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              min(0x10000000 / PAGE_SIZE, totalram_pages() / 16)
  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         struct {
 211                 unsigned long long reads;
 212                 unsigned long long read_hits;
 213                 unsigned long long writes;
 214                 unsigned long long write_hits_uncommitted;
 215                 unsigned long long write_hits_committed;
 216                 unsigned long long writes_around;
 217                 unsigned long long writes_allocate;
 218                 unsigned long long writes_blocked_on_freelist;
 219                 unsigned long long flushes;
 220                 unsigned long long discards;
 221         } stats;
 222 };
 223
 224 #define WB_LIST_INLINE          16
 225
 226 struct writeback_struct {
 227         struct list_head endio_entry;
 228         struct dm_writecache *wc;
 229         struct wc_entry **wc_list;
 230         unsigned wc_list_n;
 231         struct wc_entry *wc_list_inline[WB_LIST_INLINE];
 232         struct bio bio;
 233 };
 234
 235 struct copy_struct {
 236         struct list_head endio_entry;
 237         struct dm_writecache *wc;
 238         struct wc_entry *e;
 239         unsigned n_entries;
 240         int error;
 241 };
 242
 243 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(dm_writecache_throttle,
 244                                             "A percentage of time allocated for data copying");
 245
 246 static void wc_lock(struct dm_writecache *wc)
 247 {
 248         mutex_lock(&wc->lock);
 249 }
 250
 251 static void wc_unlock(struct dm_writecache *wc)
 252 {
 253         mutex_unlock(&wc->lock);
 254 }
 255
 256 #ifdef DM_WRITECACHE_HAS_PMEM
 257 static int persistent_memory_claim(struct dm_writecache *wc)
 258 {
 259         int r;
 260         loff_t s;
 261         long p, da;
 262         pfn_t pfn;
 263         int id;
 264         struct page **pages;
 265         sector_t offset;
 266
 267         wc->memory_vmapped = false;
 268
 269         s = wc->memory_map_size;
 270         p = s >> PAGE_SHIFT;
 271         if (!p) {
 272                 r = -EINVAL;
 273                 goto err1;
 274         }
 275         if (p != s >> PAGE_SHIFT) {
 276                 r = -EOVERFLOW;
 277                 goto err1;
 278         }
 279
 280         offset = get_start_sect(wc->ssd_dev->bdev);
 281         if (offset & (PAGE_SIZE / 512 - 1)) {
 282                 r = -EINVAL;
 283                 goto err1;
 284         }
 285         offset >>= PAGE_SHIFT - 9;
 286
 287         id = dax_read_lock();
 288
 289         da = dax_direct_access(wc->ssd_dev->dax_dev, offset, p, &wc->memory_map, &pfn);
 290         if (da < 0) {
 291                 wc->memory_map = NULL;
 292                 r = da;
 293                 goto err2;
 294         }
 295         if (!pfn_t_has_page(pfn)) {
 296                 wc->memory_map = NULL;
 297                 r = -EOPNOTSUPP;
 298                 goto err2;
 299         }
 300         if (da != p) {
 301                 long i;
 302                 wc->memory_map = NULL;
 303                 pages = kvmalloc_array(p, sizeof(struct page *), GFP_KERNEL);
 304                 if (!pages) {
 305                         r = -ENOMEM;
 306                         goto err2;
 307                 }
 308                 i = 0;
 309                 do {
 310                         long daa;
 311                         daa = dax_direct_access(wc->ssd_dev->dax_dev, offset + i, p - i,
 312                                                 NULL, &pfn);
 313                         if (daa <= 0) {
 314                                 r = daa ? daa : -EINVAL;
 315                                 goto err3;
 316                         }
 317                         if (!pfn_t_has_page(pfn)) {
 318                                 r = -EOPNOTSUPP;
 319                                 goto err3;
 320                         }
 321                         while (daa-- && i < p) {
 322                                 pages[i++] = pfn_t_to_page(pfn);
 323                                 pfn.val++;
 324                                 if (!(i & 15))
 325                                         cond_resched();
 326                         }
 327                 } while (i < p);
 328                 wc->memory_map = vmap(pages, p, VM_MAP, PAGE_KERNEL);
 329                 if (!wc->memory_map) {
 330                         r = -ENOMEM;
 331                         goto err3;
 332                 }
 333                 kvfree(pages);
 334                 wc->memory_vmapped = true;
 335         }
 336
 337         dax_read_unlock(id);
 338
 339         wc->memory_map += (size_t)wc->start_sector << SECTOR_SHIFT;
 340         wc->memory_map_size -= (size_t)wc->start_sector << SECTOR_SHIFT;
 341
 342         return 0;
 343 err3:
 344         kvfree(pages);
 345 err2:
 346         dax_read_unlock(id);
 347 err1:
 348         return r;
 349 }
 350 #else
 351 static int persistent_memory_claim(struct dm_writecache *wc)
 352 {
 353         return -EOPNOTSUPP;
 354 }
 355 #endif
 356
 357 static void persistent_memory_release(struct dm_writecache *wc)
 358 {
 359         if (wc->memory_vmapped)
 360                 vunmap(wc->memory_map - ((size_t)wc->start_sector << SECTOR_SHIFT));
 361 }
 362
 363 static struct page *persistent_memory_page(void *addr)
 364 {
 365         if (is_vmalloc_addr(addr))
 366                 return vmalloc_to_page(addr);
 367         else
 368                 return virt_to_page(addr);
 369 }
 370
 371 static unsigned persistent_memory_page_offset(void *addr)
 372 {
 373         return (unsigned long)addr & (PAGE_SIZE - 1);
 374 }
 375
 376 static void persistent_memory_flush_cache(void *ptr, size_t size)
 377 {
 378         if (is_vmalloc_addr(ptr))
 379                 flush_kernel_vmap_range(ptr, size);
 380 }
 381
 382 static void persistent_memory_invalidate_cache(void *ptr, size_t size)
 383 {
 384         if (is_vmalloc_addr(ptr))
 385                 invalidate_kernel_vmap_range(ptr, size);
 386 }
 387
 388 static struct wc_memory_superblock *sb(struct dm_writecache *wc)
 389 {
 390         return wc->memory_map;
 391 }
 392
 393 static struct wc_memory_entry *memory_entry(struct dm_writecache *wc, struct wc_entry *e)
 394 {
 395         return &sb(wc)->entries[e->index];
 396 }
 397
 398 static void *memory_data(struct dm_writecache *wc, struct wc_entry *e)
 399 {
 400         return (char *)wc->block_start + (e->index << wc->block_size_bits);
 401 }
 402
 403 static sector_t cache_sector(struct dm_writecache *wc, struct wc_entry *e)
 404 {
 405         return wc->start_sector + wc->metadata_sectors +
 406                 ((sector_t)e->index << (wc->block_size_bits - SECTOR_SHIFT));
 407 }
 408
 409 static uint64_t read_original_sector(struct dm_writecache *wc, struct wc_entry *e)
 410 {
 411 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 412         return e->original_sector;
 413 #else
 414         return le64_to_cpu(memory_entry(wc, e)->original_sector);
 415 #endif
 416 }
 417
 418 static uint64_t read_seq_count(struct dm_writecache *wc, struct wc_entry *e)
 419 {
 420 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 421         return e->seq_count;
 422 #else
 423         return le64_to_cpu(memory_entry(wc, e)->seq_count);
 424 #endif
 425 }
 426
 427 static void clear_seq_count(struct dm_writecache *wc, struct wc_entry *e)
 428 {
