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 #include <linux/delay.h>
  19 #include "dm-io-tracker.h"
  20
  21 #define DM_MSG_PREFIX "writecache"
  22
  23 #define HIGH_WATERMARK                  50
  24 #define LOW_WATERMARK                   45
  25 #define MAX_WRITEBACK_JOBS              0
  26 #define ENDIO_LATENCY                   16
  27 #define WRITEBACK_LATENCY               64
  28 #define AUTOCOMMIT_BLOCKS_SSD           65536
  29 #define AUTOCOMMIT_BLOCKS_PMEM          64
  30 #define AUTOCOMMIT_MSEC                 1000
  31 #define MAX_AGE_DIV                     16
  32 #define MAX_AGE_UNSPECIFIED             -1UL
  33 #define PAUSE_WRITEBACK                 (HZ * 3)
  34
  35 #define BITMAP_GRANULARITY      65536
  36 #if BITMAP_GRANULARITY < PAGE_SIZE
  37 #undef BITMAP_GRANULARITY
  38 #define BITMAP_GRANULARITY      PAGE_SIZE
  39 #endif
  40
  41 #if IS_ENABLED(CONFIG_ARCH_HAS_PMEM_API) && IS_ENABLED(CONFIG_DAX_DRIVER)
  42 #define DM_WRITECACHE_HAS_PMEM
  43 #endif
  44
  45 #ifdef DM_WRITECACHE_HAS_PMEM
  46 #define pmem_assign(dest, src)                                  \
  47 do {                                                            \
  48         typeof(dest) uniq = (src);                              \
  49         memcpy_flushcache(&(dest), &uniq, sizeof(dest));        \
  50 } while (0)
  51 #else
  52 #define pmem_assign(dest, src)  ((dest) = (src))
  53 #endif
  54
  55 #if IS_ENABLED(CONFIG_ARCH_HAS_COPY_MC) && defined(DM_WRITECACHE_HAS_PMEM)
  56 #define DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
  57 #endif
  58
  59 #define MEMORY_SUPERBLOCK_MAGIC         0x23489321
  60 #define MEMORY_SUPERBLOCK_VERSION       1
  61
  62 struct wc_memory_entry {
  63         __le64 original_sector;
  64         __le64 seq_count;
  65 };
  66
  67 struct wc_memory_superblock {
  68         union {
  69                 struct {
  70                         __le32 magic;
  71                         __le32 version;
  72                         __le32 block_size;
  73                         __le32 pad;
  74                         __le64 n_blocks;
  75                         __le64 seq_count;
  76                 };
  77                 __le64 padding[8];
  78         };
  79         struct wc_memory_entry entries[];
  80 };
  81
  82 struct wc_entry {
  83         struct rb_node rb_node;
  84         struct list_head lru;
  85         unsigned short wc_list_contiguous;
  86         bool write_in_progress
  87 #if BITS_PER_LONG == 64
  88                 :1
  89 #endif
  90         ;
  91         unsigned long index
  92 #if BITS_PER_LONG == 64
  93                 :47
  94 #endif
  95         ;
  96         unsigned long age;
  97 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
  98         uint64_t original_sector;
  99         uint64_t seq_count;
 100 #endif
 101 };
 102
 103 #ifdef DM_WRITECACHE_HAS_PMEM
 104 #define WC_MODE_PMEM(wc)                        ((wc)->pmem_mode)
 105 #define WC_MODE_FUA(wc)                         ((wc)->writeback_fua)
 106 #else
 107 #define WC_MODE_PMEM(wc)                        false
 108 #define WC_MODE_FUA(wc)                         false
 109 #endif
 110 #define WC_MODE_SORT_FREELIST(wc)               (!WC_MODE_PMEM(wc))
 111
 112 struct dm_writecache {
 113         struct mutex lock;
 114         struct list_head lru;
 115         union {
 116                 struct list_head freelist;
 117                 struct {
 118                         struct rb_root freetree;
 119                         struct wc_entry *current_free;
 120                 };
 121         };
 122         struct rb_root tree;
 123
 124         size_t freelist_size;
 125         size_t writeback_size;
 126         size_t freelist_high_watermark;
 127         size_t freelist_low_watermark;
 128         unsigned long max_age;
 129         unsigned long pause;
 130
 131         unsigned uncommitted_blocks;
 132         unsigned autocommit_blocks;
 133         unsigned max_writeback_jobs;
 134
 135         int error;
 136
 137         unsigned long autocommit_jiffies;
 138         struct timer_list autocommit_timer;
 139         struct wait_queue_head freelist_wait;
 140
 141         struct timer_list max_age_timer;
 142
 143         atomic_t bio_in_progress[2];
 144         struct wait_queue_head bio_in_progress_wait[2];
 145
 146         struct dm_target *ti;
 147         struct dm_dev *dev;
 148         struct dm_dev *ssd_dev;
 149         sector_t start_sector;
 150         void *memory_map;
 151         uint64_t memory_map_size;
 152         size_t metadata_sectors;
 153         size_t n_blocks;
 154         uint64_t seq_count;
 155         sector_t data_device_sectors;
 156         void *block_start;
 157         struct wc_entry *entries;
 158         unsigned block_size;
 159         unsigned char block_size_bits;
 160
 161         bool pmem_mode:1;
 162         bool writeback_fua:1;
 163
 164         bool overwrote_committed:1;
 165         bool memory_vmapped:1;
 166
 167         bool start_sector_set:1;
 168         bool high_wm_percent_set:1;
 169         bool low_wm_percent_set:1;
 170         bool max_writeback_jobs_set:1;
 171         bool autocommit_blocks_set:1;
 172         bool autocommit_time_set:1;
 173         bool max_age_set:1;
 174         bool writeback_fua_set:1;
 175         bool flush_on_suspend:1;
 176         bool cleaner:1;
 177         bool cleaner_set:1;
 178         bool metadata_only:1;
 179         bool pause_set:1;
 180
 181         unsigned high_wm_percent_value;
 182         unsigned low_wm_percent_value;
 183         unsigned autocommit_time_value;
 184         unsigned max_age_value;
 185         unsigned pause_value;
 186
 187         unsigned writeback_all;
 188         struct workqueue_struct *writeback_wq;
 189         struct work_struct writeback_work;
 190         struct work_struct flush_work;
 191
 192         struct dm_io_tracker iot;
 193
 194         struct dm_io_client *dm_io;
 195
 196         raw_spinlock_t endio_list_lock;
 197         struct list_head endio_list;
 198         struct task_struct *endio_thread;
 199
 200         struct task_struct *flush_thread;
 201         struct bio_list flush_list;
 202
 203         struct dm_kcopyd_client *dm_kcopyd;
 204         unsigned long *dirty_bitmap;
 205         unsigned dirty_bitmap_size;
 206
 207         struct bio_set bio_set;
 208         mempool_t copy_pool;
 209 };
 210
 211 #define WB_LIST_INLINE          16
 212
 213 struct writeback_struct {
 214         struct list_head endio_entry;
 215         struct dm_writecache *wc;
 216         struct wc_entry **wc_list;
 217         unsigned wc_list_n;
 218         struct wc_entry *wc_list_inline[WB_LIST_INLINE];
 219         struct bio bio;
 220 };
 221
 222 struct copy_struct {
 223         struct list_head endio_entry;
 224         struct dm_writecache *wc;
 225         struct wc_entry *e;
 226         unsigned n_entries;
 227         int error;
 228 };
 229
 230 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(dm_writecache_throttle,
 231                                             "A percentage of time allocated for data copying");
 232
 233 static void wc_lock(struct dm_writecache *wc)
 234 {
 235         mutex_lock(&wc->lock);
 236 }
 237
 238 static void wc_unlock(struct dm_writecache *wc)
 239 {
 240         mutex_unlock(&wc->lock);
 241 }
 242
 243 #ifdef DM_WRITECACHE_HAS_PMEM
 244 static int persistent_memory_claim(struct dm_writecache *wc)
 245 {
 246         int r;
 247         loff_t s;
 248         long p, da;
 249         pfn_t pfn;
 250         int id;
 251         struct page **pages;
 252         sector_t offset;
 253
 254         wc->memory_vmapped = false;
 255
 256         s = wc->memory_map_size;
 257         p = s >> PAGE_SHIFT;
 258         if (!p) {
 259                 r = -EINVAL;
 260                 goto err1;
 261         }
 262         if (p != s >> PAGE_SHIFT) {
 263                 r = -EOVERFLOW;
 264                 goto err1;
 265         }
 266
 267         offset = get_start_sect(wc->ssd_dev->bdev);
 268         if (offset & (PAGE_SIZE / 512 - 1)) {
 269                 r = -EINVAL;
 270                 goto err1;
 271         }
 272         offset >>= PAGE_SHIFT - 9;
 273
 274         id = dax_read_lock();
 275
 276         da = dax_direct_access(wc->ssd_dev->dax_dev, offset, p, &wc->memory_map, &pfn);
 277         if (da < 0) {
 278                 wc->memory_map = NULL;
 279                 r = da;
 280                 goto err2;
 281         }
 282         if (!pfn_t_has_page(pfn)) {
 283                 wc->memory_map = NULL;
 284                 r = -EOPNOTSUPP;
 285                 goto err2;
 286         }
 287         if (da != p) {
 288                 long i;
 289                 wc->memory_map = NULL;
 290                 pages = kvmalloc_array(p, sizeof(struct page *), GFP_KERNEL);
 291                 if (!pages) {
 292                         r = -ENOMEM;
 293                         goto err2;
 294                 }
 295                 i = 0;
 296                 do {
 297                         long daa;
 298                         daa = dax_direct_access(wc->ssd_dev->dax_dev, offset + i, p - i,
 299                                                 NULL, &pfn);
 300                         if (daa <= 0) {
 301                                 r = daa ? daa : -EINVAL;
 302                                 goto err3;
 303                         }
 304                         if (!pfn_t_has_page(pfn)) {
 305                                 r = -EOPNOTSUPP;
 306                                 goto err3;
 307                         }
 308                         while (daa-- && i < p) {
 309                                 pages[i++] = pfn_t_to_page(pfn);
 310                                 pfn.val++;
 311                                 if (!(i & 15))
 312                                         cond_resched();
 313                         }
 314                 } while (i < p);
 315                 wc->memory_map = vmap(pages, p, VM_MAP, PAGE_KERNEL);
 316                 if (!wc->memory_map) {
 317                         r = -ENOMEM;
 318                         goto err3;
 319                 }
 320                 kvfree(pages);
 321                 wc->memory_vmapped = true;
 322         }
 323
 324         dax_read_unlock(id);
 325
 326         wc->memory_map += (size_t)wc->start_sector << SECTOR_SHIFT;
 327         wc->memory_map_size -= (size_t)wc->start_sector << SECTOR_SHIFT;
 328
 329         return 0;
 330 err3:
 331         kvfree(pages);
 332 err2:
 333         dax_read_unlock(id);
 334 err1:
 335         return r;
 336 }
 337 #else
 338 static int persistent_memory_claim(struct dm_writecache *wc)
 339 {
 340         return -EOPNOTSUPP;
 341 }
 342 #endif
 343
 344 static void persistent_memory_release(struct dm_writecache *wc)
 345 {
 346         if (wc->memory_vmapped)
 347                 vunmap(wc->memory_map - ((size_t)wc->start_sector << SECTOR_SHIFT));
 348 }
 349
 350 static struct page *persistent_memory_page(void *addr)
 351 {
 352         if (is_vmalloc_addr(addr))
 353                 return vmalloc_to_page(addr);
 354         else
 355                 return virt_to_page(addr);
 356 }
 357
 358 static unsigned persistent_memory_page_offset(void *addr)
 359 {
 360         return (unsigned long)addr & (PAGE_SIZE - 1);
 361 }
 362
 363 static void persistent_memory_flush_cache(void *ptr, size_t size)
 364 {
 365         if (is_vmalloc_addr(ptr))
 366                 flush_kernel_vmap_range(ptr, size);
 367 }
 368
 369 static void persistent_memory_invalidate_cache(void *ptr, size_t size)
 370 {
 371         if (is_vmalloc_addr(ptr))
 372                 invalidate_kernel_vmap_range(ptr, size);
 373 }
 374
 375 static struct wc_memory_superblock *sb(struct dm_writecache *wc)
 376 {
 377         return wc->memory_map;
 378 }
 379
 380 static struct wc_memory_entry *memory_entry(struct dm_writecache *wc, struct wc_entry *e)
 381 {
 382         return &sb(wc)->entries[e->index];
 383 }
 384
 385 static void *memory_data(struct dm_writecache *wc, struct wc_entry *e)
 386 {
 387         return (char *)wc->block_start + (e->index << wc->block_size_bits);
 388 }
 389
 390 static sector_t cache_sector(struct dm_writecache *wc, struct wc_entry *e)
 391 {
 392         return wc->start_sector + wc->metadata_sectors +
 393                 ((sector_t)e->index << (wc->block_size_bits - SECTOR_SHIFT));
 394 }
 395
 396 static uint64_t read_original_sector(struct dm_writecache *wc, struct wc_entry *e)
 397 {
 398 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 399         return e->original_sector;
 400 #else
 401         return le64_to_cpu(memory_entry(wc, e)->original_sector);
 402 #endif
 403 }
 404
 405 static uint64_t read_seq_count(struct dm_writecache *wc, struct wc_entry *e)
 406 {
 407 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 408         return e->seq_count;
 409 #else
 410         return le64_to_cpu(memory_entry(wc, e)->seq_count);
 411 #endif
 412 }
 413
 414 static void clear_seq_count(struct dm_writecache *wc, struct wc_entry *e)
 415 {
 416 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 417         e->seq_count = -1;
 418 #endif
 419         pmem_assign(memory_entry(wc, e)->seq_count, cpu_to_le64(-1));
 420 }
 421
 422 static void write_original_sector_seq_count(struct dm_writecache *wc, struct wc_entry *e,
