 return;
 820
 821         while (read_original_sector(wc, e) < end) {
 822                 struct rb_node *node = rb_next(&e->rb_node);
 823
 824                 if (likely(!e->write_in_progress)) {
 825                         if (!discarded_something) {
 826                                 writecache_wait_for_ios(wc, READ);
 827                                 writecache_wait_for_ios(wc, WRITE);
 828                                 discarded_something = true;
 829                         }
 830                         if (!writecache_entry_is_committed(wc, e))
 831                                 wc->uncommitted_blocks--;
 832                         writecache_free_entry(wc, e);
 833                 }
 834
 835                 if (!node)
 836                         break;
 837
 838                 e = container_of(node, struct wc_entry, rb_node);
 839         }
 840
 841         if (discarded_something)
 842                 writecache_commit_flushed(wc, false);
 843 }
 844
 845 static bool writecache_wait_for_writeback(struct dm_writecache *wc)
 846 {
 847         if (wc->writeback_size) {
 848                 writecache_wait_on_freelist(wc);
 849                 return true;
 850         }
 851         return false;
 852 }
 853
 854 static void writecache_suspend(struct dm_target *ti)
 855 {
 856         struct dm_writecache *wc = ti->private;
 857         bool flush_on_suspend;
 858
 859         del_timer_sync(&wc->autocommit_timer);
 860
 861         wc_lock(wc);
 862         writecache_flush(wc);
 863         flush_on_suspend = wc->flush_on_suspend;
 864         if (flush_on_suspend) {
 865                 wc->flush_on_suspend = false;
 866                 wc->writeback_all++;
 867                 queue_work(wc->writeback_wq, &wc->writeback_work);
 868         }
 869         wc_unlock(wc);
 870
 871         drain_workqueue(wc->writeback_wq);
 872
 873         wc_lock(wc);
 874         if (flush_on_suspend)
 875                 wc->writeback_all--;
 876         while (writecache_wait_for_writeback(wc));
 877
 878         if (WC_MODE_PMEM(wc))
 879                 persistent_memory_flush_cache(wc->memory_map, wc->memory_map_size);
 880
 881         writecache_poison_lists(wc);
 882
 883         wc_unlock(wc);
 884 }
 885
 886 static int writecache_alloc_entries(struct dm_writecache *wc)
 887 {
 888         size_t b;
 889
 890         if (wc->entries)
 891                 return 0;
 892         wc->entries = vmalloc(array_size(sizeof(struct wc_entry), wc->n_blocks));
 893         if (!wc->entries)
 894                 return -ENOMEM;
 895         for (b = 0; b < wc->n_blocks; b++) {
 896                 struct wc_entry *e = &wc->entries[b];
 897                 e->index = b;
 898                 e->write_in_progress = false;
 899                 cond_resched();
 900         }
 901
 902         return 0;
 903 }
 904
 905 static int writecache_read_metadata(struct dm_writecache *wc, sector_t n_sectors)
 906 {
 907         struct dm_io_region region;
 908         struct dm_io_request req;
 909
 910         region.bdev = wc->ssd_dev->bdev;
 911         region.sector = wc->start_sector;
 912         region.count = n_sectors;
 913         req.bi_op = REQ_OP_READ;
 914         req.bi_op_flags = REQ_SYNC;
 915         req.mem.type = DM_IO_VMA;
 916         req.mem.ptr.vma = (char *)wc->memory_map;
 917         req.client = wc->dm_io;
 918         req.notify.fn = NULL;
 919
 920         return dm_io(&req, 1, &region, NULL);
 921 }
 922
 923 static void writecache_resume(struct dm_target *ti)
 924 {
 925         struct dm_writecache *wc = ti->private;
 926         size_t b;
 927         bool need_flush = false;
 928         __le64 sb_seq_count;
 929         int r;
 930
 931         wc_lock(wc);
 932
 933         wc->data_device_sectors = i_size_read(wc->dev->bdev->bd_inode) >> SECTOR_SHIFT;
 934
 935         if (WC_MODE_PMEM(wc)) {
 936                 persistent_memory_invalidate_cache(wc->memory_map, wc->memory_map_size);
 937         } else {
 938                 r = writecache_read_metadata(wc, wc->metadata_sectors);
 939                 if (r) {
 940                         size_t sb_entries_offset;
 941                         writecache_error(wc, r, "unable to read metadata: %d", r);
 942                         sb_entries_offset = offsetof(struct wc_memory_superblock, entries);
 943                         memset((char *)wc->memory_map + sb_entries_offset, -1,
 944                                (wc->metadata_sectors << SECTOR_SHIFT) - sb_entries_offset);
 945                 }
 946         }
 947
 948         wc->tree = RB_ROOT;
 949         INIT_LIST_HEAD(&wc->lru);
 950         if (WC_MODE_SORT_FREELIST(wc)) {
 951                 wc->freetree = RB_ROOT;
 952                 wc->current_free = NULL;
 953         } else {
 954                 INIT_LIST_HEAD(&wc->freelist);
 955         }
 956         wc->freelist_size = 0;
 957
 958         r = memcpy_mcsafe(&sb_seq_count, &sb(wc)->seq_count, sizeof(uint64_t));
 959         if (r) {
 960                 writecache_error(wc, r, "hardware memory error when reading superblock: %d", r);
 961                 sb_seq_count = cpu_to_le64(0);
 962         }
 963         wc->seq_count = le64_to_cpu(sb_seq_count);
 964
 965 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 966         for (b = 0; b < wc->n_blocks; b++) {
 967                 struct wc_entry *e = &wc->entries[b];
 968                 struct wc_memory_entry wme;
 969                 if (writecache_has_error(wc)) {
 970                         e->original_sector = -1;
 971                         e->seq_count = -1;
 972                         continue;
 973                 }
 974                 r = memcpy_mcsafe(&wme, memory_entry(wc, e), sizeof(struct wc_memory_entry));
 975                 if (r) {
 976                         writecache_error(wc, r, "hardware memory error when reading metadata entry %lu: %d",
 977                                          (unsigned long)b, r);
 978                         e->original_sector = -1;
 979                         e->seq_count = -1;
 980                 } else {
 981                         e->original_sector = le64_to_cpu(wme.original_sector);
 982                         e->seq_count = le64_to_cpu(wme.seq_count);
 983                 }
 984                 cond_resched();
 985         }
 986 #endif
 987         for (b = 0; b < wc->n_blocks; b++) {
 988                 struct wc_entry *e = &wc->entries[b];
 989                 if (!writecache_entry_is_committed(wc, e)) {
 990                         if (read_seq_count(wc, e) != -1) {
 991 erase_this:
 992                                 clear_seq_count(wc, e);
 993                                 need_flush = true;
 994                         }
 995                         writecache_add_to_freelist(wc, e);
 996                 } else {
 997                         struct wc_entry *old;
 998
 999                         old = writecache_find_entry(wc, read_original_sector(wc, e), 0);
1000                         if (!old) {
1001                                 writecache_insert_entry(wc, e);
1002                         } else {
1003                                 if (read_seq_count(wc, old) == read_seq_count(wc, e)) {
1004                                         writecache_error(wc, -EINVAL,
1005                                                  "two identical entries, position %llu, sector %llu, sequence %llu",
1006                                                  (unsigned long long)b, (unsigned long long)read_original_sector(wc, e),
1007                                                  (unsigned long long)read_seq_count(wc, e));
1008                                 }
1009                                 if (read_seq_count(wc, old) > read_seq_count(wc, e)) {
1010                                         goto erase_this;
1011                                 } else {
1012                                         writecache_free_entry(wc, old);
1013                                         writecache_insert_entry(wc, e);
1014                                         need_flush = true;
1015                                 }
1016                         }
1017                 }
1018                 cond_resched();
1019         }
1020
1021         if (need_flush) {
1022                 writecache_flush_all_metadata(wc);
1023                 writecache_commit_flushed(wc, false);
1024         }
1025
1026         writecache_verify_watermark(wc);
1027
1028         wc_unlock(wc);
1029 }
1030
1031 static int process_flush_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1032 {
1033         if (argc != 1)
1034                 return -EINVAL;
1035
1036         wc_lock(wc);
1037         if (dm_suspended(wc->ti)) {
1038                 wc_unlock(wc);
1039                 return -EBUSY;
1040         }
1041         if (writecache_has_error(wc)) {
1042                 wc_unlock(wc);
1043                 return -EIO;
1044         }
1045
1046         writecache_flush(wc);
1047         wc->writeback_all++;
1048         queue_work(wc->writeback_wq, &wc->writeback_work);
1049         wc_unlock(wc);
1050
1051         flush_workqueue(wc->writeback_wq);
1052
1053         wc_lock(wc);
1054         wc->writeback_all--;
1055         if (writecache_has_error(wc)) {
1056                 wc_unlock(wc);
1057                 return -EIO;
1058         }
1059         wc_unlock(wc);
1060
1061         return 0;
1062 }
1063
1064 static int process_flush_on_suspend_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1065 {
1066         if (argc != 1)
1067                 return -EINVAL;
1068
1069         wc_lock(wc);
1070         wc->flush_on_suspend = true;
1071         wc_unlock(wc);
1072
1073         return 0;
1074 }
1075
1076 static int writecache_message(struct dm_target *ti, unsigned argc, char **argv,
1077                               char *result, unsigned maxlen)
1078 {
1079         int r = -EINVAL;
1080         struct dm_writecache *wc = ti->private;
1081
1082         if (!strcasecmp(argv[0], "flush"))
1083                 r = process_flush_mesg(argc, argv, wc);
1084         else if (!strcasecmp(argv[0], "flush_on_suspend"))
1085                 r = process_flush_on_suspend_mesg(argc, argv, wc);
1086         else
1087                 DMERR("unrecognised message received: %s", argv[0]);
1088
1089         return r;
1090 }
1091
1092 static void bio_copy_block(struct dm_writecache *wc, struct bio *bio, void *data)
1093 {
1094         void *buf;
1095         unsigned long flags;
1096         unsigned size;
1097         int rw = bio_data_dir(bio);
1098         unsigned remaining_size = wc->block_size;
1099
1100         do {
1101                 struct bio_vec bv = bio_iter_iovec(bio, bio->bi_iter);
1102                 buf = bvec_kmap_irq(&bv, &flags);
1103                 size = bv.bv_len;
1104                 if (unlikely(size > remaining_size))
1105                         size = remaining_size;
1106
1107                 if (rw == READ) {
1108                         int r;
1109                         r = memcpy_mcsafe(buf, data, size);
1110                         flush_dcache_page(bio_page(bio));
1111                         if (unlikely(r)) {
1112                                 writecache_error(wc, r, "hardware memory error when reading data: %d", r);
1113                                 bio->bi_status = BLK_STS_IOERR;
1114                         }
1115                 } else {
1116                         flush_dcache_page(bio_page(bio));
1117                         memcpy_flushcache(data, buf, size);
1118                 }
1119
1120                 bvec_kunmap_irq(buf, &flags);
1121
1122                 data = (char *)data + size;
1123                 remaining_size -= size;
1124                 bio_advance(bio, size);
1125         } while (unlikely(remaining_size));
1126 }
1127
1128 static int writecache_flush_thread(void *data)
1129 {
1130         struct dm_writecache *wc = data;
1131
1132         while (1) {
1133                 struct bio *bio;
1134
1135                 wc_lock(wc);
1136                 bio = bio_list_pop(&wc->flush_list);
1137                 if (!bio) {
1138                         set_current_state(TASK_INTERRUPTIBLE);
1139                         wc_unlock(wc);
1140
1141                         if (unlikely(kthread_should_stop())) {
1142                                 set_current_state(TASK_RUNNING);
1143                                 break;
1144                         }
1145
1146                         schedule();
1147                         continue;
1148                 }
1149
1150                 if (bio_op(bio) == REQ_OP_DISCARD) {
1151                         writecache_discard(wc, bio->bi_iter.bi_sector,
1152                                            bio_end_sector(bio));
1153                         wc_unlock(wc);
1154                         bio_set_dev(bio, wc->dev->bdev);
1155                         generic_make_request(bio);
1156                 } else {
1157                         writecache_flush(wc);
1158                         wc_unlock(wc);
1159                         if (writecache_has_error(wc))
1160                                 bio->bi_status = BLK_STS_IOERR;
1161                         bio_endio(bio);
1162                 }
1163         }
1164
1165         return 0;
1166 }
1167
1168 static void writecache_offload_bio(struct dm_writecache *wc, struct bio *bio)
1169 {
1170         if (bio_list_empty(&wc->flush_list))
1171                 wake_up_process(wc->flush_thread);
1172         bio_list_add(&wc->flush_list, bio);
1173 }
1174
1175 static int writecache_map(struct dm_target *ti, struct bio *bio)
1176 {
1177         struct wc_entry *e;
1178         struct dm_writecache *wc = ti->private;
1179
1180         bio->bi_private = NULL;
1181
1182         wc_lock(wc);
1183
1184         if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1185                 if (writecache_has_error(wc))
1186                         goto unlock_error;
1187                 if (WC_MODE_PMEM(wc)) {
1188                         writecache_flush(wc);
1189                         if (writecache_has_error(wc))
1190                                 goto unlock_error;
1191                         goto unlock_submit;
1192                 } else {
1193                         writecache_offload_bio(wc, bio);
1194                         goto unlock_return;
1195                 }
1196         }
1197
1198         bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
1199
1200         if (unlikely((((unsigned)bio->bi_iter.bi_sector | bio_sectors(bio)) &
1201                                 (wc->block_size / 512 - 1)) != 0)) {
1202                 DMERR("I/O is not aligned, sector %llu, size %u, block size %u",
1203                       (unsigned long long)bio->bi_iter.bi_sector,
1204                       bio->bi_iter.bi_size, wc->block_size);
1205                 goto unlock_error;
1206         }
1207
1208         if (unlikely(bio_op(bio) == REQ_OP_DISCARD)) {
1209                 if (writecache_has_error(wc))
1210                         goto unlock_error;
1211                 if (WC_MODE_PMEM(wc)) {
1212                         writecache_discard(wc, bio->bi_iter.bi_sector, bio_end_sector(bio));
1213                         goto unlock_remap_origin;
1214                 } else {
1215                         writecache_offload_bio(wc, bio);
1216                         goto unlock_return;
1217                 }
1218         }
1219
1220         if (bio_data_dir(bio) == READ) {
1221 read_next_block:
1222                 e = writecache_find_entry(wc, bio->bi_iter.bi_sector, WFE_RETURN_FOLLOWING);
1223                 if (e && read_original_sector(wc, e) == bio->bi_iter.bi_sector) {
1224                         if (WC_MODE_PMEM(wc)) {
1225                                 bio_copy_block(wc, bio, memory_data(wc, e));
1226                                 if (bio->bi_iter.bi_size)
1227                                         goto read_next_block;
1228                                 goto unlock_submit;
1229                         } else {
1230                                 dm_accept_partial_bio(bio, wc->block_size >> SECTOR_SHIFT);
1231                                 bio_set_dev(bio, wc->ssd_dev->bdev);
1232                                 bio->bi_iter.bi_sector = cache_sector(wc, e);
1233                                 if (!writecache_entry_is_committed(wc, e))
1234                                         writecache_wait_for_ios(wc, WRITE);
1235                                 goto unlock_remap;
1236                         }
1237                 } else {
1238                         if (e) {
1239                                 sector_t next_boundary =
1240                                         read_original_sector(wc, e) - bio->bi_iter.bi_sector;
1241                                 if (next_boundary < bio->bi_iter.bi_size >> SECTOR_SHIFT) {
1242                                         dm_accept_partial_bio(bio, next_boundary);
1243                                 }
1244                         }
1245                         goto unlock_remap_origin;
1246                 }
1247         } else {
1248                 do {
1249                         if (writecache_has_error(wc))
1250                                 goto unlock_error;
1251                         e = writecache_find_entry(wc, bio->bi_iter.bi_sector, 0);
1252                         if (e) {
1253                                 if (!writecache_entry_is_committed(wc, e))
1254                                         goto bio_copy;
1255                                 if (!WC_MODE_PMEM(wc) && !e->write_in_progress) {
1256                                         wc->overwrote_committed = true;
1257                                         goto bio_copy;
1258                                 }
1259                         }
1260                         e = writecache_pop_from_freelist(wc);
1261                         if (unlikely(!e)) {
1262                                 writecache_wait_on_freelist(wc);
1263                                 continue;
1264                         }
1265                         write_original_sector_seq_count(wc, e, bio->bi_iter.bi_sector, wc->seq_count);
1266                         writecache_insert_entry(wc, e);
1267                         wc->uncommitted_blocks++;
1268 bio_copy:
1269                         if (WC_MODE_PMEM(wc)) {
1270                                 bio_copy_block(wc, bio, memory_data(wc, e));
1271                         } else {
1272                                 dm_accept_partial_bio(bio, wc->block_size >> SECTOR_SHIFT);
1273                                 bio_set_dev(bio, wc->ssd_dev->bdev);
1274                                 bio->bi_iter.bi_sector = cache_sector(wc, e);
1275                                 if (unlikely(wc->uncommitted_blocks >= wc->autocommit_blocks)) {
1276                                         wc->uncommitted_blocks = 0;
1277                                         queue_work(wc->writeback_wq, &wc->flush_work);
1278                                 } else {
1279                                         writecache_schedule_autocommit(wc);
1280                                 }
1281                                 goto unlock_remap;
1282                         }
1283                 } while (bio->bi_iter.bi_size);
1284
1285                 if (unlikely(bio->bi_opf & REQ_FUA ||
1286                              wc->uncommitted_blocks >= wc->autocommit_blocks))
1287                         writecache_flush(wc);
1288                 else
1289                         writecache_schedule_autocommit(wc);
1290                 goto unlock_submit;
1291         }
1292
1293 unlock_remap_origin:
1294         bio_set_dev(bio, wc->dev->bdev);
1295         wc_unlock(wc);
1296         return DM_MAPIO_REMAPPED;
1297
1298 unlock_remap:
1299         /* make sure that writecache_end_io decrements bio_in_progress: */
1300         bio->bi_private = (void *)1;
1301         atomic_inc(&wc->bio_in_progress[bio_data_dir(bio)]);
1302         wc_unlock(wc);
1303         return DM_MAPIO_REMAPPED;
1304
1305 unlock_submit:
1306         wc_unlock(wc);
1307         bio_endio(bio);
1308         return DM_MAPIO_SUBMITTED;
1309
1310 unlock_return:
1311         wc_unlock(wc);
1312         return DM_MAPIO_SUBMITTED;
1313
1314 unlock_error:
1315         wc_unlock(wc);
1316         bio_io_error(bio);
1317         return DM_MAPIO_SUBMITTED;
1318 }
1319
1320 static int writecache_end_io(struct dm_target *ti, struct bio *bio, blk_status_t *status)
1321 {
1322         struct dm_writecache *wc = ti->private;
1323
1324         if (bio->bi_private != NULL) {
1325                 int dir = bio_data_dir(bio);
1326                 if (atomic_dec_and_test(&wc->bio_in_progress[dir]))
1327                         if (unlikely(waitqueue_active(&wc->bio_in_progress_wait[dir])))
1328                                 wake_up(&wc->bio_in_progress_wait[dir]);
1329         }
1330         return 0;
1331 }
1332
1333 static int writecache_iterate_devices(struct dm_target *ti,
1334                                       iterate_devices_callout_fn fn, void *data)
1335 {
1336         struct dm_writecache *wc = ti->private;
1337
1338         return fn(ti, wc->dev, 0, ti->len, data);
1339 }
1340
1341 static void writecache_io_hints(struct dm_target *ti, struct queue_limits *limits)
1342 {
1343         struct dm_writecache *wc = ti->private;
1344
1345         if (limits->logical_block_size < wc->block_size)
1346                 limits->logical_block_size = wc->block_size;
1347
1348         if (limits->physical_block_size < wc->block_size)
1349                 limits->physical_block_size = wc->block_size;
1350
1351         if (limits->io_min < wc->block_size)
1352                 limits->io_min = wc->block_size;
1353 }
1354
1355
1356 static void writecache_writeback_endio(struct bio *bio)
1357 {
1358         struct writeback_struct *wb = container_of(bio, struct writeback_struct, bio);
1359         struct dm_writecache *wc = wb->wc;
1360         unsigned long flags;
1361
1362         raw_spin_lock_irqsave(&wc->endio_list_lock, flags);
1363         if (unlikely(list_empty(&wc->endio_list)))
1364                 wake_up_process(wc->endio_thread);
1365         list_add_tail(&wb->endio_entry, &wc->endio_list);
1366         raw_spin_unlock_irqrestore(&wc->endio_list_lock, flags);
1367 }
1368
1369 static void writecache_copy_endio(int read_err, unsigned long write_err, void *ptr)
1370 {
1371         struct copy_struct *c = ptr;
1372         struct dm_writecache *wc = c->wc;
1373
1374         c->error = likely(!(read_err | write_err)) ? 0 : -EIO;
1375
1376         raw_spin_lock_irq(&wc->endio_list_lock);
1377         if (unlikely(list_empty(&wc->endio_list)))
1378                 wake_up_process(wc->endio_thread);
1379         list_add_tail(&c->endio_entry, &wc->endio_list);
1380         raw_spin_unlock_irq(&wc->endio_list_lock);
1381 }
1382
1383 static void __writecache_endio_pmem(struct dm_writecache *wc, struct list_head *list)
1384 {
1385         unsigned i;
1386         struct writeback_struct *wb;
1387         struct wc_entry *e;
1388         unsigned long n_walked = 0;
1389
1390         do {
1391                 wb = list_entry(list->next, struct writeback_struct, endio_entry);
1392                 list_del(&wb->endio_entry);
1393
