drivers/md/dm-writecache.c

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