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 #define MAX_AGE_DIV                     16
  30 #define MAX_AGE_UNSPECIFIED             -1UL
  31
  32 #define BITMAP_GRANULARITY      65536
  33 #if BITMAP_GRANULARITY < PAGE_SIZE
  34 #undef BITMAP_GRANULARITY
  35 #define BITMAP_GRANULARITY      PAGE_SIZE
  36 #endif
  37
  38 #if IS_ENABLED(CONFIG_ARCH_HAS_PMEM_API) && IS_ENABLED(CONFIG_DAX_DRIVER)
  39 #define DM_WRITECACHE_HAS_PMEM
  40 #endif
  41
  42 #ifdef DM_WRITECACHE_HAS_PMEM
  43 #define pmem_assign(dest, src)                                  \
  44 do {                                                            \
  45         typeof(dest) uniq = (src);                              \
  46         memcpy_flushcache(&(dest), &uniq, sizeof(dest));        \
  47 } while (0)
  48 #else
  49 #define pmem_assign(dest, src)  ((dest) = (src))
  50 #endif
  51
  52 #if defined(__HAVE_ARCH_MEMCPY_MCSAFE) && defined(DM_WRITECACHE_HAS_PMEM)
  53 #define DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
  54 #endif
  55
  56 #define MEMORY_SUPERBLOCK_MAGIC         0x23489321
  57 #define MEMORY_SUPERBLOCK_VERSION       1
  58
  59 struct wc_memory_entry {
  60         __le64 original_sector;
  61         __le64 seq_count;
  62 };
  63
  64 struct wc_memory_superblock {
  65         union {
  66                 struct {
  67                         __le32 magic;
  68                         __le32 version;
  69                         __le32 block_size;
  70                         __le32 pad;
  71                         __le64 n_blocks;
  72                         __le64 seq_count;
  73                 };
  74                 __le64 padding[8];
  75         };
  76         struct wc_memory_entry entries[0];
  77 };
  78
  79 struct wc_entry {
  80         struct rb_node rb_node;
  81         struct list_head lru;
  82         unsigned short wc_list_contiguous;
  83         bool write_in_progress
  84 #if BITS_PER_LONG == 64
  85                 :1
  86 #endif
  87         ;
  88         unsigned long index
  89 #if BITS_PER_LONG == 64
  90                 :47
  91 #endif
  92         ;
  93         unsigned long age;
  94 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
  95         uint64_t original_sector;
  96         uint64_t seq_count;
  97 #endif
  98 };
  99
 100 #ifdef DM_WRITECACHE_HAS_PMEM
 101 #define WC_MODE_PMEM(wc)                        ((wc)->pmem_mode)
 102 #define WC_MODE_FUA(wc)                         ((wc)->writeback_fua)
 103 #else
 104 #define WC_MODE_PMEM(wc)                        false
 105 #define WC_MODE_FUA(wc)                         false
 106 #endif
 107 #define WC_MODE_SORT_FREELIST(wc)               (!WC_MODE_PMEM(wc))
 108
 109 struct dm_writecache {
 110         struct mutex lock;
 111         struct list_head lru;
 112         union {
 113                 struct list_head freelist;
 114                 struct {
 115                         struct rb_root freetree;
 116                         struct wc_entry *current_free;
 117                 };
 118         };
 119         struct rb_root tree;
 120
 121         size_t freelist_size;
 122         size_t writeback_size;
 123         size_t freelist_high_watermark;
 124         size_t freelist_low_watermark;
 125         unsigned long max_age;
 126
 127         unsigned uncommitted_blocks;
 128         unsigned autocommit_blocks;
 129         unsigned max_writeback_jobs;
 130
 131         int error;
 132
 133         unsigned long autocommit_jiffies;
 134         struct timer_list autocommit_timer;
 135         struct wait_queue_head freelist_wait;
 136
 137         struct timer_list max_age_timer;
 138
 139         atomic_t bio_in_progress[2];
 140         struct wait_queue_head bio_in_progress_wait[2];
 141
 142         struct dm_target *ti;
 143         struct dm_dev *dev;
 144         struct dm_dev *ssd_dev;
 145         sector_t start_sector;
 146         void *memory_map;
 147         uint64_t memory_map_size;
 148         size_t metadata_sectors;
 149         size_t n_blocks;
 150         uint64_t seq_count;
 151         void *block_start;
 152         struct wc_entry *entries;
 153         unsigned block_size;
 154         unsigned char block_size_bits;
 155
 156         bool pmem_mode:1;
 157         bool writeback_fua:1;
 158
 159         bool overwrote_committed:1;
 160         bool memory_vmapped:1;
 161
 162         bool high_wm_percent_set:1;
 163         bool low_wm_percent_set:1;
 164         bool max_writeback_jobs_set:1;
 165         bool autocommit_blocks_set:1;
 166         bool autocommit_time_set:1;
 167         bool writeback_fua_set:1;
 168         bool flush_on_suspend:1;
 169         bool cleaner:1;
 170
 171         unsigned writeback_all;
 172         struct workqueue_struct *writeback_wq;
 173         struct work_struct writeback_work;
 174         struct work_struct flush_work;
 175
 176         struct dm_io_client *dm_io;
 177
 178         raw_spinlock_t endio_list_lock;
 179         struct list_head endio_list;
 180         struct task_struct *endio_thread;
 181
 182         struct task_struct *flush_thread;
 183         struct bio_list flush_list;
 184
 185         struct dm_kcopyd_client *dm_kcopyd;
 186         unsigned long *dirty_bitmap;
 187         unsigned dirty_bitmap_size;
 188
 189         struct bio_set bio_set;
 190         mempool_t copy_pool;
 191 };
 192
 193 #define WB_LIST_INLINE          16
 194
 195 struct writeback_struct {
 196         struct list_head endio_entry;
 197         struct dm_writecache *wc;
 198         struct wc_entry **wc_list;
 199         unsigned wc_list_n;
 200         struct wc_entry *wc_list_inline[WB_LIST_INLINE];
 201         struct bio bio;
 202 };
 203
 204 struct copy_struct {
 205         struct list_head endio_entry;
 206         struct dm_writecache *wc;
 207         struct wc_entry *e;
 208         unsigned n_entries;
 209         int error;
 210 };
 211
 212 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(dm_writecache_throttle,
 213                                             "A percentage of time allocated for data copying");
 214
 215 static void wc_lock(struct dm_writecache *wc)
 216 {
 217         mutex_lock(&wc->lock);
 218 }
 219
 220 static void wc_unlock(struct dm_writecache *wc)
 221 {
 222         mutex_unlock(&wc->lock);
 223 }
 224
 225 #ifdef DM_WRITECACHE_HAS_PMEM
 226 static int persistent_memory_claim(struct dm_writecache *wc)
 227 {
 228         int r;
 229         loff_t s;
 230         long p, da;
 231         pfn_t pfn;
 232         int id;
 233         struct page **pages;
 234
 235         wc->memory_vmapped = false;
 236
 237         if (!wc->ssd_dev->dax_dev) {
 238                 r = -EOPNOTSUPP;
 239                 goto err1;
 240         }
 241         s = wc->memory_map_size;
 242         p = s >> PAGE_SHIFT;
 243         if (!p) {
 244                 r = -EINVAL;
 245                 goto err1;
 246         }
 247         if (p != s >> PAGE_SHIFT) {
 248                 r = -EOVERFLOW;
 249                 goto err1;
 250         }
 251
 252         id = dax_read_lock();
 253
 254         da = dax_direct_access(wc->ssd_dev->dax_dev, 0, p, &wc->memory_map, &pfn);
 255         if (da < 0) {
 256                 wc->memory_map = NULL;
 257                 r = da;
 258                 goto err2;
 259         }
 260         if (!pfn_t_has_page(pfn)) {
 261                 wc->memory_map = NULL;
 262                 r = -EOPNOTSUPP;
 263                 goto err2;
 264         }
 265         if (da != p) {
 266                 long i;
 267                 wc->memory_map = NULL;
 268                 pages = kvmalloc_array(p, sizeof(struct page *), GFP_KERNEL);
 269                 if (!pages) {
 270                         r = -ENOMEM;
 271                         goto err2;
 272                 }
 273                 i = 0;
 274                 do {
 275                         long daa;
 276                         daa = dax_direct_access(wc->ssd_dev->dax_dev, i, p - i,
 277                                                 NULL, &pfn);
 278                         if (daa <= 0) {
 279                                 r = daa ? daa : -EINVAL;
 280                                 goto err3;
 281                         }
 282                         if (!pfn_t_has_page(pfn)) {
 283                                 r = -EOPNOTSUPP;
 284                                 goto err3;
 285                         }
 286                         while (daa-- && i < p) {
 287                                 pages[i++] = pfn_t_to_page(pfn);
 288                                 pfn.val++;
 289                         }
 290                 } while (i < p);
 291                 wc->memory_map = vmap(pages, p, VM_MAP, PAGE_KERNEL);
 292                 if (!wc->memory_map) {
 293                         r = -ENOMEM;
 294                         goto err3;
 295                 }
 296                 kvfree(pages);
 297                 wc->memory_vmapped = true;
 298         }
 299
 300         dax_read_unlock(id);
 301
 302         wc->memory_map += (size_t)wc->start_sector << SECTOR_SHIFT;
 303         wc->memory_map_size -= (size_t)wc->start_sector << SECTOR_SHIFT;
 304
 305         return 0;
 306 err3:
 307         kvfree(pages);
 308 err2:
 309         dax_read_unlock(id);
 310 err1:
 311         return r;
 312 }
 313 #else
 314 static int persistent_memory_claim(struct dm_writecache *wc)
 315 {
 316         BUG();
 317 }
 318 #endif
 319
 320 static void persistent_memory_release(struct dm_writecache *wc)
 321 {
 322         if (wc->memory_vmapped)
 323                 vunmap(wc->memory_map - ((size_t)wc->start_sector << SECTOR_SHIFT));
 324 }
 325
 326 static struct page *persistent_memory_page(void *addr)
 327 {
 328         if (is_vmalloc_addr(addr))
 329                 return vmalloc_to_page(addr);
 330         else
 331                 return virt_to_page(addr);
 332 }
 333
 334 static unsigned persistent_memory_page_offset(void *addr)
 335 {
 336         return (unsigned long)addr & (PAGE_SIZE - 1);
 337 }
 338
 339 static void persistent_memory_flush_cache(void *ptr, size_t size)
 340 {
 341         if (is_vmalloc_addr(ptr))
 342                 flush_kernel_vmap_range(ptr, size);
 343 }
 344
 345 static void persistent_memory_invalidate_cache(void *ptr, size_t size)
 346 {
 347         if (is_vmalloc_addr(ptr))
 348                 invalidate_kernel_vmap_range(ptr, size);
 349 }
 350
 351 static struct wc_memory_superblock *sb(struct dm_writecache *wc)
 352 {
 353         return wc->memory_map;
 354 }
 355
 356 static struct wc_memory_entry *memory_entry(struct dm_writecache *wc, struct wc_entry *e)
 357 {
 358         return &sb(wc)->entries[e->index];
 359 }
 360
 361 static void *memory_data(struct dm_writecache *wc, struct wc_entry *e)
 362 {
 363         return (char *)wc->block_start + (e->index << wc->block_size_bits);
 364 }
 365
 366 static sector_t cache_sector(struct dm_writecache *wc, struct wc_entry *e)
 367 {
 368         return wc->start_sector + wc->metadata_sectors +
 369                 ((sector_t)e->index << (wc->block_size_bits - SECTOR_SHIFT));
 370 }
 371
 372 static uint64_t read_original_sector(struct dm_writecache *wc, struct wc_entry *e)
 373 {
 374 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 375         return e->original_sector;
 376 #else
 377         return le64_to_cpu(memory_entry(wc, e)->original_sector);
 378 #endif
 379 }
 380
 381 static uint64_t read_seq_count(struct dm_writecache *wc, struct wc_entry *e)
 382 {
 383 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 384         return e->seq_count;
 385 #else
 386         return le64_to_cpu(memory_entry(wc, e)->seq_count);
 387 #endif
 388 }
 389
 390 static void clear_seq_count(struct dm_writecache *wc, struct wc_entry *e)
 391 {
 392 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 393         e->seq_count = -1;
 394 #endif
 395         pmem_assign(memory_entry(wc, e)->seq_count, cpu_to_le64(-1));
 396 }
 397
 398 static void write_original_sector_seq_count(struct dm_writecache *wc, struct wc_entry *e,
 399                                             uint64_t original_sector, uint64_t seq_count)
 400 {
 401         struct wc_memory_entry me;
 402 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 403         e->original_sector = original_sector;
 404         e->seq_count = seq_count;
 405 #endif
 406         me.original_sector = cpu_to_le64(original_sector);
 407         me.seq_count = cpu_to_le64(seq_count);
 408         pmem_assign(*memory_entry(wc, e), me);
 409 }
 410
 411 #define writecache_error(wc, err, msg, arg...)                          \
 412 do {                                                                    \
 413         if (!cmpxchg(&(wc)->error, 0, err))                             \