 423                                             uint64_t original_sector, uint64_t seq_count)
 424 {
 425         struct wc_memory_entry me;
 426 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 427         e->original_sector = original_sector;
 428         e->seq_count = seq_count;
 429 #endif
 430         me.original_sector = cpu_to_le64(original_sector);
 431         me.seq_count = cpu_to_le64(seq_count);
 432         pmem_assign(*memory_entry(wc, e), me);
 433 }
 434
 435 #define writecache_error(wc, err, msg, arg...)                          \
 436 do {                                                                    \
 437         if (!cmpxchg(&(wc)->error, 0, err))                             \
 438                 DMERR(msg, ##arg);                                      \
 439         wake_up(&(wc)->freelist_wait);                                  \
 440 } while (0)
 441
 442 #define writecache_has_error(wc)        (unlikely(READ_ONCE((wc)->error)))
 443
 444 static void writecache_flush_all_metadata(struct dm_writecache *wc)
 445 {
 446         if (!WC_MODE_PMEM(wc))
 447                 memset(wc->dirty_bitmap, -1, wc->dirty_bitmap_size);
 448 }
 449
 450 static void writecache_flush_region(struct dm_writecache *wc, void *ptr, size_t size)
 451 {
 452         if (!WC_MODE_PMEM(wc))
 453                 __set_bit(((char *)ptr - (char *)wc->memory_map) / BITMAP_GRANULARITY,
 454                           wc->dirty_bitmap);
 455 }
 456
 457 static void writecache_disk_flush(struct dm_writecache *wc, struct dm_dev *dev);
 458
 459 struct io_notify {
 460         struct dm_writecache *wc;
 461         struct completion c;
 462         atomic_t count;
 463 };
 464
 465 static void writecache_notify_io(unsigned long error, void *context)
 466 {
 467         struct io_notify *endio = context;
 468
 469         if (unlikely(error != 0))
 470                 writecache_error(endio->wc, -EIO, "error writing metadata");
 471         BUG_ON(atomic_read(&endio->count) <= 0);
 472         if (atomic_dec_and_test(&endio->count))
 473                 complete(&endio->c);
 474 }
 475
 476 static void writecache_wait_for_ios(struct dm_writecache *wc, int direction)
 477 {
 478         wait_event(wc->bio_in_progress_wait[direction],
 479                    !atomic_read(&wc->bio_in_progress[direction]));
 480 }
 481
 482 static void ssd_commit_flushed(struct dm_writecache *wc, bool wait_for_ios)
 483 {
 484         struct dm_io_region region;
 485         struct dm_io_request req;
 486         struct io_notify endio = {
 487                 wc,
 488                 COMPLETION_INITIALIZER_ONSTACK(endio.c),
 489                 ATOMIC_INIT(1),
 490         };
 491         unsigned bitmap_bits = wc->dirty_bitmap_size * 8;
 492         unsigned i = 0;
 493
 494         while (1) {
 495                 unsigned j;
 496                 i = find_next_bit(wc->dirty_bitmap, bitmap_bits, i);
 497                 if (unlikely(i == bitmap_bits))
 498                         break;
 499                 j = find_next_zero_bit(wc->dirty_bitmap, bitmap_bits, i);
 500
 501                 region.bdev = wc->ssd_dev->bdev;
 502                 region.sector = (sector_t)i * (BITMAP_GRANULARITY >> SECTOR_SHIFT);
 503                 region.count = (sector_t)(j - i) * (BITMAP_GRANULARITY >> SECTOR_SHIFT);
 504
 505                 if (unlikely(region.sector >= wc->metadata_sectors))
 506                         break;
 507                 if (unlikely(region.sector + region.count > wc->metadata_sectors))
 508                         region.count = wc->metadata_sectors - region.sector;
 509
 510                 region.sector += wc->start_sector;
 511                 atomic_inc(&endio.count);
 512                 req.bi_op = REQ_OP_WRITE;
 513                 req.bi_op_flags = REQ_SYNC;
 514                 req.mem.type = DM_IO_VMA;
 515                 req.mem.ptr.vma = (char *)wc->memory_map + (size_t)i * BITMAP_GRANULARITY;
 516                 req.client = wc->dm_io;
 517                 req.notify.fn = writecache_notify_io;
 518                 req.notify.context = &endio;
 519
 520                 /* writing via async dm-io (implied by notify.fn above) won't return an error */
 521                 (void) dm_io(&req, 1, &region, NULL);
 522                 i = j;
 523         }
 524
 525         writecache_notify_io(0, &endio);
 526         wait_for_completion_io(&endio.c);
 527
 528         if (wait_for_ios)
 529                 writecache_wait_for_ios(wc, WRITE);
 530
 531         writecache_disk_flush(wc, wc->ssd_dev);
 532
 533         memset(wc->dirty_bitmap, 0, wc->dirty_bitmap_size);
 534 }
 535
 536 static void ssd_commit_superblock(struct dm_writecache *wc)
 537 {
 538         int r;
 539         struct dm_io_region region;
 540         struct dm_io_request req;
 541
 542         region.bdev = wc->ssd_dev->bdev;
 543         region.sector = 0;
 544         region.count = max(4096U, wc->block_size) >> SECTOR_SHIFT;
 545
 546         if (unlikely(region.sector + region.count > wc->metadata_sectors))
 547                 region.count = wc->metadata_sectors - region.sector;
 548
 549         region.sector += wc->start_sector;
 550
 551         req.bi_op = REQ_OP_WRITE;
 552         req.bi_op_flags = REQ_SYNC | REQ_FUA;
 553         req.mem.type = DM_IO_VMA;
 554         req.mem.ptr.vma = (char *)wc->memory_map;
 555         req.client = wc->dm_io;
 556         req.notify.fn = NULL;
 557         req.notify.context = NULL;
 558
 559         r = dm_io(&req, 1, &region, NULL);
 560         if (unlikely(r))
 561                 writecache_error(wc, r, "error writing superblock");
 562 }
 563
 564 static void writecache_commit_flushed(struct dm_writecache *wc, bool wait_for_ios)
 565 {
 566         if (WC_MODE_PMEM(wc))
 567                 pmem_wmb();
 568         else
 569                 ssd_commit_flushed(wc, wait_for_ios);
 570 }
 571
 572 static void writecache_disk_flush(struct dm_writecache *wc, struct dm_dev *dev)
 573 {
 574         int r;
 575         struct dm_io_region region;
 576         struct dm_io_request req;
 577
 578         region.bdev = dev->bdev;
 579         region.sector = 0;
 580         region.count = 0;
 581         req.bi_op = REQ_OP_WRITE;
 582         req.bi_op_flags = REQ_PREFLUSH;
 583         req.mem.type = DM_IO_KMEM;
 584         req.mem.ptr.addr = NULL;
 585         req.client = wc->dm_io;
 586         req.notify.fn = NULL;
 587
 588         r = dm_io(&req, 1, &region, NULL);
 589         if (unlikely(r))
 590                 writecache_error(wc, r, "error flushing metadata: %d", r);
 591 }
 592
 593 #define WFE_RETURN_FOLLOWING    1
 594 #define WFE_LOWEST_SEQ          2
 595
 596 static struct wc_entry *writecache_find_entry(struct dm_writecache *wc,
 597                                               uint64_t block, int flags)
 598 {
 599         struct wc_entry *e;
 600         struct rb_node *node = wc->tree.rb_node;
 601
 602         if (unlikely(!node))
 603                 return NULL;
 604
 605         while (1) {
 606                 e = container_of(node, struct wc_entry, rb_node);
 607                 if (read_original_sector(wc, e) == block)
 608                         break;
 609
 610                 node = (read_original_sector(wc, e) >= block ?
 611                         e->rb_node.rb_left : e->rb_node.rb_right);
 612                 if (unlikely(!node)) {
 613                         if (!(flags & WFE_RETURN_FOLLOWING))
 614                                 return NULL;
 615                         if (read_original_sector(wc, e) >= block) {
 616                                 return e;
 617                         } else {
 618                                 node = rb_next(&e->rb_node);
 619                                 if (unlikely(!node))
 620                                         return NULL;
 621                                 e = container_of(node, struct wc_entry, rb_node);
 622                                 return e;
 623                         }
 624                 }
 625         }
 626
 627         while (1) {
 628                 struct wc_entry *e2;
 629                 if (flags & WFE_LOWEST_SEQ)
 630                         node = rb_prev(&e->rb_node);
 631                 else
 632                         node = rb_next(&e->rb_node);
 633                 if (unlikely(!node))
 634                         return e;
 635                 e2 = container_of(node, struct wc_entry, rb_node);
 636                 if (read_original_sector(wc, e2) != block)
 637                         return e;
 638                 e = e2;
 639         }
 640 }
 641
 642 static void writecache_insert_entry(struct dm_writecache *wc, struct wc_entry *ins)
 643 {
 644         struct wc_entry *e;
 645         struct rb_node **node = &wc->tree.rb_node, *parent = NULL;
 646
 647         while (*node) {
 648                 e = container_of(*node, struct wc_entry, rb_node);
 649                 parent = &e->rb_node;
 650                 if (read_original_sector(wc, e) > read_original_sector(wc, ins))
 651                         node = &parent->rb_left;
 652                 else
 653                         node = &parent->rb_right;
 654         }
 655         rb_link_node(&ins->rb_node, parent, node);
 656         rb_insert_color(&ins->rb_node, &wc->tree);
 657         list_add(&ins->lru, &wc->lru);
 658         ins->age = jiffies;
 659 }
 660
 661 static void writecache_unlink(struct dm_writecache *wc, struct wc_entry *e)
 662 {
 663         list_del(&e->lru);
 664         rb_erase(&e->rb_node, &wc->tree);
 665 }
 666
 667 static void writecache_add_to_freelist(struct dm_writecache *wc, struct wc_entry *e)
 668 {
 669         if (WC_MODE_SORT_FREELIST(wc)) {
 670                 struct rb_node **node = &wc->freetree.rb_node, *parent = NULL;
 671                 if (unlikely(!*node))
 672                         wc->current_free = e;
 673                 while (*node) {
 674                         parent = *node;
 675                         if (&e->rb_node < *node)
 676                                 node = &parent->rb_left;
 677                         else
 678                                 node = &parent->rb_right;
 679                 }
 680                 rb_link_node(&e->rb_node, parent, node);
 681                 rb_insert_color(&e->rb_node, &wc->freetree);
 682         } else {
 683                 list_add_tail(&e->lru, &wc->freelist);
 684         }
 685         wc->freelist_size++;
 686 }
 687
 688 static inline void writecache_verify_watermark(struct dm_writecache *wc)
 689 {
 690         if (unlikely(wc->freelist_size + wc->writeback_size <= wc->freelist_high_watermark))
 691                 queue_work(wc->writeback_wq, &wc->writeback_work);
 692 }
 693
 694 static void writecache_max_age_timer(struct timer_list *t)
 695 {
 696         struct dm_writecache *wc = from_timer(wc, t, max_age_timer);
 697
 698         if (!dm_suspended(wc->ti) && !writecache_has_error(wc)) {
 699                 queue_work(wc->writeback_wq, &wc->writeback_work);
 700                 mod_timer(&wc->max_age_timer, jiffies + wc->max_age / MAX_AGE_DIV);
 701         }
 702 }
 703
 704 static struct wc_entry *writecache_pop_from_freelist(struct dm_writecache *wc, sector_t expected_sector)
 705 {
 706         struct wc_entry *e;
 707
 708         if (WC_MODE_SORT_FREELIST(wc)) {
 709                 struct rb_node *next;
 710                 if (unlikely(!wc->current_free))
 711                         return NULL;
 712                 e = wc->current_free;
 713                 if (expected_sector != (sector_t)-1 && unlikely(cache_sector(wc, e) != expected_sector))
 714                         return NULL;
 715                 next = rb_next(&e->rb_node);
 716                 rb_erase(&e->rb_node, &wc->freetree);
 717                 if (unlikely(!next))
 718                         next = rb_first(&wc->freetree);
 719                 wc->current_free = next ? container_of(next, struct wc_entry, rb_node) : NULL;
 720         } else {
 721                 if (unlikely(list_empty(&wc->freelist)))
 722                         return NULL;
 723                 e = container_of(wc->freelist.next, struct wc_entry, lru);
 724                 if (expected_sector != (sector_t)-1 && unlikely(cache_sector(wc, e) != expected_sector))
 725                         return NULL;
 726                 list_del(&e->lru);
 727         }
 728         wc->freelist_size--;
 729
 730         writecache_verify_watermark(wc);
 731
 732         return e;
 733 }
 734
 735 static void writecache_free_entry(struct dm_writecache *wc, struct wc_entry *e)
 736 {
 737         writecache_unlink(wc, e);
 738         writecache_add_to_freelist(wc, e);
 739         clear_seq_count(wc, e);
 740         writecache_flush_region(wc, memory_entry(wc, e), sizeof(struct wc_memory_entry));
 741         if (unlikely(waitqueue_active(&wc->freelist_wait)))
 742                 wake_up(&wc->freelist_wait);
 743 }
 744
 745 static void writecache_wait_on_freelist(struct dm_writecache *wc)
 746 {
 747         DEFINE_WAIT(wait);
 748
 749         prepare_to_wait(&wc->freelist_wait, &wait, TASK_UNINTERRUPTIBLE);
 750         wc_unlock(wc);
 751         io_schedule();
 752         finish_wait(&wc->freelist_wait, &wait);
 753         wc_lock(wc);
 754 }
 755
 756 static void writecache_poison_lists(struct dm_writecache *wc)
 757 {
 758         /*
 759          * Catch incorrect access to these values while the device is suspended.
 760          */
 761         memset(&wc->tree, -1, sizeof wc->tree);
 762         wc->lru.next = LIST_POISON1;
 763         wc->lru.prev = LIST_POISON2;
 764         wc->freelist.next = LIST_POISON1;
 765         wc->freelist.prev = LIST_POISON2;
 766 }
 767
 768 static void writecache_flush_entry(struct dm_writecache *wc, struct wc_entry *e)
 769 {
 770         writecache_flush_region(wc, memory_entry(wc, e), sizeof(struct wc_memory_entry));
 771         if (WC_MODE_PMEM(wc))
 772                 writecache_flush_region(wc, memory_data(wc, e), wc->block_size);
 773 }
 774
 775 static bool writecache_entry_is_committed(struct dm_writecache *wc, struct wc_entry *e)
 776 {
 777         return read_seq_count(wc, e) < wc->seq_count;
 778 }
 779
 780 static void writecache_flush(struct dm_writecache *wc)
 781 {
 782         struct wc_entry *e, *e2;
 783         bool need_flush_after_free;
 784
 785         wc->uncommitted_blocks = 0;
 786         del_timer(&wc->autocommit_timer);
 787
 788         if (list_empty(&wc->lru))
 789                 return;
 790
 791         e = container_of(wc->lru.next, struct wc_entry, lru);
 792         if (writecache_entry_is_committed(wc, e)) {
 793                 if (wc->overwrote_committed) {
 794                         writecache_wait_for_ios(wc, WRITE);
 795                         writecache_disk_flush(wc, wc->ssd_dev);
 796                         wc->overwrote_committed = false;
 797                 }
 798                 return;
 799         }
 800         while (1) {
 801                 writecache_flush_entry(wc, e);
 802                 if (unlikely(e->lru.next == &wc->lru))
 803                         break;
 804                 e2 = container_of(e->lru.next, struct wc_entry, lru);
 805                 if (writecache_entry_is_committed(wc, e2))
 806                         break;
 807                 e = e2;
 808                 cond_resched();
 809         }
 810         writecache_commit_flushed(wc, true);
 811
 812         wc->seq_count++;
 813         pmem_assign(sb(wc)->seq_count, cpu_to_le64(wc->seq_count));
 814         if (WC_MODE_PMEM(wc))
 815                 writecache_commit_flushed(wc, false);
 816         else
 817                 ssd_commit_superblock(wc);
 818
 819         wc->overwrote_committed = false;
 820
 821         need_flush_after_free = false;
 822         while (1) {
 823                 /* Free another committed entry with lower seq-count */
 824                 struct rb_node *rb_node = rb_prev(&e->rb_node);
 825
 826                 if (rb_node) {
 827                         e2 = container_of(rb_node, struct wc_entry, rb_node);
 828                         if (read_original_sector(wc, e2) == read_original_sector(wc, e) &&
 829                             likely(!e2->write_in_progress)) {
 830                                 writecache_free_entry(wc, e2);
 831                                 need_flush_after_free = true;
 832                         }
 833                 }
 834                 if (unlikely(e->lru.prev == &wc->lru))
 835                         break;
 836                 e = container_of(e->lru.prev, struct wc_entry, lru);
 837                 cond_resched();
 838         }
 839
 840         if (need_flush_after_free)
 841                 writecache_commit_flushed(wc, false);
 842 }
 843
 844 static void writecache_flush_work(struct work_struct *work)
 845 {
 846         struct dm_writecache *wc = container_of(work, struct dm_writecache, flush_work);
 847
 848         wc_lock(wc);
 849         writecache_flush(wc);
 850         wc_unlock(wc);
 851 }
 852
 853 static void writecache_autocommit_timer(struct timer_list *t)
 854 {
 855         struct dm_writecache *wc = from_timer(wc, t, autocommit_timer);
 856         if (!writecache_has_error(wc))
 857                 queue_work(wc->writeback_wq, &wc->flush_work);
 858 }
 859
 860 static void writecache_schedule_autocommit(struct dm_writecache *wc)
 861 {
 862         if (!timer_pending(&wc->autocommit_timer))
 863                 mod_timer(&wc->autocommit_timer, jiffies + wc->autocommit_jiffies);
 864 }
 865
 866 static void writecache_discard(struct dm_writecache *wc, sector_t start, sector_t end)
 867 {
 868         struct wc_entry *e;
 869         bool discarded_something = false;
 870
 871         e = writecache_find_entry(wc, start, WFE_RETURN_FOLLOWING | WFE_LOWEST_SEQ);
 872         if (unlikely(!e))
 873                 return;
 874
 875         while (read_original_sector(wc, e) < end) {
 876                 struct rb_node *node = rb_next(&e->rb_node);
 877
 878                 if (likely(!e->write_in_progress)) {
 879                         if (!discarded_something) {
 880                                 if (!WC_MODE_PMEM(wc)) {
 881                                         writecache_wait_for_ios(wc, READ);
 882                                         writecache_wait_for_ios(wc, WRITE);
 883                                 }
 884                                 discarded_something = true;
 885                         }
 886                         if (!writecache_entry_is_committed(wc, e))
 887                                 wc->uncommitted_blocks--;
 888                         writecache_free_entry(wc, e);
 889                 }
 890
 891                 if (unlikely(!node))
 892                         break;
 893
 894                 e = container_of(node, struct wc_entry, rb_node);
 895         }
 896
 897         if (discarded_something)
 898                 writecache_commit_flushed(wc, false);
 899 }
 900
 901 static bool writecache_wait_for_writeback(struct dm_writecache *wc)
 902 {
 903         if (wc->writeback_size) {
 904                 writecache_wait_on_freelist(wc);
 905                 return true;
 906         }
 907         return false;
 908 }
 909
 910 static void writecache_suspend(struct dm_target *ti)
 911 {
 912         struct dm_writecache *wc = ti->private;
 913         bool flush_on_suspend;
 914
 915         del_timer_sync(&wc->autocommit_timer);
 916         del_timer_sync(&wc->max_age_timer);
 917
 918         wc_lock(wc);
 919         writecache_flush(wc);
 920         flush_on_suspend = wc->flush_on_suspend;
 921         if (flush_on_suspend) {
 922                 wc->flush_on_suspend = false;
 923                 wc->writeback_all++;
 924                 queue_work(wc->writeback_wq, &wc->writeback_work);
 925         }
 926         wc_unlock(wc);
 927
 928         drain_workqueue(wc->writeback_wq);
 929
 930         wc_lock(wc);
 931         if (flush_on_suspend)
 932                 wc->writeback_all--;
 933         while (writecache_wait_for_writeback(wc));
 934
 935         if (WC_MODE_PMEM(wc))
 936                 persistent_memory_flush_cache(wc->memory_map, wc->memory_map_size);
 937
 938         writecache_poison_lists(wc);
 939
 940         wc_unlock(wc);
 941 }
 942
 943 static int writecache_alloc_entries(struct dm_writecache *wc)
 944 {
 945         size_t b;
 946
 947         if (wc->entries)
 948                 return 0;
 949         wc->entries = vmalloc(array_size(sizeof(struct wc_entry), wc->n_blocks));
 950         if (!wc->entries)
 951                 return -ENOMEM;
 952         for (b = 0; b < wc->n_blocks; b++) {
 953                 struct wc_entry *e = &wc->entries[b];
 954                 e->index = b;
 955                 e->write_in_progress = false;
 956                 cond_resched();
 957         }
 958
 959         return 0;
 960 }
 961
 962 static int writecache_read_metadata(struct dm_writecache *wc, sector_t n_sectors)
 963 {
 964         struct dm_io_region region;
 965         struct dm_io_request req;
 966
 967         region.bdev = wc->ssd_dev->bdev;
 968         region.sector = wc->start_sector;
 969         region.count = n_sectors;
 970         req.bi_op = REQ_OP_READ;
 971         req.bi_op_flags = REQ_SYNC;
 972         req.mem.type = DM_IO_VMA;
 973         req.mem.ptr.vma = (char *)wc->memory_map;
 974         req.client = wc->dm_io;
 975         req.notify.fn = NULL;
 976
 977         return dm_io(&req, 1, &region, NULL);
 978 }
 979
 980 static void writecache_resume(struct dm_target *ti)
 981 {
 982         struct dm_writecache *wc = ti->private;
 983         size_t b;
 984         bool need_flush = false;
 985         __le64 sb_seq_count;
 986         int r;
 987
 988         wc_lock(wc);
 989
 990         wc->data_device_sectors = bdev_nr_sectors(wc->dev->bdev);
 991
 992         if (WC_MODE_PMEM(wc)) {
 993                 persistent_memory_invalidate_cache(wc->memory_map, wc->memory_map_size);
 994         } else {
 995                 r = writecache_read_metadata(wc, wc->metadata_sectors);
 996                 if (r) {
 997                         size_t sb_entries_offset;
 998                         writecache_error(wc, r, "unable to read metadata: %d", r);
 999                         sb_entries_offset = offsetof(struct wc_memory_superblock, entries);
1000                         memset((char *)wc->memory_map + sb_entries_offset, -1,
1001                                (wc->metadata_sectors << SECTOR_SHIFT) - sb_entries_offset);
1002                 }
1003         }
1004
1005         wc->tree = RB_ROOT;
1006         INIT_LIST_HEAD(&wc->lru);
1007         if (WC_MODE_SORT_FREELIST(wc)) {
1008                 wc->freetree = RB_ROOT;
1009                 wc->current_free = NULL;
1010         } else {
1011                 INIT_LIST_HEAD(&wc->freelist);
1012         }
1013         wc->freelist_size = 0;
1014
1015         r = copy_mc_to_kernel(&sb_seq_count, &sb(wc)->seq_count,
1016                               sizeof(uint64_t));
1017         if (r) {
1018                 writecache_error(wc, r, "hardware memory error when reading superblock: %d", r);
1019                 sb_seq_count = cpu_to_le64(0);
1020         }
1021         wc->seq_count = le64_to_cpu(sb_seq_count);
1022
1023 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
1024         for (b = 0; b < wc->n_blocks; b++) {
1025                 struct wc_entry *e = &wc->entries[b];
1026                 struct wc_memory_entry wme;
1027                 if (writecache_has_error(wc)) {
1028                         e->original_sector = -1;
1029                         e->seq_count = -1;
1030                         continue;
1031                 }
1032                 r = copy_mc_to_kernel(&wme, memory_entry(wc, e),
1033                                       sizeof(struct wc_memory_entry));
1034                 if (r) {
1035                         writecache_error(wc, r, "hardware memory error when reading metadata entry %lu: %d",
1036                                          (unsigned long)b, r);
1037                         e->original_sector = -1;
1038                         e->seq_count = -1;
1039                 } else {
1040                         e->original_sector = le64_to_cpu(wme.original_sector);
1041                         e->seq_count = le64_to_cpu(wme.seq_count);
1042                 }
1043                 cond_resched();
1044         }
1045 #endif
1046         for (b = 0; b < wc->n_blocks; b++) {
1047                 struct wc_entry *e = &wc->entries[b];
1048                 if (!writecache_entry_is_committed(wc, e)) {
1049                         if (read_seq_count(wc, e) != -1) {
1050 erase_this:
1051                                 clear_seq_count(wc, e);
1052                                 need_flush = true;
1053                         }
1054                         writecache_add_to_freelist(wc, e);
1055                 } else {
1056                         struct wc_entry *old;
1057
1058                         old = writecache_find_entry(wc, read_original_sector(wc, e), 0);
1059                         if (!old) {
1060                                 writecache_insert_entry(wc, e);
1061                         } else {
1062                                 if (read_seq_count(wc, old) == read_seq_count(wc, e)) {
1063                                         writecache_error(wc, -EINVAL,
1064                                                  "two identical entries, position %llu, sector %llu, sequence %llu",
1065                                                  (unsigned long long)b, (unsigned long long)read_original_sector(wc, e),
1066                                                  (unsigned long long)read_seq_count(wc, e));
1067                                 }
1068                                 if (read_seq_count(wc, old) > read_seq_count(wc, e)) {
1069                                         goto erase_this;
1070                                 } else {
1071                                         writecache_free_entry(wc, old);
1072                                         writecache_insert_entry(wc, e);
1073                                         need_flush = true;
1074                                 }
1075                         }
1076                 }
1077                 cond_resched();
1078         }
1079
1080         if (need_flush) {
1081                 writecache_flush_all_metadata(wc);
1082                 writecache_commit_flushed(wc, false);
1083         }
1084
1085         writecache_verify_watermark(wc);
1086
1087         if (wc->max_age != MAX_AGE_UNSPECIFIED)
1088                 mod_timer(&wc->max_age_timer, jiffies + wc->max_age / MAX_AGE_DIV);
1089
1090         wc_unlock(wc);
1091 }
1092
1093 static int process_flush_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1094 {
1095         if (argc != 1)
1096                 return -EINVAL;
1097
1098         wc_lock(wc);
1099         if (dm_suspended(wc->ti)) {
1100                 wc_unlock(wc);
1101                 return -EBUSY;
1102         }
1103         if (writecache_has_error(wc)) {
1104                 wc_unlock(wc);
1105                 return -EIO;
1106         }
1107
1108         writecache_flush(wc);
1109         wc->writeback_all++;
1110         queue_work(wc->writeback_wq, &wc->writeback_work);
1111         wc_unlock(wc);
1112
1113         flush_workqueue(wc->writeback_wq);
1114
1115         wc_lock(wc);
1116         wc->writeback_all--;
1117         if (writecache_has_error(wc)) {
1118                 wc_unlock(wc);
1119                 return -EIO;
1120         }
1121         wc_unlock(wc);
1122
1123         return 0;
1124 }
1125
1126 static int process_flush_on_suspend_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1127 {
1128         if (argc != 1)
1129                 return -EINVAL;
1130
1131         wc_lock(wc);
1132         wc->flush_on_suspend = true;
1133         wc_unlock(wc);
1134
1135         return 0;
1136 }
1137
1138 static void activate_cleaner(struct dm_writecache *wc)
1139 {
1140         wc->flush_on_suspend = true;
1141         wc->cleaner = true;
1142         wc->freelist_high_watermark = wc->n_blocks;
1143         wc->freelist_low_watermark = wc->n_blocks;
1144 }
1145
1146 static int process_cleaner_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1147 {
1148         if (argc != 1)
1149                 return -EINVAL;
1150
1151         wc_lock(wc);
1152         activate_cleaner(wc);
1153         if (!dm_suspended(wc->ti))
1154                 writecache_verify_watermark(wc);
1155         wc_unlock(wc);
1156
1157         return 0;
1158 }
1159
1160 static int writecache_message(struct dm_target *ti, unsigned argc, char **argv,
1161                               char *result, unsigned maxlen)
1162 {
1163         int r = -EINVAL;
1164         struct dm_writecache *wc = ti->private;
1165
1166         if (!strcasecmp(argv[0], "flush"))
1167                 r = process_flush_mesg(argc, argv, wc);
1168         else if (!strcasecmp(argv[0], "flush_on_suspend"))
1169                 r = process_flush_on_suspend_mesg(argc, argv, wc);
1170         else if (!strcasecmp(argv[0], "cleaner"))
1171                 r = process_cleaner_mesg(argc, argv, wc);
1172         else
1173                 DMERR("unrecognised message received: %s", argv[0]);
1174
1175         return r;
1176 }
1177
1178 static void memcpy_flushcache_optimized(void *dest, void *source, size_t size)
1179 {
1180         /*
1181          * clflushopt performs better with block size 1024, 2048, 4096
1182          * non-temporal stores perform better with block size 512
1183          *
1184          * block size   512             1024            2048            4096
1185          * movnti       496 MB/s        642 MB/s        725 MB/s        744 MB/s
1186          * clflushopt   373 MB/s        688 MB/s        1.1 GB/s        1.2 GB/s
1187          *
1188          * We see that movnti performs better for 512-byte blocks, and
1189          * clflushopt performs better for 1024-byte and larger blocks. So, we
1190          * prefer clflushopt for sizes >= 768.
1191          *
1192          * NOTE: this happens to be the case now (with dm-writecache's single
1193          * threaded model) but re-evaluate this once memcpy_flushcache() is
1194          * enabled to use movdir64b which might invalidate this performance
1195          * advantage seen with cache-allocating-writes plus flushing.
1196          */
1197 #ifdef CONFIG_X86
1198         if (static_cpu_has(X86_FEATURE_CLFLUSHOPT) &&
1199             likely(boot_cpu_data.x86_clflush_size == 64) &&
1200             likely(size >= 768)) {
1201                 do {
1202                         memcpy((void *)dest, (void *)source, 64);
1203                         clflushopt((void *)dest);
1204                         dest += 64;
1205                         source += 64;
1206                         size -= 64;
1207                 } while (size >= 64);
1208                 return;
1209         }
1210 #endif
1211         memcpy_flushcache(dest, source, size);
1212 }
1213
1214 static void bio_copy_block(struct dm_writecache *wc, struct bio *bio, void *data)
1215 {
1216         void *buf;
1217         unsigned long flags;
1218         unsigned size;
1219         int rw = bio_data_dir(bio);
1220         unsigned remaining_size = wc->block_size;
1221
1222         do {
1223                 struct bio_vec bv = bio_iter_iovec(bio, bio->bi_iter);
1224                 buf = bvec_kmap_irq(&bv, &flags);
1225                 size = bv.bv_len;
1226                 if (unlikely(size > remaining_size))
1227                         size = remaining_size;
1228
1229                 if (rw == READ) {
1230                         int r;
1231                         r = copy_mc_to_kernel(buf, data, size);
1232                         flush_dcache_page(bio_page(bio));
1233                         if (unlikely(r)) {
1234                                 writecache_error(wc, r, "hardware memory error when reading data: %d", r);
1235                                 bio->bi_status = BLK_STS_IOERR;
1236                         }
1237                 } else {
1238                         flush_dcache_page(bio_page(bio));
1239                         memcpy_flushcache_optimized(data, buf, size);
1240                 }
1241
1242                 bvec_kunmap_irq(buf, &flags);
1243
1244                 data = (char *)data + size;
1245                 remaining_size -= size;
1246                 bio_advance(bio, size);
1247         } while (unlikely(remaining_size));
1248 }
1249
1250 static int writecache_flush_thread(void *data)
1251 {
1252         struct dm_writecache *wc = data;
1253
1254         while (1) {
1255                 struct bio *bio;
1256
1257                 wc_lock(wc);
1258                 bio = bio_list_pop(&wc->flush_list);
1259                 if (!bio) {
1260                         set_current_state(TASK_INTERRUPTIBLE);
1261                         wc_unlock(wc);
1262
1263                         if (unlikely(kthread_should_stop())) {
1264                                 set_current_state(TASK_RUNNING);
1265                                 break;
1266                         }
1267
1268                         schedule();
1269                         continue;
1270                 }
1271
1272                 if (bio_op(bio) == REQ_OP_DISCARD) {
1273                         writecache_discard(wc, bio->bi_iter.bi_sector,
1274                                            bio_end_sector(bio));
1275                         wc_unlock(wc);
1276                         bio_set_dev(bio, wc->dev->bdev);
1277                         submit_bio_noacct(bio);
1278                 } else {
1279                         writecache_flush(wc);
1280                         wc_unlock(wc);
1281                         if (writecache_has_error(wc))
1282                                 bio->bi_status = BLK_STS_IOERR;
1283                         bio_endio(bio);
1284                 }
1285         }
1286
1287         return 0;
1288 }
1289
1290 static void writecache_offload_bio(struct dm_writecache *wc, struct bio *bio)
1291 {
1292         if (bio_list_empty(&wc->flush_list))
1293                 wake_up_process(wc->flush_thread);
1294         bio_list_add(&wc->flush_list, bio);
1295 }
1296
1297 static int writecache_map(struct dm_target *ti, struct bio *bio)
1298 {
1299         struct wc_entry *e;
1300         struct dm_writecache *wc = ti->private;
1301
1302         bio->bi_private = NULL;
1303
1304         wc_lock(wc);
1305
1306         if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1307                 if (writecache_has_error(wc))
1308                         goto unlock_error;
1309                 if (WC_MODE_PMEM(wc)) {
1310                         writecache_flush(wc);
1311                         if (writecache_has_error(wc))
1312                                 goto unlock_error;
1313                         if (unlikely(wc->cleaner) || unlikely(wc->metadata_only))
1314                                 goto unlock_remap_origin;
1315                         goto unlock_submit;
1316                 } else {
1317                         if (dm_bio_get_target_bio_nr(bio))
1318                                 goto unlock_remap_origin;
1319                         writecache_offload_bio(wc, bio);
1320                         goto unlock_return;
1321                 }
1322         }
1323
1324         bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
1325
1326         if (unlikely((((unsigned)bio->bi_iter.bi_sector | bio_sectors(bio)) &
1327                                 (wc->block_size / 512 - 1)) != 0)) {
1328                 DMERR("I/O is not aligned, sector %llu, size %u, block size %u",
1329                       (unsigned long long)bio->bi_iter.bi_sector,
1330                       bio->bi_iter.bi_size, wc->block_size);
1331                 goto unlock_error;
1332         }
1333
1334         if (unlikely(bio_op(bio) == REQ_OP_DISCARD)) {
1335                 if (writecache_has_error(wc))
1336                         goto unlock_error;
1337                 if (WC_MODE_PMEM(wc)) {
1338                         writecache_discard(wc, bio->bi_iter.bi_sector, bio_end_sector(bio));
1339                         goto unlock_remap_origin;
1340                 } else {
1341                         writecache_offload_bio(wc, bio);
1342                         goto unlock_return;
1343                 }
1344         }
1345
1346         if (bio_data_dir(bio) == READ) {
1347 read_next_block:
1348                 e = writecache_find_entry(wc, bio->bi_iter.bi_sector, WFE_RETURN_FOLLOWING);
1349                 if (e && read_original_sector(wc, e) == bio->bi_iter.bi_sector) {
1350                         if (WC_MODE_PMEM(wc)) {
1351                                 bio_copy_block(wc, bio, memory_data(wc, e));
1352                                 if (bio->bi_iter.bi_size)
1353                                         goto read_next_block;
1354                                 goto unlock_submit;
1355                         } else {
1356                                 dm_accept_partial_bio(bio, wc->block_size >> SECTOR_SHIFT);
1357                                 bio_set_dev(bio, wc->ssd_dev->bdev);
1358                                 bio->bi_iter.bi_sector = cache_sector(wc, e);
1359                                 if (!writecache_entry_is_committed(wc, e))
1360                                         writecache_wait_for_ios(wc, WRITE);
1361                                 goto unlock_remap;
1362                         }
1363                 } else {
1364                         if (e) {
1365                                 sector_t next_boundary =
1366                                         read_original_sector(wc, e) - bio->bi_iter.bi_sector;
1367                                 if (next_boundary < bio->bi_iter.bi_size >> SECTOR_SHIFT) {
1368                                         dm_accept_partial_bio(bio, next_boundary);
1369                                 }
1370                         }
1371                         goto unlock_remap_origin;
1372                 }
1373         } else {
1374                 do {
1375                         bool found_entry = false;
1376                         bool search_used = false;
1377                         if (writecache_has_error(wc))
1378                                 goto unlock_error;
1379                         e = writecache_find_entry(wc, bio->bi_iter.bi_sector, 0);
1380                         if (e) {
1381                                 if (!writecache_entry_is_committed(wc, e)) {
1382                                         search_used = true;
1383                                         goto bio_copy;
1384                                 }
1385                                 if (!WC_MODE_PMEM(wc) && !e->write_in_progress) {
1386                                         wc->overwrote_committed = true;
1387                                         search_used = true;
1388                                         goto bio_copy;
1389                                 }
1390                                 found_entry = true;
1391                         } else {
1392                                 if (unlikely(wc->cleaner) ||
1393                                     (wc->metadata_only && !(bio->bi_opf & REQ_META)))
1394                                         goto direct_write;
1395                         }
1396                         e = writecache_pop_from_freelist(wc, (sector_t)-1);
1397                         if (unlikely(!e)) {
1398                                 if (!WC_MODE_PMEM(wc) && !found_entry) {
1399 direct_write:
1400                                         e = writecache_find_entry(wc, bio->bi_iter.bi_sector, WFE_RETURN_FOLLOWING);
1401                                         if (e) {
1402                                                 sector_t next_boundary = read_original_sector(wc, e) - bio->bi_iter.bi_sector;
1403                                                 BUG_ON(!next_boundary);
1404                                                 if (next_boundary < bio->bi_iter.bi_size >> SECTOR_SHIFT) {
1405                                                         dm_accept_partial_bio(bio, next_boundary);
1406                                                 }
1407                                         }
1408                                         goto unlock_remap_origin;
1409                                 }
1410                                 writecache_wait_on_freelist(wc);
1411                                 continue;
1412                         }
1413                         write_original_sector_seq_count(wc, e, bio->bi_iter.bi_sector, wc->seq_count);
1414                         writecache_insert_entry(wc, e);
1415                         wc->uncommitted_blocks++;
1416 bio_copy:
1417                         if (WC_MODE_PMEM(wc)) {
1418                                 bio_copy_block(wc, bio, memory_data(wc, e));
1419                         } else {
1420                                 unsigned bio_size = wc->block_size;
1421                                 sector_t start_cache_sec = cache_sector(wc, e);
1422                                 sector_t current_cache_sec = start_cache_sec + (bio_size >> SECTOR_SHIFT);
1423
1424                                 while (bio_size < bio->bi_iter.bi_size) {
1425                                         if (!search_used) {
1426                                                 struct wc_entry *f = writecache_pop_from_freelist(wc, current_cache_sec);
1427                                                 if (!f)
1428                                                         break;
1429                                                 write_original_sector_seq_count(wc, f, bio->bi_iter.bi_sector +
1430                                                                                 (bio_size >> SECTOR_SHIFT), wc->seq_count);
1431                                                 writecache_insert_entry(wc, f);
1432                                                 wc->uncommitted_blocks++;
1433                                         } else {
1434                                                 struct wc_entry *f;
1435                                                 struct rb_node *next = rb_next(&e->rb_node);
1436                                                 if (!next)
1437                                                         break;
1438                                                 f = container_of(next, struct wc_entry, rb_node);
1439                                                 if (f != e + 1)
1440                                                         break;
1441                                                 if (read_original_sector(wc, f) !=
1442                                                     read_original_sector(wc, e) + (wc->block_size >> SECTOR_SHIFT))
1443                                                         break;
1444                                                 if (unlikely(f->write_in_progress))
1445                                                         break;
1446                                                 if (writecache_entry_is_committed(wc, f))
1447                                                         wc->overwrote_committed = true;
1448                                                 e = f;
1449                                         }
1450                                         bio_size += wc->block_size;
1451                                         current_cache_sec += wc->block_size >> SECTOR_SHIFT;
1452                                 }
1453
1454                                 bio_set_dev(bio, wc->ssd_dev->bdev);
1455                                 bio->bi_iter.bi_sector = start_cache_sec;
1456                                 dm_accept_partial_bio(bio, bio_size >> SECTOR_SHIFT);
1457
1458                                 if (unlikely(wc->uncommitted_blocks >= wc->autocommit_blocks)) {
1459                                         wc->uncommitted_blocks = 0;
1460                                         queue_work(wc->writeback_wq, &wc->flush_work);
1461                                 } else {
1462                                         writecache_schedule_autocommit(wc);
1463                                 }
1464                                 goto unlock_remap;
1465                         }
1466                 } while (bio->bi_iter.bi_size);
1467
1468                 if (unlikely(bio->bi_opf & REQ_FUA ||
1469                              wc->uncommitted_blocks >= wc->autocommit_blocks))
1470                         writecache_flush(wc);
1471                 else
1472                         writecache_schedule_autocommit(wc);
1473                 goto unlock_submit;
1474         }
1475
1476 unlock_remap_origin:
1477         if (likely(wc->pause != 0)) {
1478                  if (bio_op(bio) == REQ_OP_WRITE) {
1479                         dm_iot_io_begin(&wc->iot, 1);
1480                         bio->bi_private = (void *)2;
1481                 }
1482         }
1483         bio_set_dev(bio, wc->dev->bdev);
1484         wc_unlock(wc);
1485         return DM_MAPIO_REMAPPED;
1486
1487 unlock_remap:
1488         /* make sure that writecache_end_io decrements bio_in_progress: */
1489         bio->bi_private = (void *)1;
1490         atomic_inc(&wc->bio_in_progress[bio_data_dir(bio)]);
1491         wc_unlock(wc);
1492         return DM_MAPIO_REMAPPED;
1493
1494 unlock_submit:
1495         wc_unlock(wc);
1496         bio_endio(bio);
1497         return DM_MAPIO_SUBMITTED;
1498
1499 unlock_return:
1500         wc_unlock(wc);
1501         return DM_MAPIO_SUBMITTED;
1502
1503 unlock_error:
1504         wc_unlock(wc);
1505         bio_io_error(bio);
1506         return DM_MAPIO_SUBMITTED;
1507 }
1508
1509 static int writecache_end_io(struct dm_target *ti, struct bio *bio, blk_status_t *status)
1510 {
1511         struct dm_writecache *wc = ti->private;
1512
1513         if (bio->bi_private == (void *)1) {
1514                 int dir = bio_data_dir(bio);
1515                 if (atomic_dec_and_test(&wc->bio_in_progress[dir]))
1516                         if (unlikely(waitqueue_active(&wc->bio_in_progress_wait[dir])))
1517                                 wake_up(&wc->bio_in_progress_wait[dir]);
1518         } else if (bio->bi_private == (void *)2) {
1519                 dm_iot_io_end(&wc->iot, 1);
1520         }
1521         return 0;
1522 }
1523
1524 static int writecache_iterate_devices(struct dm_target *ti,
1525                                       iterate_devices_callout_fn fn, void *data)
1526 {
1527         struct dm_writecache *wc = ti->private;
1528
1529         return fn(ti, wc->dev, 0, ti->len, data);
1530 }
1531
1532 static void writecache_io_hints(struct dm_target *ti, struct queue_limits *limits)
1533 {
1534         struct dm_writecache *wc = ti->private;
1535
1536         if (limits->logical_block_size < wc->block_size)
1537                 limits->logical_block_size = wc->block_size;
1538
1539         if (limits->physical_block_size < wc->block_size)
1540                 limits->physical_block_size = wc->block_size;
1541
1542         if (limits->io_min < wc->block_size)
1543                 limits->io_min = wc->block_size;
1544 }
1545
1546
1547 static void writecache_writeback_endio(struct bio *bio)
1548 {
1549         struct writeback_struct *wb = container_of(bio, struct writeback_struct, bio);
1550         struct dm_writecache *wc = wb->wc;
1551         unsigned long flags;
1552
1553         raw_spin_lock_irqsave(&wc->endio_list_lock, flags);
1554         if (unlikely(list_empty(&wc->endio_list)))
1555                 wake_up_process(wc->endio_thread);
1556         list_add_tail(&wb->endio_entry, &wc->endio_list);
1557         raw_spin_unlock_irqrestore(&wc->endio_list_lock, flags);
1558 }
1559
1560 static void writecache_copy_endio(int read_err, unsigned long write_err, void *ptr)
1561 {
1562         struct copy_struct *c = ptr;
1563         struct dm_writecache *wc = c->wc;
1564
1565         c->error = likely(!(read_err | write_err)) ? 0 : -EIO;
1566
1567         raw_spin_lock_irq(&wc->endio_list_lock);
1568         if (unlikely(list_empty(&wc->endio_list)))
1569                 wake_up_process(wc->endio_thread);
1570         list_add_tail(&c->endio_entry, &wc->endio_list);
1571         raw_spin_unlock_irq(&wc->endio_list_lock);
1572 }
1573
1574 static void __writecache_endio_pmem(struct dm_writecache *wc, struct list_head *list)
1575 {
1576         unsigned i;
1577         struct writeback_struct *wb;
1578         struct wc_entry *e;
1579         unsigned long n_walked = 0;
1580
1581         do {
1582                 wb = list_entry(list->next, struct writeback_struct, endio_entry);
1583                 list_del(&wb->endio_entry);
1584
1585                 if (unlikely(wb->bio.bi_status != BLK_STS_OK))
1586                         writecache_error(wc, blk_status_to_errno(wb->bio.bi_status),
1587                                         "write error %d", wb->bio.bi_status);
1588                 i = 0;
1589                 do {
1590                         e = wb->wc_list[i];
1591                         BUG_ON(!e->write_in_progress);
1592                         e->write_in_progress = false;
1593                         INIT_LIST_HEAD(&e->lru);
1594                         if (!writecache_has_error(wc))
1595                                 writecache_free_entry(wc, e);
1596                         BUG_ON(!wc->writeback_size);
1597                         wc->writeback_size--;
1598                         n_walked++;
1599                         if (unlikely(n_walked >= ENDIO_LATENCY)) {
1600                                 writecache_commit_flushed(wc, false);
1601                                 wc_unlock(wc);
1602                                 wc_lock(wc);
1603                                 n_walked = 0;
1604                         }
1605                 } while (++i < wb->wc_list_n);
1606
1607                 if (wb->wc_list != wb->wc_list_inline)
1608                         kfree(wb->wc_list);
1609                 bio_put(&wb->bio);
1610         } while (!list_empty(list));
1611 }
1612
1613 static void __writecache_endio_ssd(struct dm_writecache *wc, struct list_head *list)
1614 {
1615         struct copy_struct *c;
1616         struct wc_entry *e;
1617
1618         do {
1619                 c = list_entry(list->next, struct copy_struct, endio_entry);
1620                 list_del(&c->endio_entry);
1621
1622                 if (unlikely(c->error))
1623                         writecache_error(wc, c->error, "copy error");
1624
1625                 e = c->e;
1626                 do {
1627                         BUG_ON(!e->write_in_progress);
1628                         e->write_in_progress = false;
1629                         INIT_LIST_HEAD(&e->lru);
1630                         if (!writecache_has_error(wc))
1631                                 writecache_free_entry(wc, e);
1632
1633                         BUG_ON(!wc->writeback_size);
1634                         wc->writeback_size--;
1635                         e++;
1636                 } while (--c->n_entries);
1637                 mempool_free(c, &wc->copy_pool);
1638         } while (!list_empty(list));
1639 }
1640
1641 static int writecache_endio_thread(void *data)
1642 {
1643         struct dm_writecache *wc = data;
1644
1645         while (1) {
1646                 struct list_head list;
1647
1648                 raw_spin_lock_irq(&wc->endio_list_lock);
1649                 if (!list_empty(&wc->endio_list))
1650                         goto pop_from_list;
1651                 set_current_state(TASK_INTERRUPTIBLE);
1652                 raw_spin_unlock_irq(&wc->endio_list_lock);
1653
1654                 if (unlikely(kthread_should_stop())) {
1655                         set_current_state(TASK_RUNNING);
1656                         break;
1657                 }
1658
1659                 schedule();
1660
1661                 continue;
1662
1663 pop_from_list:
1664                 list = wc->endio_list;
1665                 list.next->prev = list.prev->next = &list;
1666                 INIT_LIST_HEAD(&wc->endio_list);
1667                 raw_spin_unlock_irq(&wc->endio_list_lock);
1668
1669                 if (!WC_MODE_FUA(wc))
1670                         writecache_disk_flush(wc, wc->dev);
1671
1672                 wc_lock(wc);
1673
1674                 if (WC_MODE_PMEM(wc)) {
1675                         __writecache_endio_pmem(wc, &list);
1676                 } else {
1677                         __writecache_endio_ssd(wc, &list);
1678                         writecache_wait_for_ios(wc, READ);
1679                 }
1680
1681                 writecache_commit_flushed(wc, false);
1682
1683                 wc_unlock(wc);
1684         }
1685
1686         return 0;
1687 }
1688
1689 static bool wc_add_block(struct writeback_struct *wb, struct wc_entry *e)
1690 {
1691         struct dm_writecache *wc = wb->wc;
1692         unsigned block_size = wc->block_size;
1693         void *address = memory_data(wc, e);
1694
1695         persistent_memory_flush_cache(address, block_size);
1696
1697         if (unlikely(bio_end_sector(&wb->bio) >= wc->data_device_sectors))
1698                 return true;
1699
1700         return bio_add_page(&wb->bio, persistent_memory_page(address),
1701                             block_size, persistent_memory_page_offset(address)) != 0;
1702 }
1703
1704 struct writeback_list {
1705         struct list_head list;
1706         size_t size;
1707 };
1708
1709 static void __writeback_throttle(struct dm_writecache *wc, struct writeback_list *wbl)
1710 {
1711         if (unlikely(wc->max_writeback_jobs)) {
1712                 if (READ_ONCE(wc->writeback_size) - wbl->size >= wc->max_writeback_jobs) {
1713                         wc_lock(wc);
1714                         while (wc->writeback_size - wbl->size >= wc->max_writeback_jobs)
1715                                 writecache_wait_on_freelist(wc);
1716                         wc_unlock(wc);
1717                 }
1718         }
1719         cond_resched();
1720 }
1721
1722 static void __writecache_writeback_pmem(struct dm_writecache *wc, struct writeback_list *wbl)
1723 {
1724         struct wc_entry *e, *f;
1725         struct bio *bio;
1726         struct writeback_struct *wb;
1727         unsigned max_pages;
1728
1729         while (wbl->size) {
1730                 wbl->size--;
1731                 e = container_of(wbl->list.prev, struct wc_entry, lru);
1732                 list_del(&e->lru);
1733
1734                 max_pages = e->wc_list_contiguous;
1735
1736                 bio = bio_alloc_bioset(GFP_NOIO, max_pages, &wc->bio_set);
1737                 wb = container_of(bio, struct writeback_struct, bio);
1738                 wb->wc = wc;
1739                 bio->bi_end_io = writecache_writeback_endio;
1740                 bio_set_dev(bio, wc->dev->bdev);
1741                 bio->bi_iter.bi_sector = read_original_sector(wc, e);
1742                 if (max_pages <= WB_LIST_INLINE ||
1743                     unlikely(!(wb->wc_list = kmalloc_array(max_pages, sizeof(struct wc_entry *),
1744                                                            GFP_NOIO | __GFP_NORETRY |
1745                                                            __GFP_NOMEMALLOC | __GFP_NOWARN)))) {
1746                         wb->wc_list = wb->wc_list_inline;
1747                         max_pages = WB_LIST_INLINE;
1748                 }
1749
1750                 BUG_ON(!wc_add_block(wb, e));
1751
1752                 wb->wc_list[0] = e;
1753                 wb->wc_list_n = 1;
1754
1755                 while (wbl->size && wb->wc_list_n < max_pages) {
1756                         f = container_of(wbl->list.prev, struct wc_entry, lru);
1757                         if (read_original_sector(wc, f) !=
1758                             read_original_sector(wc, e) + (wc->block_size >> SECTOR_SHIFT))
1759                                 break;
1760                         if (!wc_add_block(wb, f))
1761                                 break;
1762                         wbl->size--;
1763                         list_del(&f->lru);
1764                         wb->wc_list[wb->wc_list_n++] = f;
1765                         e = f;
1766                 }
1767                 bio_set_op_attrs(bio, REQ_OP_WRITE, WC_MODE_FUA(wc) * REQ_FUA);
1768                 if (writecache_has_error(wc)) {
1769                         bio->bi_status = BLK_STS_IOERR;
1770                         bio_endio(bio);
1771                 } else if (unlikely(!bio_sectors(bio))) {
1772                         bio->bi_status = BLK_STS_OK;
1773                         bio_endio(bio);
1774                 } else {
1775                         submit_bio(bio);
1776                 }
1777
1778                 __writeback_throttle(wc, wbl);
1779         }
1780 }
1781
1782 static void __writecache_writeback_ssd(struct dm_writecache *wc, struct writeback_list *wbl)
1783 {
1784         struct wc_entry *e, *f;
1785         struct dm_io_region from, to;
1786         struct copy_struct *c;
1787
1788         while (wbl->size) {
1789                 unsigned n_sectors;
1790
1791                 wbl->size--;
1792                 e = container_of(wbl->list.prev, struct wc_entry, lru);
1793                 list_del(&e->lru);
1794
1795                 n_sectors = e->wc_list_contiguous << (wc->block_size_bits - SECTOR_SHIFT);
1796
1797                 from.bdev = wc->ssd_dev->bdev;
1798                 from.sector = cache_sector(wc, e);
1799                 from.count = n_sectors;
1800                 to.bdev = wc->dev->bdev;
1801                 to.sector = read_original_sector(wc, e);
1802                 to.count = n_sectors;
1803
1804                 c = mempool_alloc(&wc->copy_pool, GFP_NOIO);
1805                 c->wc = wc;
1806                 c->e = e;
1807                 c->n_entries = e->wc_list_contiguous;
1808
1809                 while ((n_sectors -= wc->block_size >> SECTOR_SHIFT)) {
1810                         wbl->size--;
1811                         f = container_of(wbl->list.prev, struct wc_entry, lru);
1812                         BUG_ON(f != e + 1);
1813                         list_del(&f->lru);
1814                         e = f;
1815                 }
1816
1817                 if (unlikely(to.sector + to.count > wc->data_device_sectors)) {
1818                         if (to.sector >= wc->data_device_sectors) {
1819                                 writecache_copy_endio(0, 0, c);
1820                                 continue;
1821                         }
1822                         from.count = to.count = wc->data_device_sectors - to.sector;
1823                 }
1824
1825                 dm_kcopyd_copy(wc->dm_kcopyd, &from, 1, &to, 0, writecache_copy_endio, c);
1826
1827                 __writeback_throttle(wc, wbl);
1828         }
1829 }
1830
1831 static void writecache_writeback(struct work_struct *work)
1832 {
1833         struct dm_writecache *wc = container_of(work, struct dm_writecache, writeback_work);
1834         struct blk_plug plug;
1835         struct wc_entry *f, *g, *e = NULL;
1836         struct rb_node *node, *next_node;
1837         struct list_head skipped;
1838         struct writeback_list wbl;
1839         unsigned long n_walked;
1840
1841         if (!WC_MODE_PMEM(wc)) {
1842                 /* Wait for any active kcopyd work on behalf of ssd writeback */
1843                 dm_kcopyd_client_flush(wc->dm_kcopyd);
1844         }
1845
1846         if (likely(wc->pause != 0)) {
1847                 while (1) {
1848                         unsigned long idle;
1849                         if (unlikely(wc->cleaner) || unlikely(wc->writeback_all) ||
1850                             unlikely(dm_suspended(wc->ti)))
1851                                 break;
1852                         idle = dm_iot_idle_time(&wc->iot);
1853                         if (idle >= wc->pause)
1854                                 break;
1855                         idle = wc->pause - idle;
1856                         if (idle > HZ)
1857                                 idle = HZ;
1858                         schedule_timeout_idle(idle);
1859                 }
1860         }
1861
1862         wc_lock(wc);
1863 restart:
1864         if (writecache_has_error(wc)) {
1865                 wc_unlock(wc);
1866                 return;
1867         }
1868
1869         if (unlikely(wc->writeback_all)) {
1870                 if (writecache_wait_for_writeback(wc))
1871                         goto restart;
1872         }
1873
1874         if (wc->overwrote_committed) {
1875                 writecache_wait_for_ios(wc, WRITE);
1876         }
1877
1878         n_walked = 0;
1879         INIT_LIST_HEAD(&skipped);
1880         INIT_LIST_HEAD(&wbl.list);
1881         wbl.size = 0;
1882         while (!list_empty(&wc->lru) &&
1883                (wc->writeback_all ||
1884                 wc->freelist_size + wc->writeback_size <= wc->freelist_low_watermark ||
1885                 (jiffies - container_of(wc->lru.prev, struct wc_entry, lru)->age >=
1886                  wc->max_age - wc->max_age / MAX_AGE_DIV))) {
1887
1888                 n_walked++;
1889                 if (unlikely(n_walked > WRITEBACK_LATENCY) &&
1890                     likely(!wc->writeback_all)) {
1891                         if (likely(!dm_suspended(wc->ti)))
1892                                 queue_work(wc->writeback_wq, &wc->writeback_work);
1893                         break;
1894                 }
1895
1896                 if (unlikely(wc->writeback_all)) {
1897                         if (unlikely(!e)) {
1898                                 writecache_flush(wc);
1899                                 e = container_of(rb_first(&wc->tree), struct wc_entry, rb_node);
1900                         } else
1901                                 e = g;
1902                 } else
1903                         e = container_of(wc->lru.prev, struct wc_entry, lru);
1904                 BUG_ON(e->write_in_progress);
1905                 if (unlikely(!writecache_entry_is_committed(wc, e))) {
1906                         writecache_flush(wc);
1907                 }
1908                 node = rb_prev(&e->rb_node);
1909                 if (node) {
1910                         f = container_of(node, struct wc_entry, rb_node);
1911                         if (unlikely(read_original_sector(wc, f) ==
1912                                      read_original_sector(wc, e))) {
1913                                 BUG_ON(!f->write_in_progress);
1914                                 list_move(&e->lru, &skipped);
1915                                 cond_resched();
1916                                 continue;
1917                         }
1918                 }
1919                 wc->writeback_size++;
1920                 list_move(&e->lru, &wbl.list);
1921                 wbl.size++;
1922                 e->write_in_progress = true;
1923                 e->wc_list_contiguous = 1;
1924
1925                 f = e;
1926
1927                 while (1) {
1928                         next_node = rb_next(&f->rb_node);
1929                         if (unlikely(!next_node))
1930                                 break;
1931                         g = container_of(next_node, struct wc_entry, rb_node);
1932                         if (unlikely(read_original_sector(wc, g) ==
1933                             read_original_sector(wc, f))) {
1934                                 f = g;
1935                                 continue;
1936                         }
1937                         if (read_original_sector(wc, g) !=
1938                             read_original_sector(wc, f) + (wc->block_size >> SECTOR_SHIFT))
1939                                 break;
1940                         if (unlikely(g->write_in_progress))
1941                                 break;
1942                         if (unlikely(!writecache_entry_is_committed(wc, g)))
1943                                 break;
1944
1945                         if (!WC_MODE_PMEM(wc)) {
1946                                 if (g != f + 1)
1947                                         break;
1948                         }
1949
1950                         n_walked++;
1951                         //if (unlikely(n_walked > WRITEBACK_LATENCY) && likely(!wc->writeback_all))
1952                         //      break;
1953
1954                         wc->writeback_size++;
1955                         list_move(&g->lru, &wbl.list);
1956                         wbl.size++;
1957                         g->write_in_progress = true;
1958                         g->wc_list_contiguous = BIO_MAX_VECS;
1959                         f = g;
1960                         e->wc_list_contiguous++;
1961                         if (unlikely(e->wc_list_contiguous == BIO_MAX_VECS)) {
1962                                 if (unlikely(wc->writeback_all)) {
1963                                         next_node = rb_next(&f->rb_node);
1964                                         if (likely(next_node))
1965                                                 g = container_of(next_node, struct wc_entry, rb_node);
1966                                 }
1967                                 break;
1968                         }
1969                 }
1970                 cond_resched();
1971         }
1972
1973         if (!list_empty(&skipped)) {
1974                 list_splice_tail(&skipped, &wc->lru);
1975                 /*
1976                  * If we didn't do any progress, we must wait until some
1977                  * writeback finishes to avoid burning CPU in a loop
1978                  */
1979                 if (unlikely(!wbl.size))
1980                         writecache_wait_for_writeback(wc);
1981         }
1982
1983         wc_unlock(wc);
1984
1985         blk_start_plug(&plug);
1986
1987         if (WC_MODE_PMEM(wc))
1988                 __writecache_writeback_pmem(wc, &wbl);
1989         else
1990                 __writecache_writeback_ssd(wc, &wbl);
1991
1992         blk_finish_plug(&plug);
1993
1994         if (unlikely(wc->writeback_all)) {
1995                 wc_lock(wc);
1996                 while (writecache_wait_for_writeback(wc));
1997                 wc_unlock(wc);
1998         }
1999 }
2000
2001 static int calculate_memory_size(uint64_t device_size, unsigned block_size,
2002                                  size_t *n_blocks_p, size_t *n_metadata_blocks_p)
2003 {
2004         uint64_t n_blocks, offset;
2005         struct wc_entry e;
2006
2007         n_blocks = device_size;
2008         do_div(n_blocks, block_size + sizeof(struct wc_memory_entry));
2009
2010         while (1) {
2011                 if (!n_blocks)
2012                         return -ENOSPC;
2013                 /* Verify the following entries[n_blocks] won't overflow */
2014                 if (n_blocks >= ((size_t)-sizeof(struct wc_memory_superblock) /
2015                                  sizeof(struct wc_memory_entry)))
2016                         return -EFBIG;
2017                 offset = offsetof(struct wc_memory_superblock, entries[n_blocks]);
2018                 offset = (offset + block_size - 1) & ~(uint64_t)(block_size - 1);
2019                 if (offset + n_blocks * block_size <= device_size)
2020                         break;
2021                 n_blocks--;
2022         }
2023
2024         /* check if the bit field overflows */
2025         e.index = n_blocks;
2026         if (e.index != n_blocks)
2027                 return -EFBIG;
2028
2029         if (n_blocks_p)
2030                 *n_blocks_p = n_blocks;
2031         if (n_metadata_blocks_p)
2032                 *n_metadata_blocks_p = offset >> __ffs(block_size);
2033         return 0;
2034 }
2035
2036 static int init_memory(struct dm_writecache *wc)
2037 {
2038         size_t b;
2039         int r;
2040
2041         r = calculate_memory_size(wc->memory_map_size, wc->block_size, &wc->n_blocks, NULL);
2042         if (r)
2043                 return r;
2044
2045         r = writecache_alloc_entries(wc);
2046         if (r)
2047                 return r;
2048
2049         for (b = 0; b < ARRAY_SIZE(sb(wc)->padding); b++)
2050                 pmem_assign(sb(wc)->padding[b], cpu_to_le64(0));
2051         pmem_assign(sb(wc)->version, cpu_to_le32(MEMORY_SUPERBLOCK_VERSION));
2052         pmem_assign(sb(wc)->block_size, cpu_to_le32(wc->block_size));
2053         pmem_assign(sb(wc)->n_blocks, cpu_to_le64(wc->n_blocks));
2054         pmem_assign(sb(wc)->seq_count, cpu_to_le64(0));
2055
2056         for (b = 0; b < wc->n_blocks; b++) {
2057                 write_original_sector_seq_count(wc, &wc->entries[b], -1, -1);
2058                 cond_resched();
2059         }
2060
2061         writecache_flush_all_metadata(wc);
2062         writecache_commit_flushed(wc, false);
2063         pmem_assign(sb(wc)->magic, cpu_to_le32(MEMORY_SUPERBLOCK_MAGIC));
2064         writecache_flush_region(wc, &sb(wc)->magic, sizeof sb(wc)->magic);
2065         writecache_commit_flushed(wc, false);
2066
2067         return 0;
2068 }
2069
2070 static void writecache_dtr(struct dm_target *ti)
2071 {
2072         struct dm_writecache *wc = ti->private;
2073
2074         if (!wc)
2075                 return;
2076
2077         if (wc->endio_thread)
2078                 kthread_stop(wc->endio_thread);
2079
2080         if (wc->flush_thread)
2081                 kthread_stop(wc->flush_thread);
2082
2083         bioset_exit(&wc->bio_set);
2084
2085         mempool_exit(&wc->copy_pool);
2086
2087         if (wc->writeback_wq)
2088                 destroy_workqueue(wc->writeback_wq);
2089
2090         if (wc->dev)
2091                 dm_put_device(ti, wc->dev);
2092
2093         if (wc->ssd_dev)
2094                 dm_put_device(ti, wc->ssd_dev);
2095
2096         vfree(wc->entries);
2097
2098         if (wc->memory_map) {
2099                 if (WC_MODE_PMEM(wc))
2100                         persistent_memory_release(wc);
2101                 else
2102                         vfree(wc->memory_map);
2103         }
2104
2105         if (wc->dm_kcopyd)
2106                 dm_kcopyd_client_destroy(wc->dm_kcopyd);
2107
2108         if (wc->dm_io)
2109                 dm_io_client_destroy(wc->dm_io);
2110
2111         vfree(wc->dirty_bitmap);
2112
2113         kfree(wc);
2114 }
2115
2116 static int writecache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2117 {
2118         struct dm_writecache *wc;
2119         struct dm_arg_set as;
2120         const char *string;
2121         unsigned opt_params;
2122         size_t offset, data_size;
2123         int i, r;
2124         char dummy;
2125         int high_wm_percent = HIGH_WATERMARK;
2126         int low_wm_percent = LOW_WATERMARK;
2127         uint64_t x;
2128         struct wc_memory_superblock s;
2129
2130         static struct dm_arg _args[] = {
2131                 {0, 18, "Invalid number of feature args"},
2132         };
2133
2134         as.argc = argc;
2135         as.argv = argv;
2136
2137         wc = kzalloc(sizeof(struct dm_writecache), GFP_KERNEL);
2138         if (!wc) {
2139                 ti->error = "Cannot allocate writecache structure";
2140                 r = -ENOMEM;
2141                 goto bad;
2142         }
2143         ti->private = wc;
2144         wc->ti = ti;
2145
2146         mutex_init(&wc->lock);
2147         wc->max_age = MAX_AGE_UNSPECIFIED;
2148         writecache_poison_lists(wc);
2149         init_waitqueue_head(&wc->freelist_wait);
2150         timer_setup(&wc->autocommit_timer, writecache_autocommit_timer, 0);
2151         timer_setup(&wc->max_age_timer, writecache_max_age_timer, 0);
2152
2153         for (i = 0; i < 2; i++) {
2154                 atomic_set(&wc->bio_in_progress[i], 0);
2155                 init_waitqueue_head(&wc->bio_in_progress_wait[i]);
2156         }
2157
2158         wc->dm_io = dm_io_client_create();
2159         if (IS_ERR(wc->dm_io)) {
2160                 r = PTR_ERR(wc->dm_io);
2161                 ti->error = "Unable to allocate dm-io client";
2162                 wc->dm_io = NULL;
2163                 goto bad;
2164         }
2165
2166         wc->writeback_wq = alloc_workqueue("writecache-writeback", WQ_MEM_RECLAIM, 1);
2167         if (!wc->writeback_wq) {
2168                 r = -ENOMEM;
2169                 ti->error = "Could not allocate writeback workqueue";
2170                 goto bad;
2171         }
2172         INIT_WORK(&wc->writeback_work, writecache_writeback);
2173         INIT_WORK(&wc->flush_work, writecache_flush_work);
2174
2175         dm_iot_init(&wc->iot);
2176
2177         raw_spin_lock_init(&wc->endio_list_lock);
2178         INIT_LIST_HEAD(&wc->endio_list);
2179         wc->endio_thread = kthread_create(writecache_endio_thread, wc, "writecache_endio");
2180         if (IS_ERR(wc->endio_thread)) {
2181                 r = PTR_ERR(wc->endio_thread);
2182                 wc->endio_thread = NULL;
2183                 ti->error = "Couldn't spawn endio thread";
2184                 goto bad;
2185         }
2186         wake_up_process(wc->endio_thread);
2187
2188         /*
2189          * Parse the mode (pmem or ssd)
2190          */
2191         string = dm_shift_arg(&as);
2192         if (!string)
2193                 goto bad_arguments;
2194
2195         if (!strcasecmp(string, "s")) {
2196                 wc->pmem_mode = false;
2197         } else if (!strcasecmp(string, "p")) {
2198 #ifdef DM_WRITECACHE_HAS_PMEM
2199                 wc->pmem_mode = true;
2200                 wc->writeback_fua = true;
2201 #else
2202                 /*
2203                  * If the architecture doesn't support persistent memory or
2204                  * the kernel doesn't support any DAX drivers, this driver can
2205                  * only be used in SSD-only mode.
2206                  */
2207                 r = -EOPNOTSUPP;
2208                 ti->error = "Persistent memory or DAX not supported on this system";
2209                 goto bad;
2210 #endif
2211         } else {
2212                 goto bad_arguments;
2213         }
2214
2215         if (WC_MODE_PMEM(wc)) {
2216                 r = bioset_init(&wc->bio_set, BIO_POOL_SIZE,
2217                                 offsetof(struct writeback_struct, bio),
2218                                 BIOSET_NEED_BVECS);
2219                 if (r) {
2220                         ti->error = "Could not allocate bio set";
2221                         goto bad;
2222                 }
2223         } else {
2224                 wc->pause = PAUSE_WRITEBACK;
2225                 r = mempool_init_kmalloc_pool(&wc->copy_pool, 1, sizeof(struct copy_struct));
2226                 if (r) {
2227                         ti->error = "Could not allocate mempool";
2228                         goto bad;
2229                 }
2230         }
2231
2232         /*
2233          * Parse the origin data device
2234          */
2235         string = dm_shift_arg(&as);
2236         if (!string)
2237                 goto bad_arguments;
2238         r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->dev);
2239         if (r) {
2240                 ti->error = "Origin data device lookup failed";
2241                 goto bad;
2242         }
2243
2244         /*
2245          * Parse cache data device (be it pmem or ssd)
2246          */
2247         string = dm_shift_arg(&as);
2248         if (!string)
2249                 goto bad_arguments;
2250
2251         r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->ssd_dev);
2252         if (r) {
2253                 ti->error = "Cache data device lookup failed";
2254                 goto bad;
2255         }
2256         wc->memory_map_size = i_size_read(wc->ssd_dev->bdev->bd_inode);
2257
2258         /*
2259          * Parse the cache block size
2260          */
2261         string = dm_shift_arg(&as);
2262         if (!string)
2263                 goto bad_arguments;
2264         if (sscanf(string, "%u%c", &wc->block_size, &dummy) != 1 ||
2265             wc->block_size < 512 || wc->block_size > PAGE_SIZE ||
2266             (wc->block_size & (wc->block_size - 1))) {
2267                 r = -EINVAL;
2268                 ti->error = "Invalid block size";
2269                 goto bad;
2270         }
2271         if (wc->block_size < bdev_logical_block_size(wc->dev->bdev) ||
2272             wc->block_size < bdev_logical_block_size(wc->ssd_dev->bdev)) {
2273                 r = -EINVAL;
2274                 ti->error = "Block size is smaller than device logical block size";
2275                 goto bad;
2276         }
2277         wc->block_size_bits = __ffs(wc->block_size);
2278
2279         wc->max_writeback_jobs = MAX_WRITEBACK_JOBS;
2280         wc->autocommit_blocks = !WC_MODE_PMEM(wc) ? AUTOCOMMIT_BLOCKS_SSD : AUTOCOMMIT_BLOCKS_PMEM;
2281         wc->autocommit_jiffies = msecs_to_jiffies(AUTOCOMMIT_MSEC);
2282
2283         /*
2284          * Parse optional arguments
2285          */
2286         r = dm_read_arg_group(_args, &as, &opt_params, &ti->error);
2287         if (r)
2288                 goto bad;
2289
2290         while (opt_params) {
2291                 string = dm_shift_arg(&as), opt_params--;
2292                 if (!strcasecmp(string, "start_sector") && opt_params >= 1) {
2293                         unsigned long long start_sector;
2294                         string = dm_shift_arg(&as), opt_params--;
2295                         if (sscanf(string, "%llu%c", &start_sector, &dummy) != 1)
2296                                 goto invalid_optional;
2297                         wc->start_sector = start_sector;
2298                         wc->start_sector_set = true;
2299                         if (wc->start_sector != start_sector ||
2300                             wc->start_sector >= wc->memory_map_size >> SECTOR_SHIFT)
2301                                 goto invalid_optional;
2302                 } else if (!strcasecmp(string, "high_watermark") && opt_params >= 1) {
2303                         string = dm_shift_arg(&as), opt_params--;
2304                         if (sscanf(string, "%d%c", &high_wm_percent, &dummy) != 1)
2305                                 goto invalid_optional;
2306                         if (high_wm_percent < 0 || high_wm_percent > 100)
2307                                 goto invalid_optional;
2308                         wc->high_wm_percent_value = high_wm_percent;
2309                         wc->high_wm_percent_set = true;
2310                 } else if (!strcasecmp(string, "low_watermark") && opt_params >= 1) {
2311                         string = dm_shift_arg(&as), opt_params--;
2312                         if (sscanf(string, "%d%c", &low_wm_percent, &dummy) != 1)
2313                                 goto invalid_optional;
2314                         if (low_wm_percent < 0 || low_wm_percent > 100)
2315                                 goto invalid_optional;
2316                         wc->low_wm_percent_value = low_wm_percent;
2317                         wc->low_wm_percent_set = true;
2318                 } else if (!strcasecmp(string, "writeback_jobs") && opt_params >= 1) {
2319                         string = dm_shift_arg(&as), opt_params--;
2320                         if (sscanf(string, "%u%c", &wc->max_writeback_jobs, &dummy) != 1)
2321                                 goto invalid_optional;
2322                         wc->max_writeback_jobs_set = true;
2323                 } else if (!strcasecmp(string, "autocommit_blocks") && opt_params >= 1) {
2324                         string = dm_shift_arg(&as), opt_params--;
2325                         if (sscanf(string, "%u%c", &wc->autocommit_blocks, &dummy) != 1)
2326                                 goto invalid_optional;
2327                         wc->autocommit_blocks_set = true;
2328                 } else if (!strcasecmp(string, "autocommit_time") && opt_params >= 1) {
2329                         unsigned autocommit_msecs;
2330                         string = dm_shift_arg(&as), opt_params--;
2331                         if (sscanf(string, "%u%c", &autocommit_msecs, &dummy) != 1)
2332                                 goto invalid_optional;
2333                         if (autocommit_msecs > 3600000)
2334                                 goto invalid_optional;
2335                         wc->autocommit_jiffies = msecs_to_jiffies(autocommit_msecs);
2336                         wc->autocommit_time_value = autocommit_msecs;
2337                         wc->autocommit_time_set = true;
2338                 } else if (!strcasecmp(string, "max_age") && opt_params >= 1) {
2339                         unsigned max_age_msecs;
2340                         string = dm_shift_arg(&as), opt_params--;
2341                         if (sscanf(string, "%u%c", &max_age_msecs, &dummy) != 1)
2342                                 goto invalid_optional;
2343                         if (max_age_msecs > 86400000)
2344                                 goto invalid_optional;
2345                         wc->max_age = msecs_to_jiffies(max_age_msecs);
2346                         wc->max_age_set = true;
2347                         wc->max_age_value = max_age_msecs;
2348                 } else if (!strcasecmp(string, "cleaner")) {
2349                         wc->cleaner_set = true;
2350                         wc->cleaner = true;
2351                 } else if (!strcasecmp(string, "fua")) {
2352                         if (WC_MODE_PMEM(wc)) {
2353                                 wc->writeback_fua = true;
2354                                 wc->writeback_fua_set = true;
2355                         } else goto invalid_optional;
2356                 } else if (!strcasecmp(string, "nofua")) {
2357                         if (WC_MODE_PMEM(wc)) {
2358                                 wc->writeback_fua = false;
2359                                 wc->writeback_fua_set = true;
2360                         } else goto invalid_optional;
2361                 } else if (!strcasecmp(string, "metadata_only")) {
2362                         wc->metadata_only = true;
2363                 } else if (!strcasecmp(string, "pause_writeback") && opt_params >= 1) {
2364                         unsigned pause_msecs;
2365                         if (WC_MODE_PMEM(wc))
2366                                 goto invalid_optional;
2367                         string = dm_shift_arg(&as), opt_params--;
2368                         if (sscanf(string, "%u%c", &pause_msecs, &dummy) != 1)
2369                                 goto invalid_optional;
2370                         if (pause_msecs > 60000)
2371                                 goto invalid_optional;
2372                         wc->pause = msecs_to_jiffies(pause_msecs);
2373                         wc->pause_set = true;
2374                         wc->pause_value = pause_msecs;
2375                 } else {
2376 invalid_optional:
2377                         r = -EINVAL;
2378                         ti->error = "Invalid optional argument";
2379                         goto bad;
2380                 }
2381         }
2382
2383         if (high_wm_percent < low_wm_percent) {
2384                 r = -EINVAL;
2385                 ti->error = "High watermark must be greater than or equal to low watermark";
2386                 goto bad;
2387         }
2388
2389         if (WC_MODE_PMEM(wc)) {
2390                 if (!dax_synchronous(wc->ssd_dev->dax_dev)) {
2391                         r = -EOPNOTSUPP;
2392                         ti->error = "Asynchronous persistent memory not supported as pmem cache";
2393                         goto bad;
2394                 }
2395
2396                 r = persistent_memory_claim(wc);
2397                 if (r) {
2398                         ti->error = "Unable to map persistent memory for cache";
2399                         goto bad;
2400                 }
2401         } else {
2402                 size_t n_blocks, n_metadata_blocks;
2403                 uint64_t n_bitmap_bits;
2404
2405                 wc->memory_map_size -= (uint64_t)wc->start_sector << SECTOR_SHIFT;
2406
2407                 bio_list_init(&wc->flush_list);
2408                 wc->flush_thread = kthread_create(writecache_flush_thread, wc, "dm_writecache_flush");
2409                 if (IS_ERR(wc->flush_thread)) {
2410                         r = PTR_ERR(wc->flush_thread);
2411                         wc->flush_thread = NULL;
2412                         ti->error = "Couldn't spawn flush thread";
2413                         goto bad;
2414                 }
2415                 wake_up_process(wc->flush_thread);
2416
2417                 r = calculate_memory_size(wc->memory_map_size, wc->block_size,
2418                                           &n_blocks, &n_metadata_blocks);
2419                 if (r) {
2420                         ti->error = "Invalid device size";
2421                         goto bad;
2422                 }
2423
2424                 n_bitmap_bits = (((uint64_t)n_metadata_blocks << wc->block_size_bits) +
2425                                  BITMAP_GRANULARITY - 1) / BITMAP_GRANULARITY;
2426                 /* this is limitation of test_bit functions */
2427                 if (n_bitmap_bits > 1U << 31) {
2428                         r = -EFBIG;
2429                         ti->error = "Invalid device size";
2430                         goto bad;
2431                 }
2432
2433                 wc->memory_map = vmalloc(n_metadata_blocks << wc->block_size_bits);
2434                 if (!wc->memory_map) {
2435                         r = -ENOMEM;
2436                         ti->error = "Unable to allocate memory for metadata";
2437                         goto bad;
2438                 }
2439
2440                 wc->dm_kcopyd = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2441                 if (IS_ERR(wc->dm_kcopyd)) {
2442                         r = PTR_ERR(wc->dm_kcopyd);
2443                         ti->error = "Unable to allocate dm-kcopyd client";
2444                         wc->dm_kcopyd = NULL;
2445                         goto bad;
2446                 }
2447
2448                 wc->metadata_sectors = n_metadata_blocks << (wc->block_size_bits - SECTOR_SHIFT);
2449                 wc->dirty_bitmap_size = (n_bitmap_bits + BITS_PER_LONG - 1) /
2450                         BITS_PER_LONG * sizeof(unsigned long);
2451                 wc->dirty_bitmap = vzalloc(wc->dirty_bitmap_size);
2452                 if (!wc->dirty_bitmap) {
2453                         r = -ENOMEM;
2454                         ti->error = "Unable to allocate dirty bitmap";
2455                         goto bad;
2456                 }
2457
2458                 r = writecache_read_metadata(wc, wc->block_size >> SECTOR_SHIFT);
2459                 if (r) {
2460                         ti->error = "Unable to read first block of metadata";
2461                         goto bad;
2462                 }
2463         }
2464
2465         r = copy_mc_to_kernel(&s, sb(wc), sizeof(struct wc_memory_superblock));
2466         if (r) {
2467                 ti->error = "Hardware memory error when reading superblock";
2468                 goto bad;
2469         }
2470         if (!le32_to_cpu(s.magic) && !le32_to_cpu(s.version)) {
2471                 r = init_memory(wc);
2472                 if (r) {
2473                         ti->error = "Unable to initialize device";
2474                         goto bad;
2475                 }
2476                 r = copy_mc_to_kernel(&s, sb(wc),
2477                                       sizeof(struct wc_memory_superblock));
2478                 if (r) {
2479                         ti->error = "Hardware memory error when reading superblock";
2480                         goto bad;
2481                 }
2482         }
2483
2484         if (le32_to_cpu(s.magic) != MEMORY_SUPERBLOCK_MAGIC) {
2485                 ti->error = "Invalid magic in the superblock";
2486                 r = -EINVAL;
2487                 goto bad;
2488         }
2489
2490         if (le32_to_cpu(s.version) != MEMORY_SUPERBLOCK_VERSION) {
2491                 ti->error = "Invalid version in the superblock";
2492                 r = -EINVAL;
2493                 goto bad;
2494         }
2495
2496         if (le32_to_cpu(s.block_size) != wc->block_size) {
2497                 ti->error = "Block size does not match superblock";
2498                 r = -EINVAL;
2499                 goto bad;
2500         }
2501
2502         wc->n_blocks = le64_to_cpu(s.n_blocks);
2503
2504         offset = wc->n_blocks * sizeof(struct wc_memory_entry);
2505         if (offset / sizeof(struct wc_memory_entry) != le64_to_cpu(sb(wc)->n_blocks)) {
2506 overflow:
2507                 ti->error = "Overflow in size calculation";
2508                 r = -EINVAL;
2509                 goto bad;
2510         }
2511         offset += sizeof(struct wc_memory_superblock);
2512         if (offset < sizeof(struct wc_memory_superblock))
2513                 goto overflow;
2514         offset = (offset + wc->block_size - 1) & ~(size_t)(wc->block_size - 1);
2515         data_size = wc->n_blocks * (size_t)wc->block_size;
2516         if (!offset || (data_size / wc->block_size != wc->n_blocks) ||
2517             (offset + data_size < offset))
2518                 goto overflow;
2519         if (offset + data_size > wc->memory_map_size) {
2520                 ti->error = "Memory area is too small";
2521                 r = -EINVAL;
2522                 goto bad;
2523         }
2524
2525         wc->metadata_sectors = offset >> SECTOR_SHIFT;
2526         wc->block_start = (char *)sb(wc) + offset;
2527
2528         x = (uint64_t)wc->n_blocks * (100 - high_wm_percent);
2529         x += 50;
2530         do_div(x, 100);
2531         wc->freelist_high_watermark = x;
2532         x = (uint64_t)wc->n_blocks * (100 - low_wm_percent);
2533         x += 50;
2534         do_div(x, 100);
2535         wc->freelist_low_watermark = x;
2536
2537         if (wc->cleaner)
2538                 activate_cleaner(wc);
2539
2540         r = writecache_alloc_entries(wc);
2541         if (r) {
2542                 ti->error = "Cannot allocate memory";
2543                 goto bad;
2544         }
2545
2546         ti->num_flush_bios = WC_MODE_PMEM(wc) ? 1 : 2;
2547         ti->flush_supported = true;
2548         ti->num_discard_bios = 1;
2549
2550         if (WC_MODE_PMEM(wc))
2551                 persistent_memory_flush_cache(wc->memory_map, wc->memory_map_size);
2552
2553         return 0;
2554
2555 bad_arguments:
2556         r = -EINVAL;
2557         ti->error = "Bad arguments";
2558 bad:
2559         writecache_dtr(ti);
2560         return r;
2561 }
2562
2563 static void writecache_status(struct dm_target *ti, status_type_t type,
2564                               unsigned status_flags, char *result, unsigned maxlen)
2565 {
2566         struct dm_writecache *wc = ti->private;
2567         unsigned extra_args;
2568         unsigned sz = 0;
2569
2570         switch (type) {
2571         case STATUSTYPE_INFO:
2572                 DMEMIT("%ld %llu %llu %llu", writecache_has_error(wc),
2573                        (unsigned long long)wc->n_blocks, (unsigned long long)wc->freelist_size,
2574                        (unsigned long long)wc->writeback_size);
2575                 break;
2576         case STATUSTYPE_TABLE:
2577                 DMEMIT("%c %s %s %u ", WC_MODE_PMEM(wc) ? 'p' : 's',
2578                                 wc->dev->name, wc->ssd_dev->name, wc->block_size);
2579                 extra_args = 0;
2580                 if (wc->start_sector_set)
2581                         extra_args += 2;
2582                 if (wc->high_wm_percent_set)
2583                         extra_args += 2;
2584                 if (wc->low_wm_percent_set)
2585                         extra_args += 2;
2586                 if (wc->max_writeback_jobs_set)
2587                         extra_args += 2;
2588                 if (wc->autocommit_blocks_set)
2589                         extra_args += 2;
2590                 if (wc->autocommit_time_set)
2591                         extra_args += 2;
2592                 if (wc->max_age_set)
2593                         extra_args += 2;
2594                 if (wc->cleaner_set)
2595                         extra_args++;
2596                 if (wc->writeback_fua_set)
2597                         extra_args++;
2598                 if (wc->metadata_only)
2599                         extra_args++;
2600                 if (wc->pause_set)
2601                         extra_args += 2;
2602
2603                 DMEMIT("%u", extra_args);
2604                 if (wc->start_sector_set)
2605                         DMEMIT(" start_sector %llu", (unsigned long long)wc->start_sector);
2606                 if (wc->high_wm_percent_set)
2607                         DMEMIT(" high_watermark %u", wc->high_wm_percent_value);
2608                 if (wc->low_wm_percent_set)
2609                         DMEMIT(" low_watermark %u", wc->low_wm_percent_value);
2610                 if (wc->max_writeback_jobs_set)
2611                         DMEMIT(" writeback_jobs %u", wc->max_writeback_jobs);
2612                 if (wc->autocommit_blocks_set)
2613                         DMEMIT(" autocommit_blocks %u", wc->autocommit_blocks);
2614                 if (wc->autocommit_time_set)
2615                         DMEMIT(" autocommit_time %u", wc->autocommit_time_value);
2616                 if (wc->max_age_set)
2617                         DMEMIT(" max_age %u", wc->max_age_value);
2618                 if (wc->cleaner_set)
2619                         DMEMIT(" cleaner");
2620                 if (wc->writeback_fua_set)
2621                         DMEMIT(" %sfua", wc->writeback_fua ? "" : "no");
2622                 if (wc->metadata_only)
2623                         DMEMIT(" metadata_only");
2624                 if (wc->pause_set)
2625                         DMEMIT(" pause_writeback %u", wc->pause_value);
2626                 break;
2627         }
2628 }
2629
2630 static struct target_type writecache_target = {
2631         .name                   = "writecache",
2632         .version                = {1, 5, 0},
2633         .module                 = THIS_MODULE,
2634         .ctr                    = writecache_ctr,
2635         .dtr                    = writecache_dtr,
2636         .status                 = writecache_status,
2637         .postsuspend            = writecache_suspend,
2638         .resume                 = writecache_resume,
2639         .message                = writecache_message,
2640         .map                    = writecache_map,
2641         .end_io                 = writecache_end_io,
2642         .iterate_devices        = writecache_iterate_devices,
2643         .io_hints               = writecache_io_hints,
2644 };
2645
2646 static int __init dm_writecache_init(void)
2647 {
2648         int r;
2649
2650         r = dm_register_target(&writecache_target);
2651         if (r < 0) {
2652                 DMERR("register failed %d", r);
2653                 return r;
2654         }
2655
2656         return 0;
2657 }
2658
2659 static void __exit dm_writecache_exit(void)
2660 {
2661         dm_unregister_target(&writecache_target);
2662 }
2663
2664 module_init(dm_writecache_init);
2665 module_exit(dm_writecache_exit);
2666
2667 MODULE_DESCRIPTION(DM_NAME " writecache target");
2668 MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
2669 MODULE_LICENSE("GPL");