1394                 if (unlikely(wb->bio.bi_status != BLK_STS_OK))
1395                         writecache_error(wc, blk_status_to_errno(wb->bio.bi_status),
1396                                         "write error %d", wb->bio.bi_status);
1397                 i = 0;
1398                 do {
1399                         e = wb->wc_list[i];
1400                         BUG_ON(!e->write_in_progress);
1401                         e->write_in_progress = false;
1402                         INIT_LIST_HEAD(&e->lru);
1403                         if (!writecache_has_error(wc))
1404                                 writecache_free_entry(wc, e);
1405                         BUG_ON(!wc->writeback_size);
1406                         wc->writeback_size--;
1407                         n_walked++;
1408                         if (unlikely(n_walked >= ENDIO_LATENCY)) {
1409                                 writecache_commit_flushed(wc, false);
1410                                 wc_unlock(wc);
1411                                 wc_lock(wc);
1412                                 n_walked = 0;
1413                         }
1414                 } while (++i < wb->wc_list_n);
1415
1416                 if (wb->wc_list != wb->wc_list_inline)
1417                         kfree(wb->wc_list);
1418                 bio_put(&wb->bio);
1419         } while (!list_empty(list));
1420 }
1421
1422 static void __writecache_endio_ssd(struct dm_writecache *wc, struct list_head *list)
1423 {
1424         struct copy_struct *c;
1425         struct wc_entry *e;
1426
1427         do {
1428                 c = list_entry(list->next, struct copy_struct, endio_entry);
1429                 list_del(&c->endio_entry);
1430
1431                 if (unlikely(c->error))
1432                         writecache_error(wc, c->error, "copy error");
1433
1434                 e = c->e;
1435                 do {
1436                         BUG_ON(!e->write_in_progress);
1437                         e->write_in_progress = false;
1438                         INIT_LIST_HEAD(&e->lru);
1439                         if (!writecache_has_error(wc))
1440                                 writecache_free_entry(wc, e);
1441
1442                         BUG_ON(!wc->writeback_size);
1443                         wc->writeback_size--;
1444                         e++;
1445                 } while (--c->n_entries);
1446                 mempool_free(c, &wc->copy_pool);
1447         } while (!list_empty(list));
1448 }
1449
1450 static int writecache_endio_thread(void *data)
1451 {
1452         struct dm_writecache *wc = data;
1453
1454         while (1) {
1455                 struct list_head list;
1456
1457                 raw_spin_lock_irq(&wc->endio_list_lock);
1458                 if (!list_empty(&wc->endio_list))
1459                         goto pop_from_list;
1460                 set_current_state(TASK_INTERRUPTIBLE);
1461                 raw_spin_unlock_irq(&wc->endio_list_lock);
1462
1463                 if (unlikely(kthread_should_stop())) {
1464                         set_current_state(TASK_RUNNING);
1465                         break;
1466                 }
1467
1468                 schedule();
1469
1470                 continue;
1471
1472 pop_from_list:
1473                 list = wc->endio_list;
1474                 list.next->prev = list.prev->next = &list;
1475                 INIT_LIST_HEAD(&wc->endio_list);
1476                 raw_spin_unlock_irq(&wc->endio_list_lock);
1477
1478                 if (!WC_MODE_FUA(wc))
1479                         writecache_disk_flush(wc, wc->dev);
1480
1481                 wc_lock(wc);
1482
1483                 if (WC_MODE_PMEM(wc)) {
1484                         __writecache_endio_pmem(wc, &list);
1485                 } else {
1486                         __writecache_endio_ssd(wc, &list);
1487                         writecache_wait_for_ios(wc, READ);
1488                 }
1489
1490                 writecache_commit_flushed(wc, false);
1491
1492                 wc_unlock(wc);
1493         }
1494
1495         return 0;
1496 }
1497
1498 static bool wc_add_block(struct writeback_struct *wb, struct wc_entry *e, gfp_t gfp)
1499 {
1500         struct dm_writecache *wc = wb->wc;
1501         unsigned block_size = wc->block_size;
1502         void *address = memory_data(wc, e);
1503
1504         persistent_memory_flush_cache(address, block_size);
1505
1506         if (unlikely(bio_end_sector(&wb->bio) >= wc->data_device_sectors))
1507                 return true;
1508
1509         return bio_add_page(&wb->bio, persistent_memory_page(address),
1510                             block_size, persistent_memory_page_offset(address)) != 0;
1511 }
1512
1513 struct writeback_list {
1514         struct list_head list;
1515         size_t size;
1516 };
1517
1518 static void __writeback_throttle(struct dm_writecache *wc, struct writeback_list *wbl)
1519 {
1520         if (unlikely(wc->max_writeback_jobs)) {
1521                 if (READ_ONCE(wc->writeback_size) - wbl->size >= wc->max_writeback_jobs) {
1522                         wc_lock(wc);
1523                         while (wc->writeback_size - wbl->size >= wc->max_writeback_jobs)
1524                                 writecache_wait_on_freelist(wc);
1525                         wc_unlock(wc);
1526                 }
1527         }
1528         cond_resched();
1529 }
1530
1531 static void __writecache_writeback_pmem(struct dm_writecache *wc, struct writeback_list *wbl)
1532 {
1533         struct wc_entry *e, *f;
1534         struct bio *bio;
1535         struct writeback_struct *wb;
1536         unsigned max_pages;
1537
1538         while (wbl->size) {
1539                 wbl->size--;
1540                 e = container_of(wbl->list.prev, struct wc_entry, lru);
1541                 list_del(&e->lru);
1542
1543                 max_pages = e->wc_list_contiguous;
1544
1545                 bio = bio_alloc_bioset(GFP_NOIO, max_pages, &wc->bio_set);
1546                 wb = container_of(bio, struct writeback_struct, bio);
1547                 wb->wc = wc;
1548                 wb->bio.bi_end_io = writecache_writeback_endio;
1549                 bio_set_dev(&wb->bio, wc->dev->bdev);
1550                 wb->bio.bi_iter.bi_sector = read_original_sector(wc, e);
1551                 wb->page_offset = PAGE_SIZE;
1552                 if (max_pages <= WB_LIST_INLINE ||
1553                     unlikely(!(wb->wc_list = kmalloc_array(max_pages, sizeof(struct wc_entry *),
1554                                                            GFP_NOIO | __GFP_NORETRY |
1555                                                            __GFP_NOMEMALLOC | __GFP_NOWARN)))) {
1556                         wb->wc_list = wb->wc_list_inline;
1557                         max_pages = WB_LIST_INLINE;
1558                 }
1559
1560                 BUG_ON(!wc_add_block(wb, e, GFP_NOIO));
1561
1562                 wb->wc_list[0] = e;
1563                 wb->wc_list_n = 1;
1564
1565                 while (wbl->size && wb->wc_list_n < max_pages) {
1566                         f = container_of(wbl->list.prev, struct wc_entry, lru);
1567                         if (read_original_sector(wc, f) !=
1568                             read_original_sector(wc, e) + (wc->block_size >> SECTOR_SHIFT))
1569                                 break;
1570                         if (!wc_add_block(wb, f, GFP_NOWAIT | __GFP_NOWARN))
1571                                 break;
1572                         wbl->size--;
1573                         list_del(&f->lru);
1574                         wb->wc_list[wb->wc_list_n++] = f;
1575                         e = f;
1576                 }
1577                 bio_set_op_attrs(&wb->bio, REQ_OP_WRITE, WC_MODE_FUA(wc) * REQ_FUA);
1578                 if (writecache_has_error(wc)) {
1579                         bio->bi_status = BLK_STS_IOERR;
1580                         bio_endio(&wb->bio);
1581                 } else if (unlikely(!bio_sectors(&wb->bio))) {
1582                         bio->bi_status = BLK_STS_OK;
1583                         bio_endio(&wb->bio);
1584                 } else {
1585                         submit_bio(&wb->bio);
1586                 }
1587
1588                 __writeback_throttle(wc, wbl);
1589         }
1590 }
1591
1592 static void __writecache_writeback_ssd(struct dm_writecache *wc, struct writeback_list *wbl)
1593 {
1594         struct wc_entry *e, *f;
1595         struct dm_io_region from, to;
1596         struct copy_struct *c;
1597
1598         while (wbl->size) {
1599                 unsigned n_sectors;
1600
1601                 wbl->size--;
1602                 e = container_of(wbl->list.prev, struct wc_entry, lru);
1603                 list_del(&e->lru);
1604
1605                 n_sectors = e->wc_list_contiguous << (wc->block_size_bits - SECTOR_SHIFT);
1606
1607                 from.bdev = wc->ssd_dev->bdev;
1608                 from.sector = cache_sector(wc, e);
1609                 from.count = n_sectors;
1610                 to.bdev = wc->dev->bdev;
1611                 to.sector = read_original_sector(wc, e);
1612                 to.count = n_sectors;
1613
1614                 c = mempool_alloc(&wc->copy_pool, GFP_NOIO);
1615                 c->wc = wc;
1616                 c->e = e;
1617                 c->n_entries = e->wc_list_contiguous;
1618
1619                 while ((n_sectors -= wc->block_size >> SECTOR_SHIFT)) {
1620                         wbl->size--;
1621                         f = container_of(wbl->list.prev, struct wc_entry, lru);
1622                         BUG_ON(f != e + 1);
1623                         list_del(&f->lru);
1624                         e = f;
1625                 }
1626
1627                 if (unlikely(to.sector + to.count > wc->data_device_sectors)) {
1628                         if (to.sector >= wc->data_device_sectors) {
1629                                 writecache_copy_endio(0, 0, c);
1630                                 continue;
1631                         }
1632                         from.count = to.count = wc->data_device_sectors - to.sector;
1633                 }
1634
1635                 dm_kcopyd_copy(wc->dm_kcopyd, &from, 1, &to, 0, writecache_copy_endio, c);
1636
1637                 __writeback_throttle(wc, wbl);
1638         }
1639 }
1640
1641 static void writecache_writeback(struct work_struct *work)
1642 {
1643         struct dm_writecache *wc = container_of(work, struct dm_writecache, writeback_work);
1644         struct blk_plug plug;
1645         struct wc_entry *e, *f, *g;
1646         struct rb_node *node, *next_node;
1647         struct list_head skipped;
1648         struct writeback_list wbl;
1649         unsigned long n_walked;
1650
1651         wc_lock(wc);
1652 restart:
1653         if (writecache_has_error(wc)) {
1654                 wc_unlock(wc);
1655                 return;
1656         }
1657
1658         if (unlikely(wc->writeback_all)) {
1659                 if (writecache_wait_for_writeback(wc))
1660                         goto restart;
1661         }
1662
1663         if (wc->overwrote_committed) {
1664                 writecache_wait_for_ios(wc, WRITE);
1665         }
1666
1667         n_walked = 0;
1668         INIT_LIST_HEAD(&skipped);
1669         INIT_LIST_HEAD(&wbl.list);
1670         wbl.size = 0;
1671         while (!list_empty(&wc->lru) &&
1672                (wc->writeback_all ||
1673                 wc->freelist_size + wc->writeback_size <= wc->freelist_low_watermark)) {
1674
1675                 n_walked++;
1676                 if (unlikely(n_walked > WRITEBACK_LATENCY) &&
1677                     likely(!wc->writeback_all) && likely(!dm_suspended(wc->ti))) {
1678                         queue_work(wc->writeback_wq, &wc->writeback_work);
1679                         break;
1680                 }
1681
1682                 e = container_of(wc->lru.prev, struct wc_entry, lru);
1683                 BUG_ON(e->write_in_progress);
1684                 if (unlikely(!writecache_entry_is_committed(wc, e))) {
1685                         writecache_flush(wc);
1686                 }
1687                 node = rb_prev(&e->rb_node);
1688                 if (node) {
1689                         f = container_of(node, struct wc_entry, rb_node);
1690                         if (unlikely(read_original_sector(wc, f) ==
1691                                      read_original_sector(wc, e))) {
1692                                 BUG_ON(!f->write_in_progress);
1693                                 list_del(&e->lru);
1694                                 list_add(&e->lru, &skipped);
1695                                 cond_resched();
1696                                 continue;
1697                         }
1698                 }
1699                 wc->writeback_size++;
1700                 list_del(&e->lru);
1701                 list_add(&e->lru, &wbl.list);
1702                 wbl.size++;
1703                 e->write_in_progress = true;
1704                 e->wc_list_contiguous = 1;
1705
1706                 f = e;
1707
1708                 while (1) {
1709                         next_node = rb_next(&f->rb_node);
1710                         if (unlikely(!next_node))
1711                                 break;
1712                         g = container_of(next_node, struct wc_entry, rb_node);
1713                         if (read_original_sector(wc, g) ==
1714                             read_original_sector(wc, f)) {
1715                                 f = g;
1716                                 continue;
1717                         }
1718                         if (read_original_sector(wc, g) !=
1719                             read_original_sector(wc, f) + (wc->block_size >> SECTOR_SHIFT))
1720                                 break;
1721                         if (unlikely(g->write_in_progress))
1722                                 break;
1723                         if (unlikely(!writecache_entry_is_committed(wc, g)))
1724                                 break;
1725
1726                         if (!WC_MODE_PMEM(wc)) {
1727                                 if (g != f + 1)
1728                                         break;
1729                         }
1730
1731                         n_walked++;
1732                         //if (unlikely(n_walked > WRITEBACK_LATENCY) && likely(!wc->writeback_all))
1733                         //      break;
1734
1735                         wc->writeback_size++;
1736                         list_del(&g->lru);
1737                         list_add(&g->lru, &wbl.list);
1738                         wbl.size++;
1739                         g->write_in_progress = true;
1740                         g->wc_list_contiguous = BIO_MAX_PAGES;
1741                         f = g;
1742                         e->wc_list_contiguous++;
1743                         if (unlikely(e->wc_list_contiguous == BIO_MAX_PAGES))
1744                                 break;
1745                 }
1746                 cond_resched();
1747         }
1748
1749         if (!list_empty(&skipped)) {
1750                 list_splice_tail(&skipped, &wc->lru);
1751                 /*
1752                  * If we didn't do any progress, we must wait until some
1753                  * writeback finishes to avoid burning CPU in a loop
1754                  */
1755                 if (unlikely(!wbl.size))
1756                         writecache_wait_for_writeback(wc);
1757         }
1758
1759         wc_unlock(wc);
1760
1761         blk_start_plug(&plug);
1762
1763         if (WC_MODE_PMEM(wc))
1764                 __writecache_writeback_pmem(wc, &wbl);
1765         else
1766                 __writecache_writeback_ssd(wc, &wbl);
1767
1768         blk_finish_plug(&plug);
1769
1770         if (unlikely(wc->writeback_all)) {
1771                 wc_lock(wc);
1772                 while (writecache_wait_for_writeback(wc));
1773                 wc_unlock(wc);
1774         }
1775 }
1776
1777 static int calculate_memory_size(uint64_t device_size, unsigned block_size,
1778                                  size_t *n_blocks_p, size_t *n_metadata_blocks_p)
1779 {
1780         uint64_t n_blocks, offset;
1781         struct wc_entry e;
1782
1783         n_blocks = device_size;
1784         do_div(n_blocks, block_size + sizeof(struct wc_memory_entry));
1785
1786         while (1) {
1787                 if (!n_blocks)
1788                         return -ENOSPC;
1789                 /* Verify the following entries[n_blocks] won't overflow */
1790                 if (n_blocks >= ((size_t)-sizeof(struct wc_memory_superblock) /
1791                                  sizeof(struct wc_memory_entry)))
1792                         return -EFBIG;
1793                 offset = offsetof(struct wc_memory_superblock, entries[n_blocks]);
1794                 offset = (offset + block_size - 1) & ~(uint64_t)(block_size - 1);
1795                 if (offset + n_blocks * block_size <= device_size)
1796                         break;
1797                 n_blocks--;
1798         }
1799
1800         /* check if the bit field overflows */
1801         e.index = n_blocks;
1802         if (e.index != n_blocks)
1803                 return -EFBIG;
1804
1805         if (n_blocks_p)
1806                 *n_blocks_p = n_blocks;
1807         if (n_metadata_blocks_p)
1808                 *n_metadata_blocks_p = offset >> __ffs(block_size);
1809         return 0;
1810 }
1811
1812 static int init_memory(struct dm_writecache *wc)
1813 {
1814         size_t b;
1815         int r;
1816
1817         r = calculate_memory_size(wc->memory_map_size, wc->block_size, &wc->n_blocks, NULL);
1818         if (r)
1819                 return r;
1820
1821         r = writecache_alloc_entries(wc);
1822         if (r)
1823                 return r;
1824
1825         for (b = 0; b < ARRAY_SIZE(sb(wc)->padding); b++)
1826                 pmem_assign(sb(wc)->padding[b], cpu_to_le64(0));
1827         pmem_assign(sb(wc)->version, cpu_to_le32(MEMORY_SUPERBLOCK_VERSION));
1828         pmem_assign(sb(wc)->block_size, cpu_to_le32(wc->block_size));
1829         pmem_assign(sb(wc)->n_blocks, cpu_to_le64(wc->n_blocks));
1830         pmem_assign(sb(wc)->seq_count, cpu_to_le64(0));
1831
1832         for (b = 0; b < wc->n_blocks; b++) {
1833                 write_original_sector_seq_count(wc, &wc->entries[b], -1, -1);
1834                 cond_resched();
1835         }
1836
1837         writecache_flush_all_metadata(wc);
1838         writecache_commit_flushed(wc, false);
1839         pmem_assign(sb(wc)->magic, cpu_to_le32(MEMORY_SUPERBLOCK_MAGIC));
1840         writecache_flush_region(wc, &sb(wc)->magic, sizeof sb(wc)->magic);
1841         writecache_commit_flushed(wc, false);
1842
1843         return 0;
1844 }
1845
1846 static void writecache_dtr(struct dm_target *ti)
1847 {
1848         struct dm_writecache *wc = ti->private;
1849
1850         if (!wc)
1851                 return;
1852
1853         if (wc->endio_thread)
1854                 kthread_stop(wc->endio_thread);
1855
1856         if (wc->flush_thread)
1857                 kthread_stop(wc->flush_thread);
1858
1859         bioset_exit(&wc->bio_set);
1860
1861         mempool_exit(&wc->copy_pool);
1862
1863         if (wc->writeback_wq)
1864                 destroy_workqueue(wc->writeback_wq);
1865
1866         if (wc->dev)
1867                 dm_put_device(ti, wc->dev);
1868
1869         if (wc->ssd_dev)
1870                 dm_put_device(ti, wc->ssd_dev);
1871
1872         if (wc->entries)
1873                 vfree(wc->entries);
1874
1875         if (wc->memory_map) {
1876                 if (WC_MODE_PMEM(wc))
1877                         persistent_memory_release(wc);
1878                 else
1879                         vfree(wc->memory_map);
1880         }
1881
1882         if (wc->dm_kcopyd)
1883                 dm_kcopyd_client_destroy(wc->dm_kcopyd);
1884
1885         if (wc->dm_io)
1886                 dm_io_client_destroy(wc->dm_io);
1887
1888         if (wc->dirty_bitmap)
1889                 vfree(wc->dirty_bitmap);
1890
1891         kfree(wc);
1892 }
1893
1894 static int writecache_ctr(struct dm_target *ti, unsigned argc, char **argv)
1895 {
1896         struct dm_writecache *wc;
1897         struct dm_arg_set as;
1898         const char *string;
1899         unsigned opt_params;
1900         size_t offset, data_size;
1901         int i, r;
1902         char dummy;
1903         int high_wm_percent = HIGH_WATERMARK;
1904         int low_wm_percent = LOW_WATERMARK;
1905         uint64_t x;
1906         struct wc_memory_superblock s;
1907
1908         static struct dm_arg _args[] = {
1909                 {0, 16, "Invalid number of feature args"},
1910         };
1911
1912         as.argc = argc;
1913         as.argv = argv;
1914
1915         wc = kzalloc(sizeof(struct dm_writecache), GFP_KERNEL);
1916         if (!wc) {
1917                 ti->error = "Cannot allocate writecache structure";
1918                 r = -ENOMEM;
1919                 goto bad;
1920         }
1921         ti->private = wc;
1922         wc->ti = ti;
1923
1924         mutex_init(&wc->lock);
1925         writecache_poison_lists(wc);
1926         init_waitqueue_head(&wc->freelist_wait);
1927         timer_setup(&wc->autocommit_timer, writecache_autocommit_timer, 0);
1928
1929         for (i = 0; i < 2; i++) {
1930                 atomic_set(&wc->bio_in_progress[i], 0);
1931                 init_waitqueue_head(&wc->bio_in_progress_wait[i]);
1932         }
1933
1934         wc->dm_io = dm_io_client_create();
1935         if (IS_ERR(wc->dm_io)) {
1936                 r = PTR_ERR(wc->dm_io);
1937                 ti->error = "Unable to allocate dm-io client";
1938                 wc->dm_io = NULL;
1939                 goto bad;
1940         }
1941
1942         wc->writeback_wq = alloc_workqueue("writecache-writeabck", WQ_MEM_RECLAIM, 1);
1943         if (!wc->writeback_wq) {
1944                 r = -ENOMEM;
1945                 ti->error = "Could not allocate writeback workqueue";
1946                 goto bad;
1947         }
1948         INIT_WORK(&wc->writeback_work, writecache_writeback);
1949         INIT_WORK(&wc->flush_work, writecache_flush_work);
1950
1951         raw_spin_lock_init(&wc->endio_list_lock);
1952         INIT_LIST_HEAD(&wc->endio_list);
1953         wc->endio_thread = kthread_create(writecache_endio_thread, wc, "writecache_endio");
1954         if (IS_ERR(wc->endio_thread)) {
1955                 r = PTR_ERR(wc->endio_thread);
1956                 wc->endio_thread = NULL;
1957                 ti->error = "Couldn't spawn endio thread";
1958                 goto bad;
1959         }
1960         wake_up_process(wc->endio_thread);
1961
1962         /*
1963          * Parse the mode (pmem or ssd)
1964          */
1965         string = dm_shift_arg(&as);
1966         if (!string)
1967                 goto bad_arguments;
1968
1969         if (!strcasecmp(string, "s")) {
1970                 wc->pmem_mode = false;
1971         } else if (!strcasecmp(string, "p")) {
1972 #ifdef DM_WRITECACHE_HAS_PMEM
1973                 wc->pmem_mode = true;
1974                 wc->writeback_fua = true;
1975 #else
1976                 /*
1977                  * If the architecture doesn't support persistent memory or
1978                  * the kernel doesn't support any DAX drivers, this driver can
1979                  * only be used in SSD-only mode.
1980                  */
1981                 r = -EOPNOTSUPP;
1982                 ti->error = "Persistent memory or DAX not supported on this system";
1983                 goto bad;
1984 #endif
1985         } else {
1986                 goto bad_arguments;
1987         }
1988
1989         if (WC_MODE_PMEM(wc)) {
1990                 r = bioset_init(&wc->bio_set, BIO_POOL_SIZE,
1991                                 offsetof(struct writeback_struct, bio),
1992                                 BIOSET_NEED_BVECS);
1993                 if (r) {
1994                         ti->error = "Could not allocate bio set";
1995                         goto bad;
1996                 }
1997         } else {
1998                 r = mempool_init_kmalloc_pool(&wc->copy_pool, 1, sizeof(struct copy_struct));
1999                 if (r) {
2000                         ti->error = "Could not allocate mempool";
2001                         goto bad;
2002                 }
2003         }
2004
2005         /*
2006          * Parse the origin data device
2007          */
2008         string = dm_shift_arg(&as);
2009         if (!string)
2010                 goto bad_arguments;
2011         r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->dev);
2012         if (r) {
2013                 ti->error = "Origin data device lookup failed";
2014                 goto bad;
2015         }
2016
2017         /*
2018          * Parse cache data device (be it pmem or ssd)
2019          */
2020         string = dm_shift_arg(&as);
2021         if (!string)
2022                 goto bad_arguments;
2023
2024         r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->ssd_dev);
2025         if (r) {
2026                 ti->error = "Cache data device lookup failed";
2027                 goto bad;
2028         }
2029         wc->memory_map_size = i_size_read(wc->ssd_dev->bdev->bd_inode);
2030
2031         /*
2032          * Parse the cache block size
2033          */
2034         string = dm_shift_arg(&as);
2035         if (!string)
2036                 goto bad_arguments;
2037         if (sscanf(string, "%u%c", &wc->block_size, &dummy) != 1 ||
2038             wc->block_size < 512 || wc->block_size > PAGE_SIZE ||
2039             (wc->block_size & (wc->block_size - 1))) {
2040                 r = -EINVAL;
2041                 ti->error = "Invalid block size";
2042                 goto bad;
2043         }
2044         if (wc->block_size < bdev_logical_block_size(wc->dev->bdev) ||
2045             wc->block_size < bdev_logical_block_size(wc->ssd_dev->bdev)) {
2046                 r = -EINVAL;
2047                 ti->error = "Block size is smaller than device logical block size";
2048                 goto bad;
2049         }
2050         wc->block_size_bits = __ffs(wc->block_size);
2051
2052         wc->max_writeback_jobs = MAX_WRITEBACK_JOBS;
2053         wc->autocommit_blocks = !WC_MODE_PMEM(wc) ? AUTOCOMMIT_BLOCKS_SSD : AUTOCOMMIT_BLOCKS_PMEM;
2054         wc->autocommit_jiffies = msecs_to_jiffies(AUTOCOMMIT_MSEC);
2055
2056         /*
2057          * Parse optional arguments
2058          */
2059         r = dm_read_arg_group(_args, &as, &opt_params, &ti->error);
2060         if (r)
2061                 goto bad;
2062
2063         while (opt_params) {
2064                 string = dm_shift_arg(&as), opt_params--;
2065                 if (!strcasecmp(string, "start_sector") && opt_params >= 1) {
2066                         unsigned long long start_sector;
2067                         string = dm_shift_arg(&as), opt_params--;
2068                         if (sscanf(string, "%llu%c", &start_sector, &dummy) != 1)
2069                                 goto invalid_optional;
2070                         wc->start_sector = start_sector;
2071                         wc->start_sector_set = true;
2072                         if (wc->start_sector != start_sector ||
2073                             wc->start_sector >= wc->memory_map_size >> SECTOR_SHIFT)
2074                                 goto invalid_optional;
2075                 } else if (!strcasecmp(string, "high_watermark") && opt_params >= 1) {
2076                         string = dm_shift_arg(&as), opt_params--;
2077                         if (sscanf(string, "%d%c", &high_wm_percent, &dummy) != 1)
2078                                 goto invalid_optional;
2079                         if (high_wm_percent < 0 || high_wm_percent > 100)
2080                                 goto invalid_optional;
2081                         wc->high_wm_percent_value = high_wm_percent;
2082                         wc->high_wm_percent_set = true;
2083                 } else if (!strcasecmp(string, "low_watermark") && opt_params >= 1) {
2084                         string = dm_shift_arg(&as), opt_params--;
2085                         if (sscanf(string, "%d%c", &low_wm_percent, &dummy) != 1)
2086                                 goto invalid_optional;
2087                         if (low_wm_percent < 0 || low_wm_percent > 100)
2088                                 goto invalid_optional;
2089                         wc->low_wm_percent_value = low_wm_percent;
2090                         wc->low_wm_percent_set = true;
2091                 } else if (!strcasecmp(string, "writeback_jobs") && opt_params >= 1) {
2092                         string = dm_shift_arg(&as), opt_params--;
2093                         if (sscanf(string, "%u%c", &wc->max_writeback_jobs, &dummy) != 1)
2094                                 goto invalid_optional;
2095                         wc->max_writeback_jobs_set = true;
2096                 } else if (!strcasecmp(string, "autocommit_blocks") && opt_params >= 1) {
2097                         string = dm_shift_arg(&as), opt_params--;
2098                         if (sscanf(string, "%u%c", &wc->autocommit_blocks, &dummy) != 1)
2099                                 goto invalid_optional;
2100                         wc->autocommit_blocks_set = true;
2101                 } else if (!strcasecmp(string, "autocommit_time") && opt_params >= 1) {
2102                         unsigned autocommit_msecs;
2103                         string = dm_shift_arg(&as), opt_params--;
2104                         if (sscanf(string, "%u%c", &autocommit_msecs, &dummy) != 1)
2105                                 goto invalid_optional;
2106                         if (autocommit_msecs > 3600000)
2107                                 goto invalid_optional;
2108                         wc->autocommit_jiffies = msecs_to_jiffies(autocommit_msecs);
2109                         wc->autocommit_time_value = autocommit_msecs;
2110                         wc->autocommit_time_set = true;
2111                 } else if (!strcasecmp(string, "fua")) {
2112                         if (WC_MODE_PMEM(wc)) {
2113                                 wc->writeback_fua = true;
2114                                 wc->writeback_fua_set = true;
2115                         } else goto invalid_optional;
2116                 } else if (!strcasecmp(string, "nofua")) {
2117                         if (WC_MODE_PMEM(wc)) {
2118                                 wc->writeback_fua = false;
2119                                 wc->writeback_fua_set = true;
2120                         } else goto invalid_optional;
2121                 } else {
2122 invalid_optional:
2123                         r = -EINVAL;
2124                         ti->error = "Invalid optional argument";
2125                         goto bad;
2126                 }
2127         }
2128
2129         if (high_wm_percent < low_wm_percent) {
2130                 r = -EINVAL;
2131                 ti->error = "High watermark must be greater than or equal to low watermark";
2132                 goto bad;
2133         }
2134
2135         if (WC_MODE_PMEM(wc)) {
2136                 r = persistent_memory_claim(wc);
2137                 if (r) {
2138                         ti->error = "Unable to map persistent memory for cache";
2139                         goto bad;
2140                 }
2141         } else {
2142                 size_t n_blocks, n_metadata_blocks;
2143                 uint64_t n_bitmap_bits;
2144
2145                 wc->memory_map_size -= (uint64_t)wc->start_sector << SECTOR_SHIFT;
2146
2147                 bio_list_init(&wc->flush_list);
2148                 wc->flush_thread = kthread_create(writecache_flush_thread, wc, "dm_writecache_flush");
2149                 if (IS_ERR(wc->flush_thread)) {
2150                         r = PTR_ERR(wc->flush_thread);
2151                         wc->flush_thread = NULL;
2152                         ti->error = "Couldn't spawn endio thread";
2153                         goto bad;
2154                 }
2155                 wake_up_process(wc->flush_thread);
2156
2157                 r = calculate_memory_size(wc->memory_map_size, wc->block_size,
2158                                           &n_blocks, &n_metadata_blocks);
2159                 if (r) {
2160                         ti->error = "Invalid device size";
2161                         goto bad;
2162                 }
2163
2164                 n_bitmap_bits = (((uint64_t)n_metadata_blocks << wc->block_size_bits) +
2165                                  BITMAP_GRANULARITY - 1) / BITMAP_GRANULARITY;
2166                 /* this is limitation of test_bit functions */
2167                 if (n_bitmap_bits > 1U << 31) {
2168                         r = -EFBIG;
2169                         ti->error = "Invalid device size";
2170                         goto bad;
2171                 }
2172
2173                 wc->memory_map = vmalloc(n_metadata_blocks << wc->block_size_bits);
2174                 if (!wc->memory_map) {
2175                         r = -ENOMEM;
2176                         ti->error = "Unable to allocate memory for metadata";
2177                         goto bad;
2178                 }
2179
2180                 wc->dm_kcopyd = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2181                 if (IS_ERR(wc->dm_kcopyd)) {
2182                         r = PTR_ERR(wc->dm_kcopyd);
2183                         ti->error = "Unable to allocate dm-kcopyd client";
2184                         wc->dm_kcopyd = NULL;
2185                         goto bad;
2186                 }
2187
2188                 wc->metadata_sectors = n_metadata_blocks << (wc->block_size_bits - SECTOR_SHIFT);
2189                 wc->dirty_bitmap_size = (n_bitmap_bits + BITS_PER_LONG - 1) /
2190                         BITS_PER_LONG * sizeof(unsigned long);
2191                 wc->dirty_bitmap = vzalloc(wc->dirty_bitmap_size);
2192                 if (!wc->dirty_bitmap) {
2193                         r = -ENOMEM;
2194                         ti->error = "Unable to allocate dirty bitmap";
2195                         goto bad;
2196                 }
2197
2198                 r = writecache_read_metadata(wc, wc->block_size >> SECTOR_SHIFT);
2199                 if (r) {
2200                         ti->error = "Unable to read first block of metadata";
2201                         goto bad;
2202                 }
2203         }
2204
2205         r = memcpy_mcsafe(&s, sb(wc), sizeof(struct wc_memory_superblock));
2206         if (r) {
2207                 ti->error = "Hardware memory error when reading superblock";
2208                 goto bad;
2209         }
2210         if (!le32_to_cpu(s.magic) && !le32_to_cpu(s.version)) {
2211                 r = init_memory(wc);
2212                 if (r) {
2213                         ti->error = "Unable to initialize device";
2214                         goto bad;
2215                 }
2216                 r = memcpy_mcsafe(&s, sb(wc), sizeof(struct wc_memory_superblock));
2217                 if (r) {
2218                         ti->error = "Hardware memory error when reading superblock";
2219                         goto bad;
2220                 }
2221         }
2222
2223         if (le32_to_cpu(s.magic) != MEMORY_SUPERBLOCK_MAGIC) {
2224                 ti->error = "Invalid magic in the superblock";
2225                 r = -EINVAL;
2226                 goto bad;
2227         }
2228
2229         if (le32_to_cpu(s.version) != MEMORY_SUPERBLOCK_VERSION) {
2230                 ti->error = "Invalid version in the superblock";
2231                 r = -EINVAL;
2232                 goto bad;
2233         }
2234
2235         if (le32_to_cpu(s.block_size) != wc->block_size) {
2236                 ti->error = "Block size does not match superblock";
2237                 r = -EINVAL;
2238                 goto bad;
2239         }
2240
2241         wc->n_blocks = le64_to_cpu(s.n_blocks);
2242
2243         offset = wc->n_blocks * sizeof(struct wc_memory_entry);
2244         if (offset / sizeof(struct wc_memory_entry) != le64_to_cpu(sb(wc)->n_blocks)) {
2245 overflow:
2246                 ti->error = "Overflow in size calculation";
2247                 r = -EINVAL;
2248                 goto bad;
2249         }
2250         offset += sizeof(struct wc_memory_superblock);
2251         if (offset < sizeof(struct wc_memory_superblock))
2252                 goto overflow;
2253         offset = (offset + wc->block_size - 1) & ~(size_t)(wc->block_size - 1);
2254         data_size = wc->n_blocks * (size_t)wc->block_size;
2255         if (!offset || (data_size / wc->block_size != wc->n_blocks) ||
2256             (offset + data_size < offset))
2257                 goto overflow;
2258         if (offset + data_size > wc->memory_map_size) {
2259                 ti->error = "Memory area is too small";
2260                 r = -EINVAL;
2261                 goto bad;
2262         }
2263
2264         wc->metadata_sectors = offset >> SECTOR_SHIFT;
2265         wc->block_start = (char *)sb(wc) + offset;
2266
2267         x = (uint64_t)wc->n_blocks * (100 - high_wm_percent);
2268         x += 50;
2269         do_div(x, 100);
2270         wc->freelist_high_watermark = x;
2271         x = (uint64_t)wc->n_blocks * (100 - low_wm_percent);
2272         x += 50;
2273         do_div(x, 100);
2274         wc->freelist_low_watermark = x;
2275
2276         r = writecache_alloc_entries(wc);
2277         if (r) {
2278                 ti->error = "Cannot allocate memory";
2279                 goto bad;
2280         }
2281
2282         ti->num_flush_bios = 1;
2283         ti->flush_supported = true;
2284         ti->num_discard_bios = 1;
2285
2286         if (WC_MODE_PMEM(wc))
2287                 persistent_memory_flush_cache(wc->memory_map, wc->memory_map_size);
2288
2289         return 0;
2290
2291 bad_arguments:
2292         r = -EINVAL;
2293         ti->error = "Bad arguments";
2294 bad:
2295         writecache_dtr(ti);
2296         return r;
2297 }
2298
2299 static void writecache_status(struct dm_target *ti, status_type_t type,
2300                               unsigned status_flags, char *result, unsigned maxlen)
2301 {
2302         struct dm_writecache *wc = ti->private;
2303         unsigned extra_args;
2304         unsigned sz = 0;
2305
2306         switch (type) {
2307         case STATUSTYPE_INFO:
2308                 DMEMIT("%ld %llu %llu %llu", writecache_has_error(wc),
2309                        (unsigned long long)wc->n_blocks, (unsigned long long)wc->freelist_size,
2310                        (unsigned long long)wc->writeback_size);
2311                 break;
2312         case STATUSTYPE_TABLE:
2313                 DMEMIT("%c %s %s %u ", WC_MODE_PMEM(wc) ? 'p' : 's',
2314                                 wc->dev->name, wc->ssd_dev->name, wc->block_size);
2315                 extra_args = 0;
2316                 if (wc->start_sector_set)
2317                         extra_args += 2;
2318                 if (wc->high_wm_percent_set)
2319                         extra_args += 2;
2320                 if (wc->low_wm_percent_set)
2321                         extra_args += 2;
2322                 if (wc->max_writeback_jobs_set)
2323                         extra_args += 2;
2324                 if (wc->autocommit_blocks_set)
2325                         extra_args += 2;
2326                 if (wc->autocommit_time_set)
2327                         extra_args += 2;
2328                 if (wc->writeback_fua_set)
2329                         extra_args++;
2330
2331                 DMEMIT("%u", extra_args);
2332                 if (wc->start_sector_set)
2333                         DMEMIT(" start_sector %llu", (unsigned long long)wc->start_sector);
2334                 if (wc->high_wm_percent_set)
2335                         DMEMIT(" high_watermark %u", wc->high_wm_percent_value);
2336                 if (wc->low_wm_percent_set)
2337                         DMEMIT(" low_watermark %u", wc->low_wm_percent_value);
2338                 if (wc->max_writeback_jobs_set)
2339                         DMEMIT(" writeback_jobs %u", wc->max_writeback_jobs);
2340                 if (wc->autocommit_blocks_set)
2341                         DMEMIT(" autocommit_blocks %u", wc->autocommit_blocks);
2342                 if (wc->autocommit_time_set)
2343                         DMEMIT(" autocommit_time %u", wc->autocommit_time_value);
2344                 if (wc->writeback_fua_set)
2345                         DMEMIT(" %sfua", wc->writeback_fua ? "" : "no");
2346                 break;
2347         }
2348 }
2349
2350 static struct target_type writecache_target = {
2351         .name                   = "writecache",
2352         .version                = {1, 1, 1},
2353         .module                 = THIS_MODULE,
2354         .ctr                    = writecache_ctr,
2355         .dtr                    = writecache_dtr,
2356         .status                 = writecache_status,
2357         .postsuspend            = writecache_suspend,
2358         .resume                 = writecache_resume,
2359         .message                = writecache_message,
2360         .map                    = writecache_map,
2361         .end_io                 = writecache_end_io,
2362         .iterate_devices        = writecache_iterate_devices,
2363         .io_hints               = writecache_io_hints,
2364 };
2365
2366 static int __init dm_writecache_init(void)
2367 {
2368         int r;
2369
2370         r = dm_register_target(&writecache_target);
2371         if (r < 0) {
2372                 DMERR("register failed %d", r);
2373                 return r;
2374         }
2375
2376         return 0;
2377 }
2378
2379 static void __exit dm_writecache_exit(void)
2380 {
2381         dm_unregister_target(&writecache_target);
2382 }
2383
2384 module_init(dm_writecache_init);
2385 module_exit(dm_writecache_exit);
2386
2387 MODULE_DESCRIPTION(DM_NAME " writecache target");
2388 MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
2389 MODULE_LICENSE("GPL");