 414                 DMERR(msg, ##arg);                                      \
 415         wake_up(&(wc)->freelist_wait);                                  \
 416 } while (0)
 417
 418 #define writecache_has_error(wc)        (unlikely(READ_ONCE((wc)->error)))
 419
 420 static void writecache_flush_all_metadata(struct dm_writecache *wc)
 421 {
 422         if (!WC_MODE_PMEM(wc))
 423                 memset(wc->dirty_bitmap, -1, wc->dirty_bitmap_size);
 424 }
 425
 426 static void writecache_flush_region(struct dm_writecache *wc, void *ptr, size_t size)
 427 {
 428         if (!WC_MODE_PMEM(wc))
 429                 __set_bit(((char *)ptr - (char *)wc->memory_map) / BITMAP_GRANULARITY,
 430                           wc->dirty_bitmap);
 431 }
 432
 433 static void writecache_disk_flush(struct dm_writecache *wc, struct dm_dev *dev);
 434
 435 struct io_notify {
 436         struct dm_writecache *wc;
 437         struct completion c;
 438         atomic_t count;
 439 };
 440
 441 static void writecache_notify_io(unsigned long error, void *context)
 442 {
 443         struct io_notify *endio = context;
 444
 445         if (unlikely(error != 0))
 446                 writecache_error(endio->wc, -EIO, "error writing metadata");
 447         BUG_ON(atomic_read(&endio->count) <= 0);
 448         if (atomic_dec_and_test(&endio->count))
 449                 complete(&endio->c);
 450 }
 451
 452 static void writecache_wait_for_ios(struct dm_writecache *wc, int direction)
 453 {
 454         wait_event(wc->bio_in_progress_wait[direction],
 455                    !atomic_read(&wc->bio_in_progress[direction]));
 456 }
 457
 458 static void ssd_commit_flushed(struct dm_writecache *wc, bool wait_for_ios)
 459 {
 460         struct dm_io_region region;
 461         struct dm_io_request req;
 462         struct io_notify endio = {
 463                 wc,
 464                 COMPLETION_INITIALIZER_ONSTACK(endio.c),
 465                 ATOMIC_INIT(1),
 466         };
 467         unsigned bitmap_bits = wc->dirty_bitmap_size * 8;
 468         unsigned i = 0;
 469
 470         while (1) {
 471                 unsigned j;
 472                 i = find_next_bit(wc->dirty_bitmap, bitmap_bits, i);
 473                 if (unlikely(i == bitmap_bits))
 474                         break;
 475                 j = find_next_zero_bit(wc->dirty_bitmap, bitmap_bits, i);
 476
 477                 region.bdev = wc->ssd_dev->bdev;
 478                 region.sector = (sector_t)i * (BITMAP_GRANULARITY >> SECTOR_SHIFT);
 479                 region.count = (sector_t)(j - i) * (BITMAP_GRANULARITY >> SECTOR_SHIFT);
 480
 481                 if (unlikely(region.sector >= wc->metadata_sectors))
 482                         break;
 483                 if (unlikely(region.sector + region.count > wc->metadata_sectors))
 484                         region.count = wc->metadata_sectors - region.sector;
 485
 486                 region.sector += wc->start_sector;
 487                 atomic_inc(&endio.count);
 488                 req.bi_op = REQ_OP_WRITE;
 489                 req.bi_op_flags = REQ_SYNC;
 490                 req.mem.type = DM_IO_VMA;
 491                 req.mem.ptr.vma = (char *)wc->memory_map + (size_t)i * BITMAP_GRANULARITY;
 492                 req.client = wc->dm_io;
 493                 req.notify.fn = writecache_notify_io;
 494                 req.notify.context = &endio;
 495
 496                 /* writing via async dm-io (implied by notify.fn above) won't return an error */
 497                 (void) dm_io(&req, 1, &region, NULL);
 498                 i = j;
 499         }
 500
 501         writecache_notify_io(0, &endio);
 502         wait_for_completion_io(&endio.c);
 503
 504         if (wait_for_ios)
 505                 writecache_wait_for_ios(wc, WRITE);
 506
 507         writecache_disk_flush(wc, wc->ssd_dev);
 508
 509         memset(wc->dirty_bitmap, 0, wc->dirty_bitmap_size);
 510 }
 511
 512 static void ssd_commit_superblock(struct dm_writecache *wc)
 513 {
 514         int r;
 515         struct dm_io_region region;
 516         struct dm_io_request req;
 517
 518         region.bdev = wc->ssd_dev->bdev;
 519         region.sector = 0;
 520         region.count = PAGE_SIZE;
 521
 522         if (unlikely(region.sector + region.count > wc->metadata_sectors))
 523                 region.count = wc->metadata_sectors - region.sector;
 524
 525         region.sector += wc->start_sector;
 526
 527         req.bi_op = REQ_OP_WRITE;
 528         req.bi_op_flags = REQ_SYNC | REQ_FUA;
 529         req.mem.type = DM_IO_VMA;
 530         req.mem.ptr.vma = (char *)wc->memory_map;
 531         req.client = wc->dm_io;
 532         req.notify.fn = NULL;
 533         req.notify.context = NULL;
 534
 535         r = dm_io(&req, 1, &region, NULL);
 536         if (unlikely(r))
 537                 writecache_error(wc, r, "error writing superblock");
 538 }
 539
 540 static void writecache_commit_flushed(struct dm_writecache *wc, bool wait_for_ios)
 541 {
 542         if (WC_MODE_PMEM(wc))
 543                 wmb();
 544         else
 545                 ssd_commit_flushed(wc, wait_for_ios);
 546 }
 547
 548 static void writecache_disk_flush(struct dm_writecache *wc, struct dm_dev *dev)
 549 {
 550         int r;
 551         struct dm_io_region region;
 552         struct dm_io_request req;
 553
 554         region.bdev = dev->bdev;
 555         region.sector = 0;
 556         region.count = 0;
 557         req.bi_op = REQ_OP_WRITE;
 558         req.bi_op_flags = REQ_PREFLUSH;
 559         req.mem.type = DM_IO_KMEM;
 560         req.mem.ptr.addr = NULL;
 561         req.client = wc->dm_io;
 562         req.notify.fn = NULL;
 563
 564         r = dm_io(&req, 1, &region, NULL);
 565         if (unlikely(r))
 566                 writecache_error(wc, r, "error flushing metadata: %d", r);
 567 }
 568
 569 #define WFE_RETURN_FOLLOWING    1
 570 #define WFE_LOWEST_SEQ          2
 571
 572 static struct wc_entry *writecache_find_entry(struct dm_writecache *wc,
 573                                               uint64_t block, int flags)
 574 {
 575         struct wc_entry *e;
 576         struct rb_node *node = wc->tree.rb_node;
 577
 578         if (unlikely(!node))
 579                 return NULL;
 580
 581         while (1) {
 582                 e = container_of(node, struct wc_entry, rb_node);
 583                 if (read_original_sector(wc, e) == block)
 584                         break;
 585
 586                 node = (read_original_sector(wc, e) >= block ?
 587                         e->rb_node.rb_left : e->rb_node.rb_right);
 588                 if (unlikely(!node)) {
 589                         if (!(flags & WFE_RETURN_FOLLOWING))
 590                                 return NULL;
 591                         if (read_original_sector(wc, e) >= block) {
 592                                 return e;
 593                         } else {
 594                                 node = rb_next(&e->rb_node);
 595                                 if (unlikely(!node))
 596                                         return NULL;
 597                                 e = container_of(node, struct wc_entry, rb_node);
 598                                 return e;
 599                         }
 600                 }
 601         }
 602
 603         while (1) {
 604                 struct wc_entry *e2;
 605                 if (flags & WFE_LOWEST_SEQ)
 606                         node = rb_prev(&e->rb_node);
 607                 else
 608                         node = rb_next(&e->rb_node);
 609                 if (unlikely(!node))
 610                         return e;
 611                 e2 = container_of(node, struct wc_entry, rb_node);
 612                 if (read_original_sector(wc, e2) != block)
 613                         return e;
 614                 e = e2;
 615         }
 616 }
 617
 618 static void writecache_insert_entry(struct dm_writecache *wc, struct wc_entry *ins)
 619 {
 620         struct wc_entry *e;
 621         struct rb_node **node = &wc->tree.rb_node, *parent = NULL;
 622
 623         while (*node) {
 624                 e = container_of(*node, struct wc_entry, rb_node);
 625                 parent = &e->rb_node;
 626                 if (read_original_sector(wc, e) > read_original_sector(wc, ins))
 627                         node = &parent->rb_left;
 628                 else
 629                         node = &parent->rb_right;
 630         }
 631         rb_link_node(&ins->rb_node, parent, node);
 632         rb_insert_color(&ins->rb_node, &wc->tree);
 633         list_add(&ins->lru, &wc->lru);
 634         ins->age = jiffies;
 635 }
 636
 637 static void writecache_unlink(struct dm_writecache *wc, struct wc_entry *e)
 638 {
 639         list_del(&e->lru);
 640         rb_erase(&e->rb_node, &wc->tree);
 641 }
 642
 643 static void writecache_add_to_freelist(struct dm_writecache *wc, struct wc_entry *e)
 644 {
 645         if (WC_MODE_SORT_FREELIST(wc)) {
 646                 struct rb_node **node = &wc->freetree.rb_node, *parent = NULL;
 647                 if (unlikely(!*node))
 648                         wc->current_free = e;
 649                 while (*node) {
 650                         parent = *node;
 651                         if (&e->rb_node < *node)
 652                                 node = &parent->rb_left;
 653                         else
 654                                 node = &parent->rb_right;
 655                 }
 656                 rb_link_node(&e->rb_node, parent, node);
 657                 rb_insert_color(&e->rb_node, &wc->freetree);
 658         } else {
 659                 list_add_tail(&e->lru, &wc->freelist);
 660         }
 661         wc->freelist_size++;
 662 }
 663
 664 static inline void writecache_verify_watermark(struct dm_writecache *wc)
 665 {
 666         if (unlikely(wc->freelist_size + wc->writeback_size <= wc->freelist_high_watermark))
 667                 queue_work(wc->writeback_wq, &wc->writeback_work);
 668 }
 669
 670 static void writecache_max_age_timer(struct timer_list *t)
 671 {
 672         struct dm_writecache *wc = from_timer(wc, t, max_age_timer);
 673
 674         if (!dm_suspended(wc->ti) && !writecache_has_error(wc)) {
 675                 queue_work(wc->writeback_wq, &wc->writeback_work);
 676                 mod_timer(&wc->max_age_timer, jiffies + wc->max_age / MAX_AGE_DIV);
 677         }
 678 }
 679
 680 static struct wc_entry *writecache_pop_from_freelist(struct dm_writecache *wc, sector_t expected_sector)
 681 {
 682         struct wc_entry *e;
 683
 684         if (WC_MODE_SORT_FREELIST(wc)) {
 685                 struct rb_node *next;
 686                 if (unlikely(!wc->current_free))
 687                         return NULL;
 688                 e = wc->current_free;
 689                 if (expected_sector != (sector_t)-1 && unlikely(cache_sector(wc, e) != expected_sector))
 690                         return NULL;
 691                 next = rb_next(&e->rb_node);
 692                 rb_erase(&e->rb_node, &wc->freetree);
 693                 if (unlikely(!next))
 694                         next = rb_first(&wc->freetree);
 695                 wc->current_free = next ? container_of(next, struct wc_entry, rb_node) : NULL;
 696         } else {
 697                 if (unlikely(list_empty(&wc->freelist)))
 698                         return NULL;
 699                 e = container_of(wc->freelist.next, struct wc_entry, lru);
 700                 if (expected_sector != (sector_t)-1 && unlikely(cache_sector(wc, e) != expected_sector))
 701                         return NULL;
 702                 list_del(&e->lru);
 703         }
 704         wc->freelist_size--;
 705
 706         writecache_verify_watermark(wc);
 707
 708         return e;
 709 }
 710
 711 static void writecache_free_entry(struct dm_writecache *wc, struct wc_entry *e)
 712 {
 713         writecache_unlink(wc, e);
 714         writecache_add_to_freelist(wc, e);
 715         clear_seq_count(wc, e);
 716         writecache_flush_region(wc, memory_entry(wc, e), sizeof(struct wc_memory_entry));
 717         if (unlikely(waitqueue_active(&wc->freelist_wait)))
 718                 wake_up(&wc->freelist_wait);
 719 }
 720
 721 static void writecache_wait_on_freelist(struct dm_writecache *wc)
 722 {
 723         DEFINE_WAIT(wait);
 724
 725         prepare_to_wait(&wc->freelist_wait, &wait, TASK_UNINTERRUPTIBLE);
 726         wc_unlock(wc);
 727         io_schedule();
 728         finish_wait(&wc->freelist_wait, &wait);
 729         wc_lock(wc);
 730 }
 731
 732 static void writecache_poison_lists(struct dm_writecache *wc)
 733 {
 734         /*
 735          * Catch incorrect access to these values while the device is suspended.
 736          */
 737         memset(&wc->tree, -1, sizeof wc->tree);
 738         wc->lru.next = LIST_POISON1;
 739         wc->lru.prev = LIST_POISON2;
 740         wc->freelist.next = LIST_POISON1;
 741         wc->freelist.prev = LIST_POISON2;
 742 }
 743
 744 static void writecache_flush_entry(struct dm_writecache *wc, struct wc_entry *e)
 745 {
 746         writecache_flush_region(wc, memory_entry(wc, e), sizeof(struct wc_memory_entry));
 747         if (WC_MODE_PMEM(wc))
 748                 writecache_flush_region(wc, memory_data(wc, e), wc->block_size);
 749 }
 750
 751 static bool writecache_entry_is_committed(struct dm_writecache *wc, struct wc_entry *e)
 752 {
 753         return read_seq_count(wc, e) < wc->seq_count;
 754 }
 755
 756 static void writecache_flush(struct dm_writecache *wc)
 757 {
 758         struct wc_entry *e, *e2;
 759         bool need_flush_after_free;
 760
 761         wc->uncommitted_blocks = 0;
 762         del_timer(&wc->autocommit_timer);
 763
 764         if (list_empty(&wc->lru))
 765                 return;
 766
 767         e = container_of(wc->lru.next, struct wc_entry, lru);
 768         if (writecache_entry_is_committed(wc, e)) {
 769                 if (wc->overwrote_committed) {
 770                         writecache_wait_for_ios(wc, WRITE);
 771                         writecache_disk_flush(wc, wc->ssd_dev);
 772                         wc->overwrote_committed = false;
 773                 }
 774                 return;
 775         }
 776         while (1) {
 777                 writecache_flush_entry(wc, e);
 778                 if (unlikely(e->lru.next == &wc->lru))
 779                         break;
 780                 e2 = container_of(e->lru.next, struct wc_entry, lru);
 781                 if (writecache_entry_is_committed(wc, e2))
 782                         break;
 783                 e = e2;
 784                 cond_resched();
 785         }
 786         writecache_commit_flushed(wc, true);
 787
 788         wc->seq_count++;
 789         pmem_assign(sb(wc)->seq_count, cpu_to_le64(wc->seq_count));
 790         if (WC_MODE_PMEM(wc))
 791                 writecache_commit_flushed(wc, false);
 792         else
 793                 ssd_commit_superblock(wc);
 794
 795         wc->overwrote_committed = false;
 796
 797         need_flush_after_free = false;
 798         while (1) {
 799                 /* Free another committed entry with lower seq-count */
 800                 struct rb_node *rb_node = rb_prev(&e->rb_node);
 801
 802                 if (rb_node) {
 803                         e2 = container_of(rb_node, struct wc_entry, rb_node);
 804                         if (read_original_sector(wc, e2) == read_original_sector(wc, e) &&
 805                             likely(!e2->write_in_progress)) {
 806                                 writecache_free_entry(wc, e2);
 807                                 need_flush_after_free = true;
 808                         }
 809                 }
 810                 if (unlikely(e->lru.prev == &wc->lru))
 811                         break;
 812                 e = container_of(e->lru.prev, struct wc_entry, lru);
 813                 cond_resched();
 814         }
 815
 816         if (need_flush_after_free)
 817                 writecache_commit_flushed(wc, false);
 818 }
 819
 820 static void writecache_flush_work(struct work_struct *work)
 821 {
 822         struct dm_writecache *wc = container_of(work, struct dm_writecache, flush_work);
 823
 824         wc_lock(wc);
 825         writecache_flush(wc);
 826         wc_unlock(wc);
 827 }
 828
 829 static void writecache_autocommit_timer(struct timer_list *t)
 830 {
 831         struct dm_writecache *wc = from_timer(wc, t, autocommit_timer);
 832         if (!writecache_has_error(wc))
 833                 queue_work(wc->writeback_wq, &wc->flush_work);
 834 }
 835
 836 static void writecache_schedule_autocommit(struct dm_writecache *wc)
 837 {
 838         if (!timer_pending(&wc->autocommit_timer))
 839                 mod_timer(&wc->autocommit_timer, jiffies + wc->autocommit_jiffies);
 840 }
 841
 842 static void writecache_discard(struct dm_writecache *wc, sector_t start, sector_t end)
 843 {
 844         struct wc_entry *e;
 845         bool discarded_something = false;
 846
 847         e = writecache_find_entry(wc, start, WFE_RETURN_FOLLOWING | WFE_LOWEST_SEQ);
 848         if (unlikely(!e))
 849                 return;
 850
 851         while (read_original_sector(wc, e) < end) {
 852                 struct rb_node *node = rb_next(&e->rb_node);
 853
 854                 if (likely(!e->write_in_progress)) {
 855                         if (!discarded_something) {
 856                                 writecache_wait_for_ios(wc, READ);
 857                                 writecache_wait_for_ios(wc, WRITE);
 858                                 discarded_something = true;
 859                         }
 860                         writecache_free_entry(wc, e);
 861                 }
 862
 863                 if (unlikely(!node))
 864                         break;
 865
 866                 e = container_of(node, struct wc_entry, rb_node);
 867         }
 868
 869         if (discarded_something)
 870                 writecache_commit_flushed(wc, false);
 871 }
 872
 873 static bool writecache_wait_for_writeback(struct dm_writecache *wc)
 874 {
 875         if (wc->writeback_size) {
 876                 writecache_wait_on_freelist(wc);
 877                 return true;
 878         }
 879         return false;
 880 }
 881
 882 static void writecache_suspend(struct dm_target *ti)
 883 {
 884         struct dm_writecache *wc = ti->private;
 885         bool flush_on_suspend;
 886
 887         del_timer_sync(&wc->autocommit_timer);
 888         del_timer_sync(&wc->max_age_timer);
 889
 890         wc_lock(wc);
 891         writecache_flush(wc);
 892         flush_on_suspend = wc->flush_on_suspend;
 893         if (flush_on_suspend) {
 894                 wc->flush_on_suspend = false;
 895                 wc->writeback_all++;
 896                 queue_work(wc->writeback_wq, &wc->writeback_work);
 897         }
 898         wc_unlock(wc);
 899
 900         drain_workqueue(wc->writeback_wq);
 901
 902         wc_lock(wc);
 903         if (flush_on_suspend)
 904                 wc->writeback_all--;
 905         while (writecache_wait_for_writeback(wc));
 906
 907         if (WC_MODE_PMEM(wc))
 908                 persistent_memory_flush_cache(wc->memory_map, wc->memory_map_size);
 909
 910         writecache_poison_lists(wc);
 911
 912         wc_unlock(wc);
 913 }
 914
 915 static int writecache_alloc_entries(struct dm_writecache *wc)
 916 {
 917         size_t b;
 918
 919         if (wc->entries)
 920                 return 0;
 921         wc->entries = vmalloc(array_size(sizeof(struct wc_entry), wc->n_blocks));
 922         if (!wc->entries)
 923                 return -ENOMEM;
 924         for (b = 0; b < wc->n_blocks; b++) {
 925                 struct wc_entry *e = &wc->entries[b];
 926                 e->index = b;
 927                 e->write_in_progress = false;
 928                 cond_resched();
 929         }
 930
 931         return 0;
 932 }
 933
 934 static int writecache_read_metadata(struct dm_writecache *wc, sector_t n_sectors)
 935 {
 936         struct dm_io_region region;
 937         struct dm_io_request req;
 938
 939         region.bdev = wc->ssd_dev->bdev;
 940         region.sector = wc->start_sector;
 941         region.count = n_sectors;
 942         req.bi_op = REQ_OP_READ;
 943         req.bi_op_flags = REQ_SYNC;
 944         req.mem.type = DM_IO_VMA;
 945         req.mem.ptr.vma = (char *)wc->memory_map;
 946         req.client = wc->dm_io;
 947         req.notify.fn = NULL;
 948
 949         return dm_io(&req, 1, &region, NULL);
 950 }
 951
 952 static void writecache_resume(struct dm_target *ti)
 953 {
 954         struct dm_writecache *wc = ti->private;
 955         size_t b;
 956         bool need_flush = false;
 957         __le64 sb_seq_count;
 958         int r;
 959
 960         wc_lock(wc);
 961
 962         if (WC_MODE_PMEM(wc)) {
 963                 persistent_memory_invalidate_cache(wc->memory_map, wc->memory_map_size);
 964         } else {
 965                 r = writecache_read_metadata(wc, wc->metadata_sectors);
 966                 if (r) {
 967                         size_t sb_entries_offset;
 968                         writecache_error(wc, r, "unable to read metadata: %d", r);
 969                         sb_entries_offset = offsetof(struct wc_memory_superblock, entries);
 970                         memset((char *)wc->memory_map + sb_entries_offset, -1,
 971                                (wc->metadata_sectors << SECTOR_SHIFT) - sb_entries_offset);
 972                 }
 973         }
 974
 975         wc->tree = RB_ROOT;
 976         INIT_LIST_HEAD(&wc->lru);
 977         if (WC_MODE_SORT_FREELIST(wc)) {
 978                 wc->freetree = RB_ROOT;
 979                 wc->current_free = NULL;
 980         } else {
 981                 INIT_LIST_HEAD(&wc->freelist);
 982         }
 983         wc->freelist_size = 0;
 984
 985         r = memcpy_mcsafe(&sb_seq_count, &sb(wc)->seq_count, sizeof(uint64_t));
 986         if (r) {
 987                 writecache_error(wc, r, "hardware memory error when reading superblock: %d", r);
 988                 sb_seq_count = cpu_to_le64(0);
 989         }
 990         wc->seq_count = le64_to_cpu(sb_seq_count);
 991
 992 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 993         for (b = 0; b < wc->n_blocks; b++) {
 994                 struct wc_entry *e = &wc->entries[b];
 995                 struct wc_memory_entry wme;
 996                 if (writecache_has_error(wc)) {
 997                         e->original_sector = -1;
 998                         e->seq_count = -1;
 999                         continue;
1000                 }
1001                 r = memcpy_mcsafe(&wme, memory_entry(wc, e), sizeof(struct wc_memory_entry));
1002                 if (r) {
1003                         writecache_error(wc, r, "hardware memory error when reading metadata entry %lu: %d",
1004                                          (unsigned long)b, r);
1005                         e->original_sector = -1;
1006                         e->seq_count = -1;
1007                 } else {
1008                         e->original_sector = le64_to_cpu(wme.original_sector);
1009                         e->seq_count = le64_to_cpu(wme.seq_count);
1010                 }
1011                 cond_resched();
1012         }
1013 #endif
1014         for (b = 0; b < wc->n_blocks; b++) {
1015                 struct wc_entry *e = &wc->entries[b];
1016                 if (!writecache_entry_is_committed(wc, e)) {
1017                         if (read_seq_count(wc, e) != -1) {
1018 erase_this:
1019                                 clear_seq_count(wc, e);
1020                                 need_flush = true;
1021                         }
1022                         writecache_add_to_freelist(wc, e);
1023                 } else {
1024                         struct wc_entry *old;
1025
1026                         old = writecache_find_entry(wc, read_original_sector(wc, e), 0);
1027                         if (!old) {
1028                                 writecache_insert_entry(wc, e);
1029                         } else {
1030                                 if (read_seq_count(wc, old) == read_seq_count(wc, e)) {
1031                                         writecache_error(wc, -EINVAL,
1032                                                  "two identical entries, position %llu, sector %llu, sequence %llu",
1033                                                  (unsigned long long)b, (unsigned long long)read_original_sector(wc, e),
1034                                                  (unsigned long long)read_seq_count(wc, e));
1035                                 }
1036                                 if (read_seq_count(wc, old) > read_seq_count(wc, e)) {
1037                                         goto erase_this;
1038                                 } else {
1039                                         writecache_free_entry(wc, old);
1040                                         writecache_insert_entry(wc, e);
1041                                         need_flush = true;
1042                                 }
1043                         }
1044                 }
1045                 cond_resched();
1046         }
1047
1048         if (need_flush) {
1049                 writecache_flush_all_metadata(wc);
1050                 writecache_commit_flushed(wc, false);
1051         }
1052
1053         writecache_verify_watermark(wc);
1054
1055         if (wc->max_age != MAX_AGE_UNSPECIFIED)
1056                 mod_timer(&wc->max_age_timer, jiffies + wc->max_age / MAX_AGE_DIV);
1057
1058         wc_unlock(wc);
1059 }
1060
1061 static int process_flush_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1062 {
1063         if (argc != 1)
1064                 return -EINVAL;
1065
1066         wc_lock(wc);
1067         if (dm_suspended(wc->ti)) {
1068                 wc_unlock(wc);
1069                 return -EBUSY;
1070         }
1071         if (writecache_has_error(wc)) {
1072                 wc_unlock(wc);
1073                 return -EIO;
1074         }
1075
1076         writecache_flush(wc);
1077         wc->writeback_all++;
1078         queue_work(wc->writeback_wq, &wc->writeback_work);
1079         wc_unlock(wc);
1080
1081         flush_workqueue(wc->writeback_wq);
1082
1083         wc_lock(wc);
1084         wc->writeback_all--;
1085         if (writecache_has_error(wc)) {
1086                 wc_unlock(wc);
1087                 return -EIO;
1088         }
1089         wc_unlock(wc);
1090
1091         return 0;
1092 }
1093
1094 static int process_flush_on_suspend_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1095 {
1096         if (argc != 1)
1097                 return -EINVAL;
1098
1099         wc_lock(wc);
1100         wc->flush_on_suspend = true;
1101         wc_unlock(wc);
1102
1103         return 0;
1104 }
1105
1106 static void activate_cleaner(struct dm_writecache *wc)
1107 {
1108         wc->flush_on_suspend = true;
1109         wc->cleaner = true;
1110         wc->freelist_high_watermark = wc->n_blocks;
1111         wc->freelist_low_watermark = wc->n_blocks;
1112 }
1113
1114 static int process_cleaner_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1115 {
1116         if (argc != 1)
1117                 return -EINVAL;
1118
1119         wc_lock(wc);
1120         activate_cleaner(wc);
1121         if (!dm_suspended(wc->ti))
1122                 writecache_verify_watermark(wc);
1123         wc_unlock(wc);
1124
1125         return 0;
1126 }
1127
1128 static int writecache_message(struct dm_target *ti, unsigned argc, char **argv,
1129                               char *result, unsigned maxlen)
1130 {
1131         int r = -EINVAL;
1132         struct dm_writecache *wc = ti->private;
1133
1134         if (!strcasecmp(argv[0], "flush"))
1135                 r = process_flush_mesg(argc, argv, wc);
1136         else if (!strcasecmp(argv[0], "flush_on_suspend"))
1137                 r = process_flush_on_suspend_mesg(argc, argv, wc);
1138         else if (!strcasecmp(argv[0], "cleaner"))
1139                 r = process_cleaner_mesg(argc, argv, wc);
1140         else
1141                 DMERR("unrecognised message received: %s", argv[0]);
1142
1143         return r;
1144 }
1145
1146 static void bio_copy_block(struct dm_writecache *wc, struct bio *bio, void *data)
1147 {
1148         void *buf;
1149         unsigned long flags;
1150         unsigned size;
1151         int rw = bio_data_dir(bio);
1152         unsigned remaining_size = wc->block_size;
1153
1154         do {
1155                 struct bio_vec bv = bio_iter_iovec(bio, bio->bi_iter);
1156                 buf = bvec_kmap_irq(&bv, &flags);
1157                 size = bv.bv_len;
1158                 if (unlikely(size > remaining_size))
1159                         size = remaining_size;
1160
1161                 if (rw == READ) {
1162                         int r;
1163                         r = memcpy_mcsafe(buf, data, size);
1164                         flush_dcache_page(bio_page(bio));
1165                         if (unlikely(r)) {
1166                                 writecache_error(wc, r, "hardware memory error when reading data: %d", r);
1167                                 bio->bi_status = BLK_STS_IOERR;
1168                         }
1169                 } else {
1170                         flush_dcache_page(bio_page(bio));
1171                         memcpy_flushcache(data, buf, size);
1172                 }
1173
1174                 bvec_kunmap_irq(buf, &flags);
1175
1176                 data = (char *)data + size;
1177                 remaining_size -= size;
1178                 bio_advance(bio, size);
1179         } while (unlikely(remaining_size));
1180 }
1181
1182 static int writecache_flush_thread(void *data)
1183 {
1184         struct dm_writecache *wc = data;
1185
1186         while (1) {
1187                 struct bio *bio;
1188
1189                 wc_lock(wc);
1190                 bio = bio_list_pop(&wc->flush_list);
1191                 if (!bio) {
1192                         set_current_state(TASK_INTERRUPTIBLE);
1193                         wc_unlock(wc);
1194
1195                         if (unlikely(kthread_should_stop())) {
1196                                 set_current_state(TASK_RUNNING);
1197                                 break;
1198                         }
1199
1200                         schedule();
1201                         continue;
1202                 }
1203
1204                 if (bio_op(bio) == REQ_OP_DISCARD) {
1205                         writecache_discard(wc, bio->bi_iter.bi_sector,
1206                                            bio_end_sector(bio));
1207                         wc_unlock(wc);
1208                         bio_set_dev(bio, wc->dev->bdev);
1209                         generic_make_request(bio);
1210                 } else {
1211                         writecache_flush(wc);
1212                         wc_unlock(wc);
1213                         if (writecache_has_error(wc))
1214                                 bio->bi_status = BLK_STS_IOERR;
1215                         bio_endio(bio);
1216                 }
1217         }
1218
1219         return 0;
1220 }
1221
1222 static void writecache_offload_bio(struct dm_writecache *wc, struct bio *bio)
1223 {
1224         if (bio_list_empty(&wc->flush_list))
1225                 wake_up_process(wc->flush_thread);
1226         bio_list_add(&wc->flush_list, bio);
1227 }
1228
1229 static int writecache_map(struct dm_target *ti, struct bio *bio)
1230 {
1231         struct wc_entry *e;
1232         struct dm_writecache *wc = ti->private;
1233
1234         bio->bi_private = NULL;
1235
1236         wc_lock(wc);
1237
1238         if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1239                 if (writecache_has_error(wc))
1240                         goto unlock_error;
1241                 if (WC_MODE_PMEM(wc)) {
1242                         writecache_flush(wc);
1243                         if (writecache_has_error(wc))
1244                                 goto unlock_error;
1245                         goto unlock_submit;
1246                 } else {
1247                         writecache_offload_bio(wc, bio);
1248                         goto unlock_return;
1249                 }
1250         }
1251
1252         bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
1253
1254         if (unlikely((((unsigned)bio->bi_iter.bi_sector | bio_sectors(bio)) &
1255                                 (wc->block_size / 512 - 1)) != 0)) {
1256                 DMERR("I/O is not aligned, sector %llu, size %u, block size %u",
1257                       (unsigned long long)bio->bi_iter.bi_sector,
1258                       bio->bi_iter.bi_size, wc->block_size);
1259                 goto unlock_error;
1260         }
1261
1262         if (unlikely(bio_op(bio) == REQ_OP_DISCARD)) {
1263                 if (writecache_has_error(wc))
1264                         goto unlock_error;
1265                 if (WC_MODE_PMEM(wc)) {
1266                         writecache_discard(wc, bio->bi_iter.bi_sector, bio_end_sector(bio));
1267                         goto unlock_remap_origin;
1268                 } else {
1269                         writecache_offload_bio(wc, bio);
1270                         goto unlock_return;
1271                 }
1272         }
1273
1274         if (bio_data_dir(bio) == READ) {
1275 read_next_block:
1276                 e = writecache_find_entry(wc, bio->bi_iter.bi_sector, WFE_RETURN_FOLLOWING);
1277                 if (e && read_original_sector(wc, e) == bio->bi_iter.bi_sector) {
1278                         if (WC_MODE_PMEM(wc)) {
1279                                 bio_copy_block(wc, bio, memory_data(wc, e));
1280                                 if (bio->bi_iter.bi_size)
1281                                         goto read_next_block;
1282                                 goto unlock_submit;
1283                         } else {
1284                                 dm_accept_partial_bio(bio, wc->block_size >> SECTOR_SHIFT);
1285                                 bio_set_dev(bio, wc->ssd_dev->bdev);
1286                                 bio->bi_iter.bi_sector = cache_sector(wc, e);
1287                                 if (!writecache_entry_is_committed(wc, e))
1288                                         writecache_wait_for_ios(wc, WRITE);
1289                                 goto unlock_remap;
1290                         }
1291                 } else {
1292                         if (e) {
1293                                 sector_t next_boundary =
1294                                         read_original_sector(wc, e) - bio->bi_iter.bi_sector;
1295                                 if (next_boundary < bio->bi_iter.bi_size >> SECTOR_SHIFT) {
1296                                         dm_accept_partial_bio(bio, next_boundary);
1297                                 }
1298                         }
1299                         goto unlock_remap_origin;
1300                 }
1301         } else {
1302                 do {
1303                         bool found_entry = false;
1304                         if (writecache_has_error(wc))
1305                                 goto unlock_error;
1306                         e = writecache_find_entry(wc, bio->bi_iter.bi_sector, 0);
1307                         if (e) {
1308                                 if (!writecache_entry_is_committed(wc, e))
1309                                         goto bio_copy;
1310                                 if (!WC_MODE_PMEM(wc) && !e->write_in_progress) {
1311                                         wc->overwrote_committed = true;
1312                                         goto bio_copy;
1313                                 }
1314                                 found_entry = true;
1315                         } else {
1316                                 if (unlikely(wc->cleaner))
1317                                         goto direct_write;
1318                         }
1319                         e = writecache_pop_from_freelist(wc, (sector_t)-1);
1320                         if (unlikely(!e)) {
1321                                 if (!found_entry) {
1322 direct_write:
1323                                         e = writecache_find_entry(wc, bio->bi_iter.bi_sector, WFE_RETURN_FOLLOWING);
1324                                         if (e) {
1325                                                 sector_t next_boundary = read_original_sector(wc, e) - bio->bi_iter.bi_sector;
1326                                                 BUG_ON(!next_boundary);
1327                                                 if (next_boundary < bio->bi_iter.bi_size >> SECTOR_SHIFT) {
1328                                                         dm_accept_partial_bio(bio, next_boundary);
1329                                                 }
1330                                         }
1331                                         goto unlock_remap_origin;
1332                                 }
1333                                 writecache_wait_on_freelist(wc);
1334                                 continue;
1335                         }
1336                         write_original_sector_seq_count(wc, e, bio->bi_iter.bi_sector, wc->seq_count);
1337                         writecache_insert_entry(wc, e);
1338                         wc->uncommitted_blocks++;
1339 bio_copy:
1340                         if (WC_MODE_PMEM(wc)) {
1341                                 bio_copy_block(wc, bio, memory_data(wc, e));
1342                         } else {
1343                                 unsigned bio_size = wc->block_size;
1344                                 sector_t start_cache_sec = cache_sector(wc, e);
1345                                 sector_t current_cache_sec = start_cache_sec + (bio_size >> SECTOR_SHIFT);
1346
1347                                 while (bio_size < bio->bi_iter.bi_size) {
1348                                         struct wc_entry *f = writecache_pop_from_freelist(wc, current_cache_sec);
1349                                         if (!f)
1350                                                 break;
1351                                         write_original_sector_seq_count(wc, f, bio->bi_iter.bi_sector +
1352                                                                         (bio_size >> SECTOR_SHIFT), wc->seq_count);
1353                                         writecache_insert_entry(wc, f);
1354                                         wc->uncommitted_blocks++;
1355                                         bio_size += wc->block_size;
1356                                         current_cache_sec += wc->block_size >> SECTOR_SHIFT;
1357                                 }
1358
1359                                 bio_set_dev(bio, wc->ssd_dev->bdev);
1360                                 bio->bi_iter.bi_sector = start_cache_sec;
1361                                 dm_accept_partial_bio(bio, bio_size >> SECTOR_SHIFT);
1362
1363                                 if (unlikely(wc->uncommitted_blocks >= wc->autocommit_blocks)) {
1364                                         wc->uncommitted_blocks = 0;
1365                                         queue_work(wc->writeback_wq, &wc->flush_work);
1366                                 } else {
1367                                         writecache_schedule_autocommit(wc);
1368                                 }
1369                                 goto unlock_remap;
1370                         }
1371                 } while (bio->bi_iter.bi_size);
1372
1373                 if (unlikely(bio->bi_opf & REQ_FUA ||
1374                              wc->uncommitted_blocks >= wc->autocommit_blocks))
1375                         writecache_flush(wc);
1376                 else
1377                         writecache_schedule_autocommit(wc);
1378                 goto unlock_submit;
1379         }
1380
1381 unlock_remap_origin:
1382         bio_set_dev(bio, wc->dev->bdev);
1383         wc_unlock(wc);
1384         return DM_MAPIO_REMAPPED;
1385
1386 unlock_remap:
1387         /* make sure that writecache_end_io decrements bio_in_progress: */
1388         bio->bi_private = (void *)1;
1389         atomic_inc(&wc->bio_in_progress[bio_data_dir(bio)]);
1390         wc_unlock(wc);
1391         return DM_MAPIO_REMAPPED;
1392
1393 unlock_submit:
1394         wc_unlock(wc);
1395         bio_endio(bio);
1396         return DM_MAPIO_SUBMITTED;
1397
1398 unlock_return:
1399         wc_unlock(wc);
1400         return DM_MAPIO_SUBMITTED;
1401
1402 unlock_error:
1403         wc_unlock(wc);
1404         bio_io_error(bio);
1405         return DM_MAPIO_SUBMITTED;
1406 }
1407
1408 static int writecache_end_io(struct dm_target *ti, struct bio *bio, blk_status_t *status)
1409 {
1410         struct dm_writecache *wc = ti->private;
1411
1412         if (bio->bi_private != NULL) {
1413                 int dir = bio_data_dir(bio);
1414                 if (atomic_dec_and_test(&wc->bio_in_progress[dir]))
1415                         if (unlikely(waitqueue_active(&wc->bio_in_progress_wait[dir])))
1416                                 wake_up(&wc->bio_in_progress_wait[dir]);
1417         }
1418         return 0;
1419 }
1420
1421 static int writecache_iterate_devices(struct dm_target *ti,
1422                                       iterate_devices_callout_fn fn, void *data)
1423 {
1424         struct dm_writecache *wc = ti->private;
1425
1426         return fn(ti, wc->dev, 0, ti->len, data);
1427 }
1428
1429 static void writecache_io_hints(struct dm_target *ti, struct queue_limits *limits)
1430 {
1431         struct dm_writecache *wc = ti->private;
1432
1433         if (limits->logical_block_size < wc->block_size)
1434                 limits->logical_block_size = wc->block_size;
1435
1436         if (limits->physical_block_size < wc->block_size)
1437                 limits->physical_block_size = wc->block_size;
1438
1439         if (limits->io_min < wc->block_size)
1440                 limits->io_min = wc->block_size;
1441 }
1442
1443
1444 static void writecache_writeback_endio(struct bio *bio)
1445 {
1446         struct writeback_struct *wb = container_of(bio, struct writeback_struct, bio);
1447         struct dm_writecache *wc = wb->wc;
1448         unsigned long flags;
1449
1450         raw_spin_lock_irqsave(&wc->endio_list_lock, flags);
1451         if (unlikely(list_empty(&wc->endio_list)))
1452                 wake_up_process(wc->endio_thread);
1453         list_add_tail(&wb->endio_entry, &wc->endio_list);
1454         raw_spin_unlock_irqrestore(&wc->endio_list_lock, flags);
1455 }
1456
1457 static void writecache_copy_endio(int read_err, unsigned long write_err, void *ptr)
1458 {
1459         struct copy_struct *c = ptr;
1460         struct dm_writecache *wc = c->wc;
1461
1462         c->error = likely(!(read_err | write_err)) ? 0 : -EIO;
1463
1464         raw_spin_lock_irq(&wc->endio_list_lock);
1465         if (unlikely(list_empty(&wc->endio_list)))
1466                 wake_up_process(wc->endio_thread);
1467         list_add_tail(&c->endio_entry, &wc->endio_list);
1468         raw_spin_unlock_irq(&wc->endio_list_lock);
1469 }
1470
1471 static void __writecache_endio_pmem(struct dm_writecache *wc, struct list_head *list)
1472 {
1473         unsigned i;
1474         struct writeback_struct *wb;
1475         struct wc_entry *e;
1476         unsigned long n_walked = 0;
1477
1478         do {
1479                 wb = list_entry(list->next, struct writeback_struct, endio_entry);
1480                 list_del(&wb->endio_entry);
1481
1482                 if (unlikely(wb->bio.bi_status != BLK_STS_OK))
1483                         writecache_error(wc, blk_status_to_errno(wb->bio.bi_status),
1484                                         "write error %d", wb->bio.bi_status);
1485                 i = 0;
1486                 do {
1487                         e = wb->wc_list[i];
1488                         BUG_ON(!e->write_in_progress);
1489                         e->write_in_progress = false;
1490                         INIT_LIST_HEAD(&e->lru);
1491                         if (!writecache_has_error(wc))
1492                                 writecache_free_entry(wc, e);
1493                         BUG_ON(!wc->writeback_size);
1494                         wc->writeback_size--;
1495                         n_walked++;
1496                         if (unlikely(n_walked >= ENDIO_LATENCY)) {
1497                                 writecache_commit_flushed(wc, false);
1498                                 wc_unlock(wc);
1499                                 wc_lock(wc);
1500                                 n_walked = 0;
1501                         }
1502                 } while (++i < wb->wc_list_n);
1503
1504                 if (wb->wc_list != wb->wc_list_inline)
1505                         kfree(wb->wc_list);
1506                 bio_put(&wb->bio);
1507         } while (!list_empty(list));
1508 }
1509
1510 static void __writecache_endio_ssd(struct dm_writecache *wc, struct list_head *list)
1511 {
1512         struct copy_struct *c;
1513         struct wc_entry *e;
1514
1515         do {
1516                 c = list_entry(list->next, struct copy_struct, endio_entry);
1517                 list_del(&c->endio_entry);
1518
1519                 if (unlikely(c->error))
1520                         writecache_error(wc, c->error, "copy error");
1521
1522                 e = c->e;
1523                 do {
1524                         BUG_ON(!e->write_in_progress);
1525                         e->write_in_progress = false;
1526                         INIT_LIST_HEAD(&e->lru);
1527                         if (!writecache_has_error(wc))
1528                                 writecache_free_entry(wc, e);
1529
1530                         BUG_ON(!wc->writeback_size);
1531                         wc->writeback_size--;
1532                         e++;
1533                 } while (--c->n_entries);
1534                 mempool_free(c, &wc->copy_pool);
1535         } while (!list_empty(list));
1536 }
1537
1538 static int writecache_endio_thread(void *data)
1539 {
1540         struct dm_writecache *wc = data;
1541
1542         while (1) {
1543                 struct list_head list;
1544
1545                 raw_spin_lock_irq(&wc->endio_list_lock);
1546                 if (!list_empty(&wc->endio_list))
1547                         goto pop_from_list;
1548                 set_current_state(TASK_INTERRUPTIBLE);
1549                 raw_spin_unlock_irq(&wc->endio_list_lock);
1550
1551                 if (unlikely(kthread_should_stop())) {
1552                         set_current_state(TASK_RUNNING);
1553                         break;
1554                 }
1555
1556                 schedule();
1557
1558                 continue;
1559
1560 pop_from_list:
1561                 list = wc->endio_list;
1562                 list.next->prev = list.prev->next = &list;
1563                 INIT_LIST_HEAD(&wc->endio_list);
1564                 raw_spin_unlock_irq(&wc->endio_list_lock);
1565
1566                 if (!WC_MODE_FUA(wc))
1567                         writecache_disk_flush(wc, wc->dev);
1568
1569                 wc_lock(wc);
1570
1571                 if (WC_MODE_PMEM(wc)) {
1572                         __writecache_endio_pmem(wc, &list);
1573                 } else {
1574                         __writecache_endio_ssd(wc, &list);
1575                         writecache_wait_for_ios(wc, READ);
1576                 }
1577
1578                 writecache_commit_flushed(wc, false);
1579
1580                 wc_unlock(wc);
1581         }
1582
1583         return 0;
1584 }
1585
1586 static bool wc_add_block(struct writeback_struct *wb, struct wc_entry *e, gfp_t gfp)
1587 {
1588         struct dm_writecache *wc = wb->wc;
1589         unsigned block_size = wc->block_size;
1590         void *address = memory_data(wc, e);
1591
1592         persistent_memory_flush_cache(address, block_size);
1593         return bio_add_page(&wb->bio, persistent_memory_page(address),
1594                             block_size, persistent_memory_page_offset(address)) != 0;
1595 }
1596
1597 struct writeback_list {
1598         struct list_head list;
1599         size_t size;
1600 };
1601
1602 static void __writeback_throttle(struct dm_writecache *wc, struct writeback_list *wbl)
1603 {
1604         if (unlikely(wc->max_writeback_jobs)) {
1605                 if (READ_ONCE(wc->writeback_size) - wbl->size >= wc->max_writeback_jobs) {
1606                         wc_lock(wc);
1607                         while (wc->writeback_size - wbl->size >= wc->max_writeback_jobs)
1608                                 writecache_wait_on_freelist(wc);
1609                         wc_unlock(wc);
1610                 }
1611         }
1612         cond_resched();
1613 }
1614
1615 static void __writecache_writeback_pmem(struct dm_writecache *wc, struct writeback_list *wbl)
1616 {
1617         struct wc_entry *e, *f;
1618         struct bio *bio;
1619         struct writeback_struct *wb;
1620         unsigned max_pages;
1621
1622         while (wbl->size) {
1623                 wbl->size--;
1624                 e = container_of(wbl->list.prev, struct wc_entry, lru);
1625                 list_del(&e->lru);
1626
1627                 max_pages = e->wc_list_contiguous;
1628
1629                 bio = bio_alloc_bioset(GFP_NOIO, max_pages, &wc->bio_set);
1630                 wb = container_of(bio, struct writeback_struct, bio);
1631                 wb->wc = wc;
1632                 bio->bi_end_io = writecache_writeback_endio;
1633                 bio_set_dev(bio, wc->dev->bdev);
1634                 bio->bi_iter.bi_sector = read_original_sector(wc, e);
1635                 if (max_pages <= WB_LIST_INLINE ||
1636                     unlikely(!(wb->wc_list = kmalloc_array(max_pages, sizeof(struct wc_entry *),
1637                                                            GFP_NOIO | __GFP_NORETRY |
1638                                                            __GFP_NOMEMALLOC | __GFP_NOWARN)))) {
1639                         wb->wc_list = wb->wc_list_inline;
1640                         max_pages = WB_LIST_INLINE;
1641                 }
1642
1643                 BUG_ON(!wc_add_block(wb, e, GFP_NOIO));
1644
1645                 wb->wc_list[0] = e;
1646                 wb->wc_list_n = 1;
1647
1648                 while (wbl->size && wb->wc_list_n < max_pages) {
1649                         f = container_of(wbl->list.prev, struct wc_entry, lru);
1650                         if (read_original_sector(wc, f) !=
1651                             read_original_sector(wc, e) + (wc->block_size >> SECTOR_SHIFT))
1652                                 break;
1653                         if (!wc_add_block(wb, f, GFP_NOWAIT | __GFP_NOWARN))
1654                                 break;
1655                         wbl->size--;
1656                         list_del(&f->lru);
1657                         wb->wc_list[wb->wc_list_n++] = f;
1658                         e = f;
1659                 }
1660                 bio_set_op_attrs(bio, REQ_OP_WRITE, WC_MODE_FUA(wc) * REQ_FUA);
1661                 if (writecache_has_error(wc)) {
1662                         bio->bi_status = BLK_STS_IOERR;
1663                         bio_endio(bio);
1664                 } else {
1665                         submit_bio(bio);
1666                 }
1667
1668                 __writeback_throttle(wc, wbl);
1669         }
1670 }
1671
1672 static void __writecache_writeback_ssd(struct dm_writecache *wc, struct writeback_list *wbl)
1673 {
1674         struct wc_entry *e, *f;
1675         struct dm_io_region from, to;
1676         struct copy_struct *c;
1677
1678         while (wbl->size) {
1679                 unsigned n_sectors;
1680
1681                 wbl->size--;
1682                 e = container_of(wbl->list.prev, struct wc_entry, lru);
1683                 list_del(&e->lru);
1684
1685                 n_sectors = e->wc_list_contiguous << (wc->block_size_bits - SECTOR_SHIFT);
1686
1687                 from.bdev = wc->ssd_dev->bdev;
1688                 from.sector = cache_sector(wc, e);
1689                 from.count = n_sectors;
1690                 to.bdev = wc->dev->bdev;
1691                 to.sector = read_original_sector(wc, e);
1692                 to.count = n_sectors;
1693
1694                 c = mempool_alloc(&wc->copy_pool, GFP_NOIO);
1695                 c->wc = wc;
1696                 c->e = e;
1697                 c->n_entries = e->wc_list_contiguous;
1698
1699                 while ((n_sectors -= wc->block_size >> SECTOR_SHIFT)) {
1700                         wbl->size--;
1701                         f = container_of(wbl->list.prev, struct wc_entry, lru);
1702                         BUG_ON(f != e + 1);
1703                         list_del(&f->lru);
1704                         e = f;
1705                 }
1706
1707                 dm_kcopyd_copy(wc->dm_kcopyd, &from, 1, &to, 0, writecache_copy_endio, c);
1708
1709                 __writeback_throttle(wc, wbl);
1710         }
1711 }
1712
1713 static void writecache_writeback(struct work_struct *work)
1714 {
1715         struct dm_writecache *wc = container_of(work, struct dm_writecache, writeback_work);
1716         struct blk_plug plug;
1717         struct wc_entry *f, *uninitialized_var(g), *e = NULL;
1718         struct rb_node *node, *next_node;
1719         struct list_head skipped;
1720         struct writeback_list wbl;
1721         unsigned long n_walked;
1722
1723         wc_lock(wc);
1724 restart:
1725         if (writecache_has_error(wc)) {
1726                 wc_unlock(wc);
1727                 return;
1728         }
1729
1730         if (unlikely(wc->writeback_all)) {
1731                 if (writecache_wait_for_writeback(wc))
1732                         goto restart;
1733         }
1734
1735         if (wc->overwrote_committed) {
1736                 writecache_wait_for_ios(wc, WRITE);
1737         }
1738
1739         n_walked = 0;
1740         INIT_LIST_HEAD(&skipped);
1741         INIT_LIST_HEAD(&wbl.list);
1742         wbl.size = 0;
1743         while (!list_empty(&wc->lru) &&
1744                (wc->writeback_all ||
1745                 wc->freelist_size + wc->writeback_size <= wc->freelist_low_watermark ||
1746                 (jiffies - container_of(wc->lru.prev, struct wc_entry, lru)->age >=
1747                  wc->max_age - wc->max_age / MAX_AGE_DIV))) {
1748
1749                 n_walked++;
1750                 if (unlikely(n_walked > WRITEBACK_LATENCY) &&
1751                     likely(!wc->writeback_all) && likely(!dm_suspended(wc->ti))) {
1752                         queue_work(wc->writeback_wq, &wc->writeback_work);
1753                         break;
1754                 }
1755
1756                 if (unlikely(wc->writeback_all)) {
1757                         if (unlikely(!e)) {
1758                                 writecache_flush(wc);
1759                                 e = container_of(rb_first(&wc->tree), struct wc_entry, rb_node);
1760                         } else
1761                                 e = g;
1762                 } else
1763                         e = container_of(wc->lru.prev, struct wc_entry, lru);
1764                 BUG_ON(e->write_in_progress);
1765                 if (unlikely(!writecache_entry_is_committed(wc, e))) {
1766                         writecache_flush(wc);
1767                 }
1768                 node = rb_prev(&e->rb_node);
1769                 if (node) {
1770                         f = container_of(node, struct wc_entry, rb_node);
1771                         if (unlikely(read_original_sector(wc, f) ==
1772                                      read_original_sector(wc, e))) {
1773                                 BUG_ON(!f->write_in_progress);
1774                                 list_del(&e->lru);
1775                                 list_add(&e->lru, &skipped);
1776                                 cond_resched();
1777                                 continue;
1778                         }
1779                 }
1780                 wc->writeback_size++;
1781                 list_del(&e->lru);
1782                 list_add(&e->lru, &wbl.list);
1783                 wbl.size++;
1784                 e->write_in_progress = true;
1785                 e->wc_list_contiguous = 1;
1786
1787                 f = e;
1788
1789                 while (1) {
1790                         next_node = rb_next(&f->rb_node);
1791                         if (unlikely(!next_node))
1792                                 break;
1793                         g = container_of(next_node, struct wc_entry, rb_node);
1794                         if (unlikely(read_original_sector(wc, g) ==
1795                             read_original_sector(wc, f))) {
1796                                 f = g;
1797                                 continue;
1798                         }
1799                         if (read_original_sector(wc, g) !=
1800                             read_original_sector(wc, f) + (wc->block_size >> SECTOR_SHIFT))
1801                                 break;
1802                         if (unlikely(g->write_in_progress))
1803                                 break;
1804                         if (unlikely(!writecache_entry_is_committed(wc, g)))
1805                                 break;
1806
1807                         if (!WC_MODE_PMEM(wc)) {
1808                                 if (g != f + 1)
1809                                         break;
1810                         }
1811
1812                         n_walked++;
1813                         //if (unlikely(n_walked > WRITEBACK_LATENCY) && likely(!wc->writeback_all))
1814                         //      break;
1815
1816                         wc->writeback_size++;
1817                         list_del(&g->lru);
1818                         list_add(&g->lru, &wbl.list);
1819                         wbl.size++;
1820                         g->write_in_progress = true;
1821                         g->wc_list_contiguous = BIO_MAX_PAGES;
1822                         f = g;
1823                         e->wc_list_contiguous++;
1824                         if (unlikely(e->wc_list_contiguous == BIO_MAX_PAGES)) {
1825                                 if (unlikely(wc->writeback_all)) {
1826                                         next_node = rb_next(&f->rb_node);
1827                                         if (likely(next_node))
1828                                                 g = container_of(next_node, struct wc_entry, rb_node);
1829                                 }
1830                                 break;
1831                         }
1832                 }
1833                 cond_resched();
1834         }
1835
1836         if (!list_empty(&skipped)) {
1837                 list_splice_tail(&skipped, &wc->lru);
1838                 /*
1839                  * If we didn't do any progress, we must wait until some
1840                  * writeback finishes to avoid burning CPU in a loop
1841                  */
1842                 if (unlikely(!wbl.size))
1843                         writecache_wait_for_writeback(wc);
1844         }
1845
1846         wc_unlock(wc);
1847
1848         blk_start_plug(&plug);
1849
1850         if (WC_MODE_PMEM(wc))
1851                 __writecache_writeback_pmem(wc, &wbl);
1852         else
1853                 __writecache_writeback_ssd(wc, &wbl);
1854
1855         blk_finish_plug(&plug);
1856
1857         if (unlikely(wc->writeback_all)) {
1858                 wc_lock(wc);
1859                 while (writecache_wait_for_writeback(wc));
1860                 wc_unlock(wc);
1861         }
1862 }
1863
1864 static int calculate_memory_size(uint64_t device_size, unsigned block_size,
1865                                  size_t *n_blocks_p, size_t *n_metadata_blocks_p)
1866 {
1867         uint64_t n_blocks, offset;
1868         struct wc_entry e;
1869
1870         n_blocks = device_size;
1871         do_div(n_blocks, block_size + sizeof(struct wc_memory_entry));
1872
1873         while (1) {
1874                 if (!n_blocks)
1875                         return -ENOSPC;
1876                 /* Verify the following entries[n_blocks] won't overflow */
1877                 if (n_blocks >= ((size_t)-sizeof(struct wc_memory_superblock) /
1878                                  sizeof(struct wc_memory_entry)))
1879                         return -EFBIG;
1880                 offset = offsetof(struct wc_memory_superblock, entries[n_blocks]);
1881                 offset = (offset + block_size - 1) & ~(uint64_t)(block_size - 1);
1882                 if (offset + n_blocks * block_size <= device_size)
1883                         break;
1884                 n_blocks--;
1885         }
1886
1887         /* check if the bit field overflows */
1888         e.index = n_blocks;
1889         if (e.index != n_blocks)
1890                 return -EFBIG;
1891
1892         if (n_blocks_p)
1893                 *n_blocks_p = n_blocks;
1894         if (n_metadata_blocks_p)
1895                 *n_metadata_blocks_p = offset >> __ffs(block_size);
1896         return 0;
1897 }
1898
1899 static int init_memory(struct dm_writecache *wc)
1900 {
1901         size_t b;
1902         int r;
1903
1904         r = calculate_memory_size(wc->memory_map_size, wc->block_size, &wc->n_blocks, NULL);
1905         if (r)
1906                 return r;
1907
1908         r = writecache_alloc_entries(wc);
1909         if (r)
1910                 return r;
1911
1912         for (b = 0; b < ARRAY_SIZE(sb(wc)->padding); b++)
1913                 pmem_assign(sb(wc)->padding[b], cpu_to_le64(0));
1914         pmem_assign(sb(wc)->version, cpu_to_le32(MEMORY_SUPERBLOCK_VERSION));
1915         pmem_assign(sb(wc)->block_size, cpu_to_le32(wc->block_size));
1916         pmem_assign(sb(wc)->n_blocks, cpu_to_le64(wc->n_blocks));
1917         pmem_assign(sb(wc)->seq_count, cpu_to_le64(0));
1918
1919         for (b = 0; b < wc->n_blocks; b++) {
1920                 write_original_sector_seq_count(wc, &wc->entries[b], -1, -1);
1921                 cond_resched();
1922         }
1923
1924         writecache_flush_all_metadata(wc);
1925         writecache_commit_flushed(wc, false);
1926         pmem_assign(sb(wc)->magic, cpu_to_le32(MEMORY_SUPERBLOCK_MAGIC));
1927         writecache_flush_region(wc, &sb(wc)->magic, sizeof sb(wc)->magic);
1928         writecache_commit_flushed(wc, false);
1929
1930         return 0;
1931 }
1932
1933 static void writecache_dtr(struct dm_target *ti)
1934 {
1935         struct dm_writecache *wc = ti->private;
1936
1937         if (!wc)
1938                 return;
1939
1940         if (wc->endio_thread)
1941                 kthread_stop(wc->endio_thread);
1942
1943         if (wc->flush_thread)
1944                 kthread_stop(wc->flush_thread);
1945
1946         bioset_exit(&wc->bio_set);
1947
1948         mempool_exit(&wc->copy_pool);
1949
1950         if (wc->writeback_wq)
1951                 destroy_workqueue(wc->writeback_wq);
1952
1953         if (wc->dev)
1954                 dm_put_device(ti, wc->dev);
1955
1956         if (wc->ssd_dev)
1957                 dm_put_device(ti, wc->ssd_dev);
1958
1959         if (wc->entries)
1960                 vfree(wc->entries);
1961
1962         if (wc->memory_map) {
1963                 if (WC_MODE_PMEM(wc))
1964                         persistent_memory_release(wc);
1965                 else
1966                         vfree(wc->memory_map);
1967         }
1968
1969         if (wc->dm_kcopyd)
1970                 dm_kcopyd_client_destroy(wc->dm_kcopyd);
1971
1972         if (wc->dm_io)
1973                 dm_io_client_destroy(wc->dm_io);
1974
1975         if (wc->dirty_bitmap)
1976                 vfree(wc->dirty_bitmap);
1977
1978         kfree(wc);
1979 }
1980
1981 static int writecache_ctr(struct dm_target *ti, unsigned argc, char **argv)
1982 {
1983         struct dm_writecache *wc;
1984         struct dm_arg_set as;
1985         const char *string;
1986         unsigned opt_params;
1987         size_t offset, data_size;
1988         int i, r;
1989         char dummy;
1990         int high_wm_percent = HIGH_WATERMARK;
1991         int low_wm_percent = LOW_WATERMARK;
1992         uint64_t x;
1993         struct wc_memory_superblock s;
1994
1995         static struct dm_arg _args[] = {
1996                 {0, 10, "Invalid number of feature args"},
1997         };
1998
1999         as.argc = argc;
2000         as.argv = argv;
2001
2002         wc = kzalloc(sizeof(struct dm_writecache), GFP_KERNEL);
2003         if (!wc) {
2004                 ti->error = "Cannot allocate writecache structure";
2005                 r = -ENOMEM;
2006                 goto bad;
2007         }
2008         ti->private = wc;
2009         wc->ti = ti;
2010
2011         mutex_init(&wc->lock);
2012         wc->max_age = MAX_AGE_UNSPECIFIED;
2013         writecache_poison_lists(wc);
2014         init_waitqueue_head(&wc->freelist_wait);
2015         timer_setup(&wc->autocommit_timer, writecache_autocommit_timer, 0);
2016         timer_setup(&wc->max_age_timer, writecache_max_age_timer, 0);
2017
2018         for (i = 0; i < 2; i++) {
2019                 atomic_set(&wc->bio_in_progress[i], 0);
2020                 init_waitqueue_head(&wc->bio_in_progress_wait[i]);
2021         }
2022
2023         wc->dm_io = dm_io_client_create();
2024         if (IS_ERR(wc->dm_io)) {
2025                 r = PTR_ERR(wc->dm_io);
2026                 ti->error = "Unable to allocate dm-io client";
2027                 wc->dm_io = NULL;
2028                 goto bad;
2029         }
2030
2031         wc->writeback_wq = alloc_workqueue("writecache-writeback", WQ_MEM_RECLAIM, 1);
2032         if (!wc->writeback_wq) {
2033                 r = -ENOMEM;
2034                 ti->error = "Could not allocate writeback workqueue";
2035                 goto bad;
2036         }
2037         INIT_WORK(&wc->writeback_work, writecache_writeback);
2038         INIT_WORK(&wc->flush_work, writecache_flush_work);
2039
2040         raw_spin_lock_init(&wc->endio_list_lock);
2041         INIT_LIST_HEAD(&wc->endio_list);
2042         wc->endio_thread = kthread_create(writecache_endio_thread, wc, "writecache_endio");
2043         if (IS_ERR(wc->endio_thread)) {
2044                 r = PTR_ERR(wc->endio_thread);
2045                 wc->endio_thread = NULL;
2046                 ti->error = "Couldn't spawn endio thread";
2047                 goto bad;
2048         }
2049         wake_up_process(wc->endio_thread);
2050
2051         /*
2052          * Parse the mode (pmem or ssd)
2053          */
2054         string = dm_shift_arg(&as);
2055         if (!string)
2056                 goto bad_arguments;
2057
2058         if (!strcasecmp(string, "s")) {
2059                 wc->pmem_mode = false;
2060         } else if (!strcasecmp(string, "p")) {
2061 #ifdef DM_WRITECACHE_HAS_PMEM
2062                 wc->pmem_mode = true;
2063                 wc->writeback_fua = true;
2064 #else
2065                 /*
2066                  * If the architecture doesn't support persistent memory or
2067                  * the kernel doesn't support any DAX drivers, this driver can
2068                  * only be used in SSD-only mode.
2069                  */
2070                 r = -EOPNOTSUPP;
2071                 ti->error = "Persistent memory or DAX not supported on this system";
2072                 goto bad;
2073 #endif
2074         } else {
2075                 goto bad_arguments;
2076         }
2077
2078         if (WC_MODE_PMEM(wc)) {
2079                 r = bioset_init(&wc->bio_set, BIO_POOL_SIZE,
2080                                 offsetof(struct writeback_struct, bio),
2081                                 BIOSET_NEED_BVECS);
2082                 if (r) {
2083                         ti->error = "Could not allocate bio set";
2084                         goto bad;
2085                 }
2086         } else {
2087                 r = mempool_init_kmalloc_pool(&wc->copy_pool, 1, sizeof(struct copy_struct));
2088                 if (r) {
2089                         ti->error = "Could not allocate mempool";
2090                         goto bad;
2091                 }
2092         }
2093
2094         /*
2095          * Parse the origin data device
2096          */
2097         string = dm_shift_arg(&as);
2098         if (!string)
2099                 goto bad_arguments;
2100         r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->dev);
2101         if (r) {
2102                 ti->error = "Origin data device lookup failed";
2103                 goto bad;
2104         }
2105
2106         /*
2107          * Parse cache data device (be it pmem or ssd)
2108          */
2109         string = dm_shift_arg(&as);
2110         if (!string)
2111                 goto bad_arguments;
2112
2113         r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->ssd_dev);
2114         if (r) {
2115                 ti->error = "Cache data device lookup failed";
2116                 goto bad;
2117         }
2118         wc->memory_map_size = i_size_read(wc->ssd_dev->bdev->bd_inode);
2119
2120         /*
2121          * Parse the cache block size
2122          */
2123         string = dm_shift_arg(&as);
2124         if (!string)
2125                 goto bad_arguments;
2126         if (sscanf(string, "%u%c", &wc->block_size, &dummy) != 1 ||
2127             wc->block_size < 512 || wc->block_size > PAGE_SIZE ||
2128             (wc->block_size & (wc->block_size - 1))) {
2129                 r = -EINVAL;
2130                 ti->error = "Invalid block size";
2131                 goto bad;
2132         }
2133         if (wc->block_size < bdev_logical_block_size(wc->dev->bdev) ||
2134             wc->block_size < bdev_logical_block_size(wc->ssd_dev->bdev)) {
2135                 r = -EINVAL;
2136                 ti->error = "Block size is smaller than device logical block size";
2137                 goto bad;
2138         }
2139         wc->block_size_bits = __ffs(wc->block_size);
2140
2141         wc->max_writeback_jobs = MAX_WRITEBACK_JOBS;
2142         wc->autocommit_blocks = !WC_MODE_PMEM(wc) ? AUTOCOMMIT_BLOCKS_SSD : AUTOCOMMIT_BLOCKS_PMEM;
2143         wc->autocommit_jiffies = msecs_to_jiffies(AUTOCOMMIT_MSEC);
2144
2145         /*
2146          * Parse optional arguments
2147          */
2148         r = dm_read_arg_group(_args, &as, &opt_params, &ti->error);
2149         if (r)
2150                 goto bad;
2151
2152         while (opt_params) {
2153                 string = dm_shift_arg(&as), opt_params--;
2154                 if (!strcasecmp(string, "start_sector") && opt_params >= 1) {
2155                         unsigned long long start_sector;
2156                         string = dm_shift_arg(&as), opt_params--;
2157                         if (sscanf(string, "%llu%c", &start_sector, &dummy) != 1)
2158                                 goto invalid_optional;
2159                         wc->start_sector = start_sector;
2160                         if (wc->start_sector != start_sector ||
2161                             wc->start_sector >= wc->memory_map_size >> SECTOR_SHIFT)
2162                                 goto invalid_optional;
2163                 } else if (!strcasecmp(string, "high_watermark") && opt_params >= 1) {
2164                         string = dm_shift_arg(&as), opt_params--;
2165                         if (sscanf(string, "%d%c", &high_wm_percent, &dummy) != 1)
2166                                 goto invalid_optional;
2167                         if (high_wm_percent < 0 || high_wm_percent > 100)
2168                                 goto invalid_optional;
2169                         wc->high_wm_percent_set = true;
2170                 } else if (!strcasecmp(string, "low_watermark") && opt_params >= 1) {
2171                         string = dm_shift_arg(&as), opt_params--;
2172                         if (sscanf(string, "%d%c", &low_wm_percent, &dummy) != 1)
2173                                 goto invalid_optional;
2174                         if (low_wm_percent < 0 || low_wm_percent > 100)
2175                                 goto invalid_optional;
2176                         wc->low_wm_percent_set = true;
2177                 } else if (!strcasecmp(string, "writeback_jobs") && opt_params >= 1) {
2178                         string = dm_shift_arg(&as), opt_params--;
2179                         if (sscanf(string, "%u%c", &wc->max_writeback_jobs, &dummy) != 1)
2180                                 goto invalid_optional;
2181                         wc->max_writeback_jobs_set = true;
2182                 } else if (!strcasecmp(string, "autocommit_blocks") && opt_params >= 1) {
2183                         string = dm_shift_arg(&as), opt_params--;
2184                         if (sscanf(string, "%u%c", &wc->autocommit_blocks, &dummy) != 1)
2185                                 goto invalid_optional;
2186                         wc->autocommit_blocks_set = true;
2187                 } else if (!strcasecmp(string, "autocommit_time") && opt_params >= 1) {
2188                         unsigned autocommit_msecs;
2189                         string = dm_shift_arg(&as), opt_params--;
2190                         if (sscanf(string, "%u%c", &autocommit_msecs, &dummy) != 1)
2191                                 goto invalid_optional;
2192                         if (autocommit_msecs > 3600000)
2193                                 goto invalid_optional;
2194                         wc->autocommit_jiffies = msecs_to_jiffies(autocommit_msecs);
2195                         wc->autocommit_time_set = true;
2196                 } else if (!strcasecmp(string, "max_age") && opt_params >= 1) {
2197                         unsigned max_age_msecs;
2198                         string = dm_shift_arg(&as), opt_params--;
2199                         if (sscanf(string, "%u%c", &max_age_msecs, &dummy) != 1)
2200                                 goto invalid_optional;
2201                         if (max_age_msecs > 86400000)
2202                                 goto invalid_optional;
2203                         wc->max_age = msecs_to_jiffies(max_age_msecs);
2204                 } else if (!strcasecmp(string, "cleaner")) {
2205                         wc->cleaner = true;
2206                 } else if (!strcasecmp(string, "fua")) {
2207                         if (WC_MODE_PMEM(wc)) {
2208                                 wc->writeback_fua = true;
2209                                 wc->writeback_fua_set = true;
2210                         } else goto invalid_optional;
2211                 } else if (!strcasecmp(string, "nofua")) {
2212                         if (WC_MODE_PMEM(wc)) {
2213                                 wc->writeback_fua = false;
2214                                 wc->writeback_fua_set = true;
2215                         } else goto invalid_optional;
2216                 } else {
2217 invalid_optional:
2218                         r = -EINVAL;
2219                         ti->error = "Invalid optional argument";
2220                         goto bad;
2221                 }
2222         }
2223
2224         if (high_wm_percent < low_wm_percent) {
2225                 r = -EINVAL;
2226                 ti->error = "High watermark must be greater than or equal to low watermark";
2227                 goto bad;
2228         }
2229
2230         if (WC_MODE_PMEM(wc)) {
2231                 r = persistent_memory_claim(wc);
2232                 if (r) {
2233                         ti->error = "Unable to map persistent memory for cache";
2234                         goto bad;
2235                 }
2236         } else {
2237                 size_t n_blocks, n_metadata_blocks;
2238                 uint64_t n_bitmap_bits;
2239
2240                 wc->memory_map_size -= (uint64_t)wc->start_sector << SECTOR_SHIFT;
2241
2242                 bio_list_init(&wc->flush_list);
2243                 wc->flush_thread = kthread_create(writecache_flush_thread, wc, "dm_writecache_flush");
2244                 if (IS_ERR(wc->flush_thread)) {
2245                         r = PTR_ERR(wc->flush_thread);
2246                         wc->flush_thread = NULL;
2247                         ti->error = "Couldn't spawn flush thread";
2248                         goto bad;
2249                 }
2250                 wake_up_process(wc->flush_thread);
2251
2252                 r = calculate_memory_size(wc->memory_map_size, wc->block_size,
2253                                           &n_blocks, &n_metadata_blocks);
2254                 if (r) {
2255                         ti->error = "Invalid device size";
2256                         goto bad;
2257                 }
2258
2259                 n_bitmap_bits = (((uint64_t)n_metadata_blocks << wc->block_size_bits) +
2260                                  BITMAP_GRANULARITY - 1) / BITMAP_GRANULARITY;
2261                 /* this is limitation of test_bit functions */
2262                 if (n_bitmap_bits > 1U << 31) {
2263                         r = -EFBIG;
2264                         ti->error = "Invalid device size";
2265                         goto bad;
2266                 }
2267
2268                 wc->memory_map = vmalloc(n_metadata_blocks << wc->block_size_bits);
2269                 if (!wc->memory_map) {
2270                         r = -ENOMEM;
2271                         ti->error = "Unable to allocate memory for metadata";
2272                         goto bad;
2273                 }
2274
2275                 wc->dm_kcopyd = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2276                 if (IS_ERR(wc->dm_kcopyd)) {
2277                         r = PTR_ERR(wc->dm_kcopyd);
2278                         ti->error = "Unable to allocate dm-kcopyd client";
2279                         wc->dm_kcopyd = NULL;
2280                         goto bad;
2281                 }
2282
2283                 wc->metadata_sectors = n_metadata_blocks << (wc->block_size_bits - SECTOR_SHIFT);
2284                 wc->dirty_bitmap_size = (n_bitmap_bits + BITS_PER_LONG - 1) /
2285                         BITS_PER_LONG * sizeof(unsigned long);
2286                 wc->dirty_bitmap = vzalloc(wc->dirty_bitmap_size);
2287                 if (!wc->dirty_bitmap) {
2288                         r = -ENOMEM;
2289                         ti->error = "Unable to allocate dirty bitmap";
2290                         goto bad;
2291                 }
2292
2293                 r = writecache_read_metadata(wc, wc->block_size >> SECTOR_SHIFT);
2294                 if (r) {
2295                         ti->error = "Unable to read first block of metadata";
2296                         goto bad;
2297                 }
2298         }
2299
2300         r = memcpy_mcsafe(&s, sb(wc), sizeof(struct wc_memory_superblock));
2301         if (r) {
2302                 ti->error = "Hardware memory error when reading superblock";
2303                 goto bad;
2304         }
2305         if (!le32_to_cpu(s.magic) && !le32_to_cpu(s.version)) {
2306                 r = init_memory(wc);
2307                 if (r) {
2308                         ti->error = "Unable to initialize device";
2309                         goto bad;
2310                 }
2311                 r = memcpy_mcsafe(&s, sb(wc), sizeof(struct wc_memory_superblock));
2312                 if (r) {
2313                         ti->error = "Hardware memory error when reading superblock";
2314                         goto bad;
2315                 }
2316         }
2317
2318         if (le32_to_cpu(s.magic) != MEMORY_SUPERBLOCK_MAGIC) {
2319                 ti->error = "Invalid magic in the superblock";
2320                 r = -EINVAL;
2321                 goto bad;
2322         }
2323
2324         if (le32_to_cpu(s.version) != MEMORY_SUPERBLOCK_VERSION) {
2325                 ti->error = "Invalid version in the superblock";
2326                 r = -EINVAL;
2327                 goto bad;
2328         }
2329
2330         if (le32_to_cpu(s.block_size) != wc->block_size) {
2331                 ti->error = "Block size does not match superblock";
2332                 r = -EINVAL;
2333                 goto bad;
2334         }
2335
2336         wc->n_blocks = le64_to_cpu(s.n_blocks);
2337
2338         offset = wc->n_blocks * sizeof(struct wc_memory_entry);
2339         if (offset / sizeof(struct wc_memory_entry) != le64_to_cpu(sb(wc)->n_blocks)) {
2340 overflow:
2341                 ti->error = "Overflow in size calculation";
2342                 r = -EINVAL;
2343                 goto bad;
2344         }
2345         offset += sizeof(struct wc_memory_superblock);
2346         if (offset < sizeof(struct wc_memory_superblock))
2347                 goto overflow;
2348         offset = (offset + wc->block_size - 1) & ~(size_t)(wc->block_size - 1);
2349         data_size = wc->n_blocks * (size_t)wc->block_size;
2350         if (!offset || (data_size / wc->block_size != wc->n_blocks) ||
2351             (offset + data_size < offset))
2352                 goto overflow;
2353         if (offset + data_size > wc->memory_map_size) {
2354                 ti->error = "Memory area is too small";
2355                 r = -EINVAL;
2356                 goto bad;
2357         }
2358
2359         wc->metadata_sectors = offset >> SECTOR_SHIFT;
2360         wc->block_start = (char *)sb(wc) + offset;
2361
2362         x = (uint64_t)wc->n_blocks * (100 - high_wm_percent);
2363         x += 50;
2364         do_div(x, 100);
2365         wc->freelist_high_watermark = x;
2366         x = (uint64_t)wc->n_blocks * (100 - low_wm_percent);
2367         x += 50;
2368         do_div(x, 100);
2369         wc->freelist_low_watermark = x;
2370
2371         if (wc->cleaner)
2372                 activate_cleaner(wc);
2373
2374         r = writecache_alloc_entries(wc);
2375         if (r) {
2376                 ti->error = "Cannot allocate memory";
2377                 goto bad;
2378         }
2379
2380         ti->num_flush_bios = 1;
2381         ti->flush_supported = true;
2382         ti->num_discard_bios = 1;
2383
2384         if (WC_MODE_PMEM(wc))
2385                 persistent_memory_flush_cache(wc->memory_map, wc->memory_map_size);
2386
2387         return 0;
2388
2389 bad_arguments:
2390         r = -EINVAL;
2391         ti->error = "Bad arguments";
2392 bad:
2393         writecache_dtr(ti);
2394         return r;
2395 }
2396
2397 static void writecache_status(struct dm_target *ti, status_type_t type,
2398                               unsigned status_flags, char *result, unsigned maxlen)
2399 {
2400         struct dm_writecache *wc = ti->private;
2401         unsigned extra_args;
2402         unsigned sz = 0;
2403         uint64_t x;
2404
2405         switch (type) {
2406         case STATUSTYPE_INFO:
2407                 DMEMIT("%ld %llu %llu %llu", writecache_has_error(wc),
2408                        (unsigned long long)wc->n_blocks, (unsigned long long)wc->freelist_size,
2409                        (unsigned long long)wc->writeback_size);
2410                 break;
2411         case STATUSTYPE_TABLE:
2412                 DMEMIT("%c %s %s %u ", WC_MODE_PMEM(wc) ? 'p' : 's',
2413                                 wc->dev->name, wc->ssd_dev->name, wc->block_size);
2414                 extra_args = 0;
2415                 if (wc->start_sector)
2416                         extra_args += 2;
2417                 if (wc->high_wm_percent_set && !wc->cleaner)
2418                         extra_args += 2;
2419                 if (wc->low_wm_percent_set && !wc->cleaner)
2420                         extra_args += 2;
2421                 if (wc->max_writeback_jobs_set)
2422                         extra_args += 2;
2423                 if (wc->autocommit_blocks_set)
2424                         extra_args += 2;
2425                 if (wc->autocommit_time_set)
2426                         extra_args += 2;
2427                 if (wc->cleaner)
2428                         extra_args++;
2429                 if (wc->writeback_fua_set)
2430                         extra_args++;
2431
2432                 DMEMIT("%u", extra_args);
2433                 if (wc->start_sector)
2434                         DMEMIT(" start_sector %llu", (unsigned long long)wc->start_sector);
2435                 if (wc->high_wm_percent_set && !wc->cleaner) {
2436                         x = (uint64_t)wc->freelist_high_watermark * 100;
2437                         x += wc->n_blocks / 2;
2438                         do_div(x, (size_t)wc->n_blocks);
2439                         DMEMIT(" high_watermark %u", 100 - (unsigned)x);
2440                 }
2441                 if (wc->low_wm_percent_set && !wc->cleaner) {
2442                         x = (uint64_t)wc->freelist_low_watermark * 100;
2443                         x += wc->n_blocks / 2;
2444                         do_div(x, (size_t)wc->n_blocks);
2445                         DMEMIT(" low_watermark %u", 100 - (unsigned)x);
2446                 }
2447                 if (wc->max_writeback_jobs_set)
2448                         DMEMIT(" writeback_jobs %u", wc->max_writeback_jobs);
2449                 if (wc->autocommit_blocks_set)
2450                         DMEMIT(" autocommit_blocks %u", wc->autocommit_blocks);
2451                 if (wc->autocommit_time_set)
2452                         DMEMIT(" autocommit_time %u", jiffies_to_msecs(wc->autocommit_jiffies));
2453                 if (wc->max_age != MAX_AGE_UNSPECIFIED)
2454                         DMEMIT(" max_age %u", jiffies_to_msecs(wc->max_age));
2455                 if (wc->cleaner)
2456                         DMEMIT(" cleaner");
2457                 if (wc->writeback_fua_set)
2458                         DMEMIT(" %sfua", wc->writeback_fua ? "" : "no");
2459                 break;
2460         }
2461 }
2462
2463 static struct target_type writecache_target = {
2464         .name                   = "writecache",
2465         .version                = {1, 3, 0},
2466         .module                 = THIS_MODULE,
2467         .ctr                    = writecache_ctr,
2468         .dtr                    = writecache_dtr,
2469         .status                 = writecache_status,
2470         .postsuspend            = writecache_suspend,
2471         .resume                 = writecache_resume,
2472         .message                = writecache_message,
2473         .map                    = writecache_map,
2474         .end_io                 = writecache_end_io,
2475         .iterate_devices        = writecache_iterate_devices,
2476         .io_hints               = writecache_io_hints,
2477 };
2478
2479 static int __init dm_writecache_init(void)
2480 {
2481         int r;
2482
2483         r = dm_register_target(&writecache_target);
2484         if (r < 0) {
2485                 DMERR("register failed %d", r);
2486                 return r;
2487         }
2488
2489         return 0;
2490 }
2491
2492 static void __exit dm_writecache_exit(void)
2493 {
2494         dm_unregister_target(&writecache_target);
2495 }
2496
2497 module_init(dm_writecache_init);
2498 module_exit(dm_writecache_exit);
2499
2500 MODULE_DESCRIPTION(DM_NAME " writecache target");
2501 MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
2502 MODULE_LICENSE("GPL");