 429 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 430         e->seq_count = -1;
 431 #endif
 432         pmem_assign(memory_entry(wc, e)->seq_count, cpu_to_le64(-1));
 433 }
 434
 435 static void write_original_sector_seq_count(struct dm_writecache *wc, struct wc_entry *e,
 436                                             uint64_t original_sector, uint64_t seq_count)
 437 {
 438         struct wc_memory_entry me;
 439 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 440         e->original_sector = original_sector;
 441         e->seq_count = seq_count;
 442 #endif
 443         me.original_sector = cpu_to_le64(original_sector);
 444         me.seq_count = cpu_to_le64(seq_count);
 445         pmem_assign(*memory_entry(wc, e), me);
 446 }
 447
 448 #define writecache_error(wc, err, msg, arg...)                          \
 449 do {                                                                    \
 450         if (!cmpxchg(&(wc)->error, 0, err))                             \
 451                 DMERR(msg, ##arg);                                      \
 452         wake_up(&(wc)->freelist_wait);                                  \
 453 } while (0)
 454
 455 #define writecache_has_error(wc)        (unlikely(READ_ONCE((wc)->error)))
 456
 457 static void writecache_flush_all_metadata(struct dm_writecache *wc)
 458 {
 459         if (!WC_MODE_PMEM(wc))
 460                 memset(wc->dirty_bitmap, -1, wc->dirty_bitmap_size);
 461 }
 462
 463 static void writecache_flush_region(struct dm_writecache *wc, void *ptr, size_t size)
 464 {
 465         if (!WC_MODE_PMEM(wc))
 466                 __set_bit(((char *)ptr - (char *)wc->memory_map) / BITMAP_GRANULARITY,
 467                           wc->dirty_bitmap);
 468 }
 469
 470 static void writecache_disk_flush(struct dm_writecache *wc, struct dm_dev *dev);
 471
 472 struct io_notify {
 473         struct dm_writecache *wc;
 474         struct completion c;
 475         atomic_t count;
 476 };
 477
 478 static void writecache_notify_io(unsigned long error, void *context)
 479 {
 480         struct io_notify *endio = context;
 481
 482         if (unlikely(error != 0))
 483                 writecache_error(endio->wc, -EIO, "error writing metadata");
 484         BUG_ON(atomic_read(&endio->count) <= 0);
 485         if (atomic_dec_and_test(&endio->count))
 486                 complete(&endio->c);
 487 }
 488
 489 static void writecache_wait_for_ios(struct dm_writecache *wc, int direction)
 490 {
 491         wait_event(wc->bio_in_progress_wait[direction],
 492                    !atomic_read(&wc->bio_in_progress[direction]));
 493 }
 494
 495 static void ssd_commit_flushed(struct dm_writecache *wc, bool wait_for_ios)
 496 {
 497         struct dm_io_region region;
 498         struct dm_io_request req;
 499         struct io_notify endio = {
 500                 wc,
 501                 COMPLETION_INITIALIZER_ONSTACK(endio.c),
 502                 ATOMIC_INIT(1),
 503         };
 504         unsigned bitmap_bits = wc->dirty_bitmap_size * 8;
 505         unsigned i = 0;
 506
 507         while (1) {
 508                 unsigned j;
 509                 i = find_next_bit(wc->dirty_bitmap, bitmap_bits, i);
 510                 if (unlikely(i == bitmap_bits))
 511                         break;
 512                 j = find_next_zero_bit(wc->dirty_bitmap, bitmap_bits, i);
 513
 514                 region.bdev = wc->ssd_dev->bdev;
 515                 region.sector = (sector_t)i * (BITMAP_GRANULARITY >> SECTOR_SHIFT);
 516                 region.count = (sector_t)(j - i) * (BITMAP_GRANULARITY >> SECTOR_SHIFT);
 517
 518                 if (unlikely(region.sector >= wc->metadata_sectors))
 519                         break;
 520                 if (unlikely(region.sector + region.count > wc->metadata_sectors))
 521                         region.count = wc->metadata_sectors - region.sector;
 522
 523                 region.sector += wc->start_sector;
 524                 atomic_inc(&endio.count);
 525                 req.bi_op = REQ_OP_WRITE;
 526                 req.bi_op_flags = REQ_SYNC;
 527                 req.mem.type = DM_IO_VMA;
 528                 req.mem.ptr.vma = (char *)wc->memory_map + (size_t)i * BITMAP_GRANULARITY;
 529                 req.client = wc->dm_io;
 530                 req.notify.fn = writecache_notify_io;
 531                 req.notify.context = &endio;
 532
 533                 /* writing via async dm-io (implied by notify.fn above) won't return an error */
 534                 (void) dm_io(&req, 1, &region, NULL);
 535                 i = j;
 536         }
 537
 538         writecache_notify_io(0, &endio);
 539         wait_for_completion_io(&endio.c);
 540
 541         if (wait_for_ios)
 542                 writecache_wait_for_ios(wc, WRITE);
 543
 544         writecache_disk_flush(wc, wc->ssd_dev);
 545
 546         memset(wc->dirty_bitmap, 0, wc->dirty_bitmap_size);
 547 }
 548
 549 static void ssd_commit_superblock(struct dm_writecache *wc)
 550 {
 551         int r;
 552         struct dm_io_region region;
 553         struct dm_io_request req;
 554
 555         region.bdev = wc->ssd_dev->bdev;
 556         region.sector = 0;
 557         region.count = max(4096U, wc->block_size) >> SECTOR_SHIFT;
 558
 559         if (unlikely(region.sector + region.count > wc->metadata_sectors))
 560                 region.count = wc->metadata_sectors - region.sector;
 561
 562         region.sector += wc->start_sector;
 563
 564         req.bi_op = REQ_OP_WRITE;
 565         req.bi_op_flags = REQ_SYNC | REQ_FUA;
 566         req.mem.type = DM_IO_VMA;
 567         req.mem.ptr.vma = (char *)wc->memory_map;
 568         req.client = wc->dm_io;
 569         req.notify.fn = NULL;
 570         req.notify.context = NULL;
 571
 572         r = dm_io(&req, 1, &region, NULL);
 573         if (unlikely(r))
 574                 writecache_error(wc, r, "error writing superblock");
 575 }
 576
 577 static void writecache_commit_flushed(struct dm_writecache *wc, bool wait_for_ios)
 578 {
 579         if (WC_MODE_PMEM(wc))
 580                 pmem_wmb();
 581         else
 582                 ssd_commit_flushed(wc, wait_for_ios);
 583 }
 584
 585 static void writecache_disk_flush(struct dm_writecache *wc, struct dm_dev *dev)
 586 {
 587         int r;
 588         struct dm_io_region region;
 589         struct dm_io_request req;
 590
 591         region.bdev = dev->bdev;
 592         region.sector = 0;
 593         region.count = 0;
 594         req.bi_op = REQ_OP_WRITE;
 595         req.bi_op_flags = REQ_PREFLUSH;
 596         req.mem.type = DM_IO_KMEM;
 597         req.mem.ptr.addr = NULL;
 598         req.client = wc->dm_io;
 599         req.notify.fn = NULL;
 600
 601         r = dm_io(&req, 1, &region, NULL);
 602         if (unlikely(r))
 603                 writecache_error(wc, r, "error flushing metadata: %d", r);
 604 }
 605
 606 #define WFE_RETURN_FOLLOWING    1
 607 #define WFE_LOWEST_SEQ          2
 608
 609 static struct wc_entry *writecache_find_entry(struct dm_writecache *wc,
 610                                               uint64_t block, int flags)
 611 {
 612         struct wc_entry *e;
 613         struct rb_node *node = wc->tree.rb_node;
 614
 615         if (unlikely(!node))
 616                 return NULL;
 617
 618         while (1) {
 619                 e = container_of(node, struct wc_entry, rb_node);
 620                 if (read_original_sector(wc, e) == block)
 621                         break;
 622
 623                 node = (read_original_sector(wc, e) >= block ?
 624                         e->rb_node.rb_left : e->rb_node.rb_right);
 625                 if (unlikely(!node)) {
 626                         if (!(flags & WFE_RETURN_FOLLOWING))
 627                                 return NULL;
 628                         if (read_original_sector(wc, e) >= block) {
 629                                 return e;
 630                         } else {
 631                                 node = rb_next(&e->rb_node);
 632                                 if (unlikely(!node))
 633                                         return NULL;
 634                                 e = container_of(node, struct wc_entry, rb_node);
 635                                 return e;
 636                         }
 637                 }
 638         }
 639
 640         while (1) {
 641                 struct wc_entry *e2;
 642                 if (flags & WFE_LOWEST_SEQ)
 643                         node = rb_prev(&e->rb_node);
 644                 else
 645                         node = rb_next(&e->rb_node);
 646                 if (unlikely(!node))
 647                         return e;
 648                 e2 = container_of(node, struct wc_entry, rb_node);
 649                 if (read_original_sector(wc, e2) != block)
 650                         return e;
 651                 e = e2;
 652         }
 653 }
 654
 655 static void writecache_insert_entry(struct dm_writecache *wc, struct wc_entry *ins)
 656 {
 657         struct wc_entry *e;
 658         struct rb_node **node = &wc->tree.rb_node, *parent = NULL;
 659
 660         while (*node) {
 661                 e = container_of(*node, struct wc_entry, rb_node);
 662                 parent = &e->rb_node;
 663                 if (read_original_sector(wc, e) > read_original_sector(wc, ins))
 664                         node = &parent->rb_left;
 665                 else
 666                         node = &parent->rb_right;
 667         }
 668         rb_link_node(&ins->rb_node, parent, node);
 669         rb_insert_color(&ins->rb_node, &wc->tree);
 670         list_add(&ins->lru, &wc->lru);
 671         ins->age = jiffies;
 672 }
 673
 674 static void writecache_unlink(struct dm_writecache *wc, struct wc_entry *e)
 675 {
 676         list_del(&e->lru);
 677         rb_erase(&e->rb_node, &wc->tree);
 678 }
 679
 680 static void writecache_add_to_freelist(struct dm_writecache *wc, struct wc_entry *e)
 681 {
 682         if (WC_MODE_SORT_FREELIST(wc)) {
 683                 struct rb_node **node = &wc->freetree.rb_node, *parent = NULL;
 684                 if (unlikely(!*node))
 685                         wc->current_free = e;
 686                 while (*node) {
 687                         parent = *node;
 688                         if (&e->rb_node < *node)
 689                                 node = &parent->rb_left;
 690                         else
 691                                 node = &parent->rb_right;
 692                 }
 693                 rb_link_node(&e->rb_node, parent, node);
 694                 rb_insert_color(&e->rb_node, &wc->freetree);
 695         } else {
 696                 list_add_tail(&e->lru, &wc->freelist);
 697         }
 698         wc->freelist_size++;
 699 }
 700
 701 static inline void writecache_verify_watermark(struct dm_writecache *wc)
 702 {
 703         if (unlikely(wc->freelist_size + wc->writeback_size <= wc->freelist_high_watermark))
 704                 queue_work(wc->writeback_wq, &wc->writeback_work);
 705 }
 706
 707 static void writecache_max_age_timer(struct timer_list *t)
 708 {
 709         struct dm_writecache *wc = from_timer(wc, t, max_age_timer);
 710
 711         if (!dm_suspended(wc->ti) && !writecache_has_error(wc)) {
 712                 queue_work(wc->writeback_wq, &wc->writeback_work);
 713                 mod_timer(&wc->max_age_timer, jiffies + wc->max_age / MAX_AGE_DIV);
 714         }
 715 }
 716
 717 static struct wc_entry *writecache_pop_from_freelist(struct dm_writecache *wc, sector_t expected_sector)
 718 {
 719         struct wc_entry *e;
 720
 721         if (WC_MODE_SORT_FREELIST(wc)) {
 722                 struct rb_node *next;
 723                 if (unlikely(!wc->current_free))
 724                         return NULL;
 725                 e = wc->current_free;
 726                 if (expected_sector != (sector_t)-1 && unlikely(cache_sector(wc, e) != expected_sector))
 727                         return NULL;
 728                 next = rb_next(&e->rb_node);
 729                 rb_erase(&e->rb_node, &wc->freetree);
 730                 if (unlikely(!next))
 731                         next = rb_first(&wc->freetree);
 732                 wc->current_free = next ? container_of(next, struct wc_entry, rb_node) : NULL;
 733         } else {
 734                 if (unlikely(list_empty(&wc->freelist)))
 735                         return NULL;
 736                 e = container_of(wc->freelist.next, struct wc_entry, lru);
 737                 if (expected_sector != (sector_t)-1 && unlikely(cache_sector(wc, e) != expected_sector))
 738                         return NULL;
 739                 list_del(&e->lru);
 740         }
 741         wc->freelist_size--;
 742
 743         writecache_verify_watermark(wc);
 744
 745         return e;
 746 }
 747
 748 static void writecache_free_entry(struct dm_writecache *wc, struct wc_entry *e)
 749 {
 750         writecache_unlink(wc, e);
 751         writecache_add_to_freelist(wc, e);
 752         clear_seq_count(wc, e);
 753         writecache_flush_region(wc, memory_entry(wc, e), sizeof(struct wc_memory_entry));
 754         if (unlikely(waitqueue_active(&wc->freelist_wait)))
 755                 wake_up(&wc->freelist_wait);
 756 }
 757
 758 static void writecache_wait_on_freelist(struct dm_writecache *wc)
 759 {
 760         DEFINE_WAIT(wait);
 761
 762         prepare_to_wait(&wc->freelist_wait, &wait, TASK_UNINTERRUPTIBLE);
 763         wc_unlock(wc);
 764         io_schedule();
 765         finish_wait(&wc->freelist_wait, &wait);
 766         wc_lock(wc);
 767 }
 768
 769 static void writecache_poison_lists(struct dm_writecache *wc)
 770 {
 771         /*
 772          * Catch incorrect access to these values while the device is suspended.
 773          */
 774         memset(&wc->tree, -1, sizeof wc->tree);
 775         wc->lru.next = LIST_POISON1;
 776         wc->lru.prev = LIST_POISON2;
 777         wc->freelist.next = LIST_POISON1;
 778         wc->freelist.prev = LIST_POISON2;
 779 }
 780
 781 static void writecache_flush_entry(struct dm_writecache *wc, struct wc_entry *e)
 782 {
 783         writecache_flush_region(wc, memory_entry(wc, e), sizeof(struct wc_memory_entry));
 784         if (WC_MODE_PMEM(wc))
 785                 writecache_flush_region(wc, memory_data(wc, e), wc->block_size);
 786 }
 787
 788 static bool writecache_entry_is_committed(struct dm_writecache *wc, struct wc_entry *e)
 789 {
 790         return read_seq_count(wc, e) < wc->seq_count;
 791 }
 792
 793 static void writecache_flush(struct dm_writecache *wc)
 794 {
 795         struct wc_entry *e, *e2;
 796         bool need_flush_after_free;
 797
 798         wc->uncommitted_blocks = 0;
 799         del_timer(&wc->autocommit_timer);
 800
 801         if (list_empty(&wc->lru))
 802                 return;
 803
 804         e = container_of(wc->lru.next, struct wc_entry, lru);
 805         if (writecache_entry_is_committed(wc, e)) {
 806                 if (wc->overwrote_committed) {
 807                         writecache_wait_for_ios(wc, WRITE);
 808                         writecache_disk_flush(wc, wc->ssd_dev);
 809                         wc->overwrote_committed = false;
 810                 }
 811                 return;
 812         }
 813         while (1) {
 814                 writecache_flush_entry(wc, e);
 815                 if (unlikely(e->lru.next == &wc->lru))
 816                         break;
 817                 e2 = container_of(e->lru.next, struct wc_entry, lru);
 818                 if (writecache_entry_is_committed(wc, e2))
 819                         break;
 820                 e = e2;
 821                 cond_resched();
 822         }
 823         writecache_commit_flushed(wc, true);
 824
 825         wc->seq_count++;
 826         pmem_assign(sb(wc)->seq_count, cpu_to_le64(wc->seq_count));
 827         if (WC_MODE_PMEM(wc))
 828                 writecache_commit_flushed(wc, false);
 829         else
 830                 ssd_commit_superblock(wc);
 831
 832         wc->overwrote_committed = false;
 833
 834         need_flush_after_free = false;
 835         while (1) {
 836                 /* Free another committed entry with lower seq-count */
 837                 struct rb_node *rb_node = rb_prev(&e->rb_node);
 838
 839                 if (rb_node) {
 840                         e2 = container_of(rb_node, struct wc_entry, rb_node);
 841                         if (read_original_sector(wc, e2) == read_original_sector(wc, e) &&
 842                             likely(!e2->write_in_progress)) {
 843                                 writecache_free_entry(wc, e2);
 844                                 need_flush_after_free = true;
 845                         }
 846                 }
 847                 if (unlikely(e->lru.prev == &wc->lru))
 848                         break;
 849                 e = container_of(e->lru.prev, struct wc_entry, lru);
 850                 cond_resched();
 851         }
 852
 853         if (need_flush_after_free)
 854                 writecache_commit_flushed(wc, false);
 855 }
 856
 857 static void writecache_flush_work(struct work_struct *work)
 858 {
 859         struct dm_writecache *wc = container_of(work, struct dm_writecache, flush_work);
 860
 861         wc_lock(wc);
 862         writecache_flush(wc);
 863         wc_unlock(wc);
 864 }
 865
 866 static void writecache_autocommit_timer(struct timer_list *t)
 867 {
 868         struct dm_writecache *wc = from_timer(wc, t, autocommit_timer);
 869         if (!writecache_has_error(wc))
 870                 queue_work(wc->writeback_wq, &wc->flush_work);
 871 }
 872
 873 static void writecache_schedule_autocommit(struct dm_writecache *wc)
 874 {
 875         if (!timer_pending(&wc->autocommit_timer))
 876                 mod_timer(&wc->autocommit_timer, jiffies + wc->autocommit_jiffies);
 877 }
 878
 879 static void writecache_discard(struct dm_writecache *wc, sector_t start, sector_t end)
 880 {
 881         struct wc_entry *e;
 882         bool discarded_something = false;
 883
 884         e = writecache_find_entry(wc, start, WFE_RETURN_FOLLOWING | WFE_LOWEST_SEQ);
 885         if (unlikely(!e))
 886                 return;
 887
 888         while (read_original_sector(wc, e) < end) {
 889                 struct rb_node *node = rb_next(&e->rb_node);
 890
 891                 if (likely(!e->write_in_progress)) {
 892                         if (!discarded_something) {
 893                                 if (!WC_MODE_PMEM(wc)) {
 894                                         writecache_wait_for_ios(wc, READ);
 895                                         writecache_wait_for_ios(wc, WRITE);
 896                                 }
 897                                 discarded_something = true;
 898                         }
 899                         if (!writecache_entry_is_committed(wc, e))
 900                                 wc->uncommitted_blocks--;
 901                         writecache_free_entry(wc, e);
 902                 }
 903
 904                 if (unlikely(!node))
 905                         break;
 906
 907                 e = container_of(node, struct wc_entry, rb_node);
 908         }
 909
 910         if (discarded_something)
 911                 writecache_commit_flushed(wc, false);
 912 }
 913
 914 static bool writecache_wait_for_writeback(struct dm_writecache *wc)
 915 {
 916         if (wc->writeback_size) {
 917                 writecache_wait_on_freelist(wc);
 918                 return true;
 919         }
 920         return false;
 921 }
 922
 923 static void writecache_suspend(struct dm_target *ti)
 924 {
 925         struct dm_writecache *wc = ti->private;
 926         bool flush_on_suspend;
 927
 928         del_timer_sync(&wc->autocommit_timer);
 929         del_timer_sync(&wc->max_age_timer);
 930
 931         wc_lock(wc);
 932         writecache_flush(wc);
 933         flush_on_suspend = wc->flush_on_suspend;
 934         if (flush_on_suspend) {
 935                 wc->flush_on_suspend = false;
 936                 wc->writeback_all++;
 937                 queue_work(wc->writeback_wq, &wc->writeback_work);
 938         }
 939         wc_unlock(wc);
 940
 941         drain_workqueue(wc->writeback_wq);
 942
 943         wc_lock(wc);
 944         if (flush_on_suspend)
 945                 wc->writeback_all--;
 946         while (writecache_wait_for_writeback(wc));
 947
 948         if (WC_MODE_PMEM(wc))
 949                 persistent_memory_flush_cache(wc->memory_map, wc->memory_map_size);
 950
 951         writecache_poison_lists(wc);
 952
 953         wc_unlock(wc);
 954 }
 955
 956 static int writecache_alloc_entries(struct dm_writecache *wc)
 957 {
 958         size_t b;
 959
 960         if (wc->entries)
 961                 return 0;
 962         wc->entries = vmalloc(array_size(sizeof(struct wc_entry), wc->n_blocks));
 963         if (!wc->entries)
 964                 return -ENOMEM;
 965         for (b = 0; b < wc->n_blocks; b++) {
 966                 struct wc_entry *e = &wc->entries[b];
 967                 e->index = b;
 968                 e->write_in_progress = false;
 969                 cond_resched();
 970         }
 971
 972         return 0;
 973 }
 974
 975 static int writecache_read_metadata(struct dm_writecache *wc, sector_t n_sectors)
 976 {
 977         struct dm_io_region region;
 978         struct dm_io_request req;
 979
 980         region.bdev = wc->ssd_dev->bdev;
 981         region.sector = wc->start_sector;
 982         region.count = n_sectors;
 983         req.bi_op = REQ_OP_READ;
 984         req.bi_op_flags = REQ_SYNC;
 985         req.mem.type = DM_IO_VMA;
 986         req.mem.ptr.vma = (char *)wc->memory_map;
 987         req.client = wc->dm_io;
 988         req.notify.fn = NULL;
 989
 990         return dm_io(&req, 1, &region, NULL);
 991 }
 992
 993 static void writecache_resume(struct dm_target *ti)
 994 {
 995         struct dm_writecache *wc = ti->private;
 996         size_t b;
 997         bool need_flush = false;
 998         __le64 sb_seq_count;
 999         int r;
1000
1001         wc_lock(wc);
1002
1003         wc->data_device_sectors = bdev_nr_sectors(wc->dev->bdev);
1004
1005         if (WC_MODE_PMEM(wc)) {
1006                 persistent_memory_invalidate_cache(wc->memory_map, wc->memory_map_size);
1007         } else {
1008                 r = writecache_read_metadata(wc, wc->metadata_sectors);
1009                 if (r) {
1010                         size_t sb_entries_offset;
1011                         writecache_error(wc, r, "unable to read metadata: %d", r);
1012                         sb_entries_offset = offsetof(struct wc_memory_superblock, entries);
1013                         memset((char *)wc->memory_map + sb_entries_offset, -1,
1014                                (wc->metadata_sectors << SECTOR_SHIFT) - sb_entries_offset);
1015                 }
1016         }
1017
1018         wc->tree = RB_ROOT;
1019         INIT_LIST_HEAD(&wc->lru);
1020         if (WC_MODE_SORT_FREELIST(wc)) {
1021                 wc->freetree = RB_ROOT;
1022                 wc->current_free = NULL;
1023         } else {
1024                 INIT_LIST_HEAD(&wc->freelist);
1025         }
1026         wc->freelist_size = 0;
1027
1028         r = copy_mc_to_kernel(&sb_seq_count, &sb(wc)->seq_count,
1029                               sizeof(uint64_t));
1030         if (r) {
1031                 writecache_error(wc, r, "hardware memory error when reading superblock: %d", r);
1032                 sb_seq_count = cpu_to_le64(0);
1033         }
1034         wc->seq_count = le64_to_cpu(sb_seq_count);
1035
1036 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
1037         for (b = 0; b < wc->n_blocks; b++) {
1038                 struct wc_entry *e = &wc->entries[b];
1039                 struct wc_memory_entry wme;
1040                 if (writecache_has_error(wc)) {
1041                         e->original_sector = -1;
1042                         e->seq_count = -1;
1043                         continue;
1044                 }
1045                 r = copy_mc_to_kernel(&wme, memory_entry(wc, e),
1046                                       sizeof(struct wc_memory_entry));
1047                 if (r) {
1048                         writecache_error(wc, r, "hardware memory error when reading metadata entry %lu: %d",
1049                                          (unsigned long)b, r);
1050                         e->original_sector = -1;
1051                         e->seq_count = -1;
1052                 } else {
1053                         e->original_sector = le64_to_cpu(wme.original_sector);
1054                         e->seq_count = le64_to_cpu(wme.seq_count);
1055                 }
1056                 cond_resched();
1057         }
1058 #endif
1059         for (b = 0; b < wc->n_blocks; b++) {
1060                 struct wc_entry *e = &wc->entries[b];
1061                 if (!writecache_entry_is_committed(wc, e)) {
1062                         if (read_seq_count(wc, e) != -1) {
1063 erase_this:
1064                                 clear_seq_count(wc, e);
1065                                 need_flush = true;
1066                         }
1067                         writecache_add_to_freelist(wc, e);
1068                 } else {
1069                         struct wc_entry *old;
1070
1071                         old = writecache_find_entry(wc, read_original_sector(wc, e), 0);
1072                         if (!old) {
1073                                 writecache_insert_entry(wc, e);
1074                         } else {
1075                                 if (read_seq_count(wc, old) == read_seq_count(wc, e)) {
1076                                         writecache_error(wc, -EINVAL,
1077                                                  "two identical entries, position %llu, sector %llu, sequence %llu",
1078                                                  (unsigned long long)b, (unsigned long long)read_original_sector(wc, e),
1079                                                  (unsigned long long)read_seq_count(wc, e));
1080                                 }
1081                                 if (read_seq_count(wc, old) > read_seq_count(wc, e)) {
1082                                         goto erase_this;
1083                                 } else {
1084                                         writecache_free_entry(wc, old);
1085                                         writecache_insert_entry(wc, e);
1086                                         need_flush = true;
1087                                 }
1088                         }
1089                 }
1090                 cond_resched();
1091         }
1092
1093         if (need_flush) {
1094                 writecache_flush_all_metadata(wc);
1095                 writecache_commit_flushed(wc, false);
1096         }
1097
1098         writecache_verify_watermark(wc);
1099
1100         if (wc->max_age != MAX_AGE_UNSPECIFIED)
1101                 mod_timer(&wc->max_age_timer, jiffies + wc->max_age / MAX_AGE_DIV);
1102
1103         wc_unlock(wc);
1104 }
1105
1106 static int process_flush_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1107 {
1108         if (argc != 1)
1109                 return -EINVAL;
1110
1111         wc_lock(wc);
1112         if (dm_suspended(wc->ti)) {
1113                 wc_unlock(wc);
1114                 return -EBUSY;
1115         }
1116         if (writecache_has_error(wc)) {
1117                 wc_unlock(wc);
1118                 return -EIO;
1119         }
1120
1121         writecache_flush(wc);
1122         wc->writeback_all++;
1123         queue_work(wc->writeback_wq, &wc->writeback_work);
1124         wc_unlock(wc);
1125
1126         flush_workqueue(wc->writeback_wq);
1127
1128         wc_lock(wc);
1129         wc->writeback_all--;
1130         if (writecache_has_error(wc)) {
1131                 wc_unlock(wc);
1132                 return -EIO;
1133         }
1134         wc_unlock(wc);
1135
1136         return 0;
1137 }
1138
1139 static int process_flush_on_suspend_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1140 {
1141         if (argc != 1)
1142                 return -EINVAL;
1143
1144         wc_lock(wc);
1145         wc->flush_on_suspend = true;
1146         wc_unlock(wc);
1147
1148         return 0;
1149 }
1150
1151 static void activate_cleaner(struct dm_writecache *wc)
1152 {
1153         wc->flush_on_suspend = true;
1154         wc->cleaner = true;
1155         wc->freelist_high_watermark = wc->n_blocks;
1156         wc->freelist_low_watermark = wc->n_blocks;
1157 }
1158
1159 static int process_cleaner_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1160 {
1161         if (argc != 1)
1162                 return -EINVAL;
1163
1164         wc_lock(wc);
1165         activate_cleaner(wc);
1166         if (!dm_suspended(wc->ti))
1167                 writecache_verify_watermark(wc);
1168         wc_unlock(wc);
1169
1170         return 0;
1171 }
1172
1173 static int process_clear_stats_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1174 {
1175         if (argc != 1)
1176                 return -EINVAL;
1177
1178         wc_lock(wc);
1179         memset(&wc->stats, 0, sizeof wc->stats);
1180         wc_unlock(wc);
1181
1182         return 0;
1183 }
1184
1185 static int writecache_message(struct dm_target *ti, unsigned argc, char **argv,
1186                               char *result, unsigned maxlen)
1187 {
1188         int r = -EINVAL;
1189         struct dm_writecache *wc = ti->private;
1190
1191         if (!strcasecmp(argv[0], "flush"))
1192                 r = process_flush_mesg(argc, argv, wc);
1193         else if (!strcasecmp(argv[0], "flush_on_suspend"))
1194                 r = process_flush_on_suspend_mesg(argc, argv, wc);
1195         else if (!strcasecmp(argv[0], "cleaner"))
1196                 r = process_cleaner_mesg(argc, argv, wc);
1197         else if (!strcasecmp(argv[0], "clear_stats"))
1198                 r = process_clear_stats_mesg(argc, argv, wc);
1199         else
1200                 DMERR("unrecognised message received: %s", argv[0]);
1201
1202         return r;
1203 }
1204
1205 static void memcpy_flushcache_optimized(void *dest, void *source, size_t size)
1206 {
1207         /*
1208          * clflushopt performs better with block size 1024, 2048, 4096
1209          * non-temporal stores perform better with block size 512
1210          *
1211          * block size   512             1024            2048            4096
1212          * movnti       496 MB/s        642 MB/s        725 MB/s        744 MB/s
1213          * clflushopt   373 MB/s        688 MB/s        1.1 GB/s        1.2 GB/s
1214          *
1215          * We see that movnti performs better for 512-byte blocks, and
1216          * clflushopt performs better for 1024-byte and larger blocks. So, we
1217          * prefer clflushopt for sizes >= 768.
1218          *
1219          * NOTE: this happens to be the case now (with dm-writecache's single
1220          * threaded model) but re-evaluate this once memcpy_flushcache() is
1221          * enabled to use movdir64b which might invalidate this performance
1222          * advantage seen with cache-allocating-writes plus flushing.
1223          */
1224 #ifdef CONFIG_X86
1225         if (static_cpu_has(X86_FEATURE_CLFLUSHOPT) &&
1226             likely(boot_cpu_data.x86_clflush_size == 64) &&
1227             likely(size >= 768)) {
1228                 do {
1229                         memcpy((void *)dest, (void *)source, 64);
1230                         clflushopt((void *)dest);
1231                         dest += 64;
1232                         source += 64;
1233                         size -= 64;
1234                 } while (size >= 64);
1235                 return;
1236         }
1237 #endif
1238         memcpy_flushcache(dest, source, size);
1239 }
1240
1241 static void bio_copy_block(struct dm_writecache *wc, struct bio *bio, void *data)
1242 {
1243         void *buf;
1244         unsigned size;
1245         int rw = bio_data_dir(bio);
1246         unsigned remaining_size = wc->block_size;
1247
1248         do {
1249                 struct bio_vec bv = bio_iter_iovec(bio, bio->bi_iter);
1250                 buf = bvec_kmap_local(&bv);
1251                 size = bv.bv_len;
1252                 if (unlikely(size > remaining_size))
1253                         size = remaining_size;
1254
1255                 if (rw == READ) {
1256                         int r;
1257                         r = copy_mc_to_kernel(buf, data, size);
1258                         flush_dcache_page(bio_page(bio));
1259                         if (unlikely(r)) {
1260                                 writecache_error(wc, r, "hardware memory error when reading data: %d", r);
1261                                 bio->bi_status = BLK_STS_IOERR;
1262                         }
1263                 } else {
1264                         flush_dcache_page(bio_page(bio));
1265                         memcpy_flushcache_optimized(data, buf, size);
1266                 }
1267
1268                 kunmap_local(buf);
1269
1270                 data = (char *)data + size;
1271                 remaining_size -= size;
1272                 bio_advance(bio, size);
1273         } while (unlikely(remaining_size));
1274 }
1275
1276 static int writecache_flush_thread(void *data)
1277 {
1278         struct dm_writecache *wc = data;
1279
1280         while (1) {
1281                 struct bio *bio;
1282
1283                 wc_lock(wc);
1284                 bio = bio_list_pop(&wc->flush_list);
1285                 if (!bio) {
1286                         set_current_state(TASK_INTERRUPTIBLE);
1287                         wc_unlock(wc);
1288
1289                         if (unlikely(kthread_should_stop())) {
1290                                 set_current_state(TASK_RUNNING);
1291                                 break;
1292                         }
1293
1294                         schedule();
1295                         continue;
1296                 }
1297
1298                 if (bio_op(bio) == REQ_OP_DISCARD) {
1299                         writecache_discard(wc, bio->bi_iter.bi_sector,
1300                                            bio_end_sector(bio));
1301                         wc_unlock(wc);
1302                         bio_set_dev(bio, wc->dev->bdev);
1303                         submit_bio_noacct(bio);
1304                 } else {
1305                         writecache_flush(wc);
1306                         wc_unlock(wc);
1307                         if (writecache_has_error(wc))
1308                                 bio->bi_status = BLK_STS_IOERR;
1309                         bio_endio(bio);
1310                 }
1311         }
1312
1313         return 0;
1314 }
1315
1316 static void writecache_offload_bio(struct dm_writecache *wc, struct bio *bio)
1317 {
1318         if (bio_list_empty(&wc->flush_list))
1319                 wake_up_process(wc->flush_thread);
1320         bio_list_add(&wc->flush_list, bio);
1321 }
1322
1323 enum wc_map_op {
1324         WC_MAP_SUBMIT,
1325         WC_MAP_REMAP,
1326         WC_MAP_REMAP_ORIGIN,
1327         WC_MAP_RETURN,
1328         WC_MAP_ERROR,
1329 };
1330
1331 static void writecache_map_remap_origin(struct dm_writecache *wc, struct bio *bio,
1332                                         struct wc_entry *e)
1333 {
1334         if (e) {
1335                 sector_t next_boundary =
1336                         read_original_sector(wc, e) - bio->bi_iter.bi_sector;
1337                 if (next_boundary < bio->bi_iter.bi_size >> SECTOR_SHIFT)
1338                         dm_accept_partial_bio(bio, next_boundary);
1339         }
1340 }
1341
1342 static enum wc_map_op writecache_map_read(struct dm_writecache *wc, struct bio *bio)
1343 {
1344         enum wc_map_op map_op;
1345         struct wc_entry *e;
1346
1347 read_next_block:
1348         wc->stats.reads++;
1349         e = writecache_find_entry(wc, bio->bi_iter.bi_sector, WFE_RETURN_FOLLOWING);
1350         if (e && read_original_sector(wc, e) == bio->bi_iter.bi_sector) {
1351                 wc->stats.read_hits++;
1352                 if (WC_MODE_PMEM(wc)) {
1353                         bio_copy_block(wc, bio, memory_data(wc, e));
1354                         if (bio->bi_iter.bi_size)
1355                                 goto read_next_block;
1356                         map_op = WC_MAP_SUBMIT;
1357                 } else {
1358                         dm_accept_partial_bio(bio, wc->block_size >> SECTOR_SHIFT);
1359                         bio_set_dev(bio, wc->ssd_dev->bdev);
1360                         bio->bi_iter.bi_sector = cache_sector(wc, e);
1361                         if (!writecache_entry_is_committed(wc, e))
1362                                 writecache_wait_for_ios(wc, WRITE);
1363                         map_op = WC_MAP_REMAP;
1364                 }
1365         } else {
1366                 writecache_map_remap_origin(wc, bio, e);
1367                 wc->stats.reads += (bio->bi_iter.bi_size - wc->block_size) >> wc->block_size_bits;
1368                 map_op = WC_MAP_REMAP_ORIGIN;
1369         }
1370
1371         return map_op;
1372 }
1373
1374 static void writecache_bio_copy_ssd(struct dm_writecache *wc, struct bio *bio,
1375                                     struct wc_entry *e, bool search_used)
1376 {
1377         unsigned bio_size = wc->block_size;
1378         sector_t start_cache_sec = cache_sector(wc, e);
1379         sector_t current_cache_sec = start_cache_sec + (bio_size >> SECTOR_SHIFT);
1380
1381         while (bio_size < bio->bi_iter.bi_size) {
1382                 if (!search_used) {
1383                         struct wc_entry *f = writecache_pop_from_freelist(wc, current_cache_sec);
1384                         if (!f)
1385                                 break;
1386                         write_original_sector_seq_count(wc, f, bio->bi_iter.bi_sector +
1387                                                         (bio_size >> SECTOR_SHIFT), wc->seq_count);
1388                         writecache_insert_entry(wc, f);
1389                         wc->uncommitted_blocks++;
1390                 } else {
1391                         struct wc_entry *f;
1392                         struct rb_node *next = rb_next(&e->rb_node);
1393                         if (!next)
1394                                 break;
1395                         f = container_of(next, struct wc_entry, rb_node);
1396                         if (f != e + 1)
1397                                 break;
1398                         if (read_original_sector(wc, f) !=
1399                             read_original_sector(wc, e) + (wc->block_size >> SECTOR_SHIFT))
1400                                 break;
1401                         if (unlikely(f->write_in_progress))
1402                                 break;
1403                         if (writecache_entry_is_committed(wc, f))
1404                                 wc->overwrote_committed = true;
1405                         e = f;
1406                 }
1407                 bio_size += wc->block_size;
1408                 current_cache_sec += wc->block_size >> SECTOR_SHIFT;
1409         }
1410
1411         bio_set_dev(bio, wc->ssd_dev->bdev);
1412         bio->bi_iter.bi_sector = start_cache_sec;
1413         dm_accept_partial_bio(bio, bio_size >> SECTOR_SHIFT);
1414
1415         wc->stats.writes += bio->bi_iter.bi_size >> wc->block_size_bits;
1416         wc->stats.writes_allocate += (bio->bi_iter.bi_size - wc->block_size) >> wc->block_size_bits;
1417
1418         if (unlikely(wc->uncommitted_blocks >= wc->autocommit_blocks)) {
1419                 wc->uncommitted_blocks = 0;
1420                 queue_work(wc->writeback_wq, &wc->flush_work);
1421         } else {
1422                 writecache_schedule_autocommit(wc);
1423         }
1424 }
1425
1426 static enum wc_map_op writecache_map_write(struct dm_writecache *wc, struct bio *bio)
1427 {
1428         struct wc_entry *e;
1429
1430         do {
1431                 bool found_entry = false;
1432                 bool search_used = false;
1433                 if (writecache_has_error(wc)) {
1434                         wc->stats.writes += bio->bi_iter.bi_size >> wc->block_size_bits;
1435                         return WC_MAP_ERROR;
1436                 }
1437                 e = writecache_find_entry(wc, bio->bi_iter.bi_sector, 0);
1438                 if (e) {
1439                         if (!writecache_entry_is_committed(wc, e)) {
1440                                 wc->stats.write_hits_uncommitted++;
1441                                 search_used = true;
1442                                 goto bio_copy;
1443                         }
1444                         wc->stats.write_hits_committed++;
1445                         if (!WC_MODE_PMEM(wc) && !e->write_in_progress) {
1446                                 wc->overwrote_committed = true;
1447                                 search_used = true;
1448                                 goto bio_copy;
1449                         }
1450                         found_entry = true;
1451                 } else {
1452                         if (unlikely(wc->cleaner) ||
1453                             (wc->metadata_only && !(bio->bi_opf & REQ_META)))
1454                                 goto direct_write;
1455                 }
1456                 e = writecache_pop_from_freelist(wc, (sector_t)-1);
1457                 if (unlikely(!e)) {
1458                         if (!WC_MODE_PMEM(wc) && !found_entry) {
1459 direct_write:
1460                                 e = writecache_find_entry(wc, bio->bi_iter.bi_sector, WFE_RETURN_FOLLOWING);
1461                                 writecache_map_remap_origin(wc, bio, e);
1462                                 wc->stats.writes_around += bio->bi_iter.bi_size >> wc->block_size_bits;
1463                                 wc->stats.writes += bio->bi_iter.bi_size >> wc->block_size_bits;
1464                                 return WC_MAP_REMAP_ORIGIN;
1465                         }
1466                         wc->stats.writes_blocked_on_freelist++;
1467                         writecache_wait_on_freelist(wc);
1468                         continue;
1469                 }
1470                 write_original_sector_seq_count(wc, e, bio->bi_iter.bi_sector, wc->seq_count);
1471                 writecache_insert_entry(wc, e);
1472                 wc->uncommitted_blocks++;
1473                 wc->stats.writes_allocate++;
1474 bio_copy:
1475                 if (WC_MODE_PMEM(wc)) {
1476                         bio_copy_block(wc, bio, memory_data(wc, e));
1477                         wc->stats.writes++;
1478                 } else {
1479                         writecache_bio_copy_ssd(wc, bio, e, search_used);
1480                         return WC_MAP_REMAP;
1481                 }
1482         } while (bio->bi_iter.bi_size);
1483
1484         if (unlikely(bio->bi_opf & REQ_FUA || wc->uncommitted_blocks >= wc->autocommit_blocks))
1485                 writecache_flush(wc);
1486         else
1487                 writecache_schedule_autocommit(wc);
1488
1489         return WC_MAP_SUBMIT;
1490 }
1491
1492 static enum wc_map_op writecache_map_flush(struct dm_writecache *wc, struct bio *bio)
1493 {
1494         if (writecache_has_error(wc))
1495                 return WC_MAP_ERROR;
1496
1497         if (WC_MODE_PMEM(wc)) {
1498                 wc->stats.flushes++;
1499                 writecache_flush(wc);
1500                 if (writecache_has_error(wc))
1501                         return WC_MAP_ERROR;
1502                 else if (unlikely(wc->cleaner) || unlikely(wc->metadata_only))
1503                         return WC_MAP_REMAP_ORIGIN;
1504                 return WC_MAP_SUBMIT;
1505         }
1506         /* SSD: */
1507         if (dm_bio_get_target_bio_nr(bio))
1508                 return WC_MAP_REMAP_ORIGIN;
1509         wc->stats.flushes++;
1510         writecache_offload_bio(wc, bio);
1511         return WC_MAP_RETURN;
1512 }
1513
1514 static enum wc_map_op writecache_map_discard(struct dm_writecache *wc, struct bio *bio)
1515 {
1516         wc->stats.discards += bio->bi_iter.bi_size >> wc->block_size_bits;
1517
1518         if (writecache_has_error(wc))
1519                 return WC_MAP_ERROR;
1520
1521         if (WC_MODE_PMEM(wc)) {
1522                 writecache_discard(wc, bio->bi_iter.bi_sector, bio_end_sector(bio));
1523                 return WC_MAP_REMAP_ORIGIN;
1524         }
1525         /* SSD: */
1526         writecache_offload_bio(wc, bio);
1527         return WC_MAP_RETURN;
1528 }
1529
1530 static int writecache_map(struct dm_target *ti, struct bio *bio)
1531 {
1532         struct dm_writecache *wc = ti->private;
1533         enum wc_map_op map_op;
1534
1535         bio->bi_private = NULL;
1536
1537         wc_lock(wc);
1538
1539         if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1540                 map_op = writecache_map_flush(wc, bio);
1541                 goto done;
1542         }
1543
1544         bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
1545
1546         if (unlikely((((unsigned)bio->bi_iter.bi_sector | bio_sectors(bio)) &
1547                                 (wc->block_size / 512 - 1)) != 0)) {
1548                 DMERR("I/O is not aligned, sector %llu, size %u, block size %u",
1549                       (unsigned long long)bio->bi_iter.bi_sector,
1550                       bio->bi_iter.bi_size, wc->block_size);
1551                 map_op = WC_MAP_ERROR;
1552                 goto done;
1553         }
1554
1555         if (unlikely(bio_op(bio) == REQ_OP_DISCARD)) {
1556                 map_op = writecache_map_discard(wc, bio);
1557                 goto done;
1558         }
1559
1560         if (bio_data_dir(bio) == READ)
1561                 map_op = writecache_map_read(wc, bio);
1562         else
1563                 map_op = writecache_map_write(wc, bio);
1564 done:
1565         switch (map_op) {
1566         case WC_MAP_REMAP_ORIGIN:
1567                 if (likely(wc->pause != 0)) {
1568                         if (bio_op(bio) == REQ_OP_WRITE) {
1569                                 dm_iot_io_begin(&wc->iot, 1);
1570                                 bio->bi_private = (void *)2;
1571                         }
1572                 }
1573                 bio_set_dev(bio, wc->dev->bdev);
1574                 wc_unlock(wc);
1575                 return DM_MAPIO_REMAPPED;
1576
1577         case WC_MAP_REMAP:
1578                 /* make sure that writecache_end_io decrements bio_in_progress: */
1579                 bio->bi_private = (void *)1;
1580                 atomic_inc(&wc->bio_in_progress[bio_data_dir(bio)]);
1581                 wc_unlock(wc);
1582                 return DM_MAPIO_REMAPPED;
1583
1584         case WC_MAP_SUBMIT:
1585                 wc_unlock(wc);
1586                 bio_endio(bio);
1587                 return DM_MAPIO_SUBMITTED;
1588
1589         case WC_MAP_RETURN:
1590                 wc_unlock(wc);
1591                 return DM_MAPIO_SUBMITTED;
1592
1593         case WC_MAP_ERROR:
1594                 wc_unlock(wc);
1595                 bio_io_error(bio);
1596                 return DM_MAPIO_SUBMITTED;
1597
1598         default:
1599                 BUG();
1600                 return -1;
1601         }
1602 }
1603
1604 static int writecache_end_io(struct dm_target *ti, struct bio *bio, blk_status_t *status)
1605 {
1606         struct dm_writecache *wc = ti->private;
1607
1608         if (bio->bi_private == (void *)1) {
1609                 int dir = bio_data_dir(bio);
1610                 if (atomic_dec_and_test(&wc->bio_in_progress[dir]))
1611                         if (unlikely(waitqueue_active(&wc->bio_in_progress_wait[dir])))
1612                                 wake_up(&wc->bio_in_progress_wait[dir]);
1613         } else if (bio->bi_private == (void *)2) {
1614                 dm_iot_io_end(&wc->iot, 1);
1615         }
1616         return 0;
1617 }
1618
1619 static int writecache_iterate_devices(struct dm_target *ti,
1620                                       iterate_devices_callout_fn fn, void *data)
1621 {
1622         struct dm_writecache *wc = ti->private;
1623
1624         return fn(ti, wc->dev, 0, ti->len, data);
1625 }
1626
1627 static void writecache_io_hints(struct dm_target *ti, struct queue_limits *limits)
1628 {
1629         struct dm_writecache *wc = ti->private;
1630
1631         if (limits->logical_block_size < wc->block_size)
1632                 limits->logical_block_size = wc->block_size;
1633
1634         if (limits->physical_block_size < wc->block_size)
1635                 limits->physical_block_size = wc->block_size;
1636
1637         if (limits->io_min < wc->block_size)
1638                 limits->io_min = wc->block_size;
1639 }
1640
1641
1642 static void writecache_writeback_endio(struct bio *bio)
1643 {
1644         struct writeback_struct *wb = container_of(bio, struct writeback_struct, bio);
1645         struct dm_writecache *wc = wb->wc;
1646         unsigned long flags;
1647
1648         raw_spin_lock_irqsave(&wc->endio_list_lock, flags);
1649         if (unlikely(list_empty(&wc->endio_list)))
1650                 wake_up_process(wc->endio_thread);
1651         list_add_tail(&wb->endio_entry, &wc->endio_list);
1652         raw_spin_unlock_irqrestore(&wc->endio_list_lock, flags);
1653 }
1654
1655 static void writecache_copy_endio(int read_err, unsigned long write_err, void *ptr)
1656 {
1657         struct copy_struct *c = ptr;
1658         struct dm_writecache *wc = c->wc;
1659
1660         c->error = likely(!(read_err | write_err)) ? 0 : -EIO;
1661
1662         raw_spin_lock_irq(&wc->endio_list_lock);
1663         if (unlikely(list_empty(&wc->endio_list)))
1664                 wake_up_process(wc->endio_thread);
1665         list_add_tail(&c->endio_entry, &wc->endio_list);
1666         raw_spin_unlock_irq(&wc->endio_list_lock);
1667 }
1668
1669 static void __writecache_endio_pmem(struct dm_writecache *wc, struct list_head *list)
1670 {
1671         unsigned i;
1672         struct writeback_struct *wb;
1673         struct wc_entry *e;
1674         unsigned long n_walked = 0;
1675
1676         do {
1677                 wb = list_entry(list->next, struct writeback_struct, endio_entry);
1678                 list_del(&wb->endio_entry);
1679
1680                 if (unlikely(wb->bio.bi_status != BLK_STS_OK))
1681                         writecache_error(wc, blk_status_to_errno(wb->bio.bi_status),
1682                                         "write error %d", wb->bio.bi_status);
1683                 i = 0;
1684                 do {
1685                         e = wb->wc_list[i];
1686                         BUG_ON(!e->write_in_progress);
1687                         e->write_in_progress = false;
1688                         INIT_LIST_HEAD(&e->lru);
1689                         if (!writecache_has_error(wc))
1690                                 writecache_free_entry(wc, e);
1691                         BUG_ON(!wc->writeback_size);
1692                         wc->writeback_size--;
1693                         n_walked++;
1694                         if (unlikely(n_walked >= ENDIO_LATENCY)) {
1695                                 writecache_commit_flushed(wc, false);
1696                                 wc_unlock(wc);
1697                                 wc_lock(wc);
1698                                 n_walked = 0;
1699                         }
1700                 } while (++i < wb->wc_list_n);
1701
1702                 if (wb->wc_list != wb->wc_list_inline)
1703                         kfree(wb->wc_list);
1704                 bio_put(&wb->bio);
1705         } while (!list_empty(list));
1706 }
1707
1708 static void __writecache_endio_ssd(struct dm_writecache *wc, struct list_head *list)
1709 {
1710         struct copy_struct *c;
1711         struct wc_entry *e;
1712
1713         do {
1714                 c = list_entry(list->next, struct copy_struct, endio_entry);
1715                 list_del(&c->endio_entry);
1716
1717                 if (unlikely(c->error))
1718                         writecache_error(wc, c->error, "copy error");
1719
1720                 e = c->e;
1721                 do {
1722                         BUG_ON(!e->write_in_progress);
1723                         e->write_in_progress = false;
1724                         INIT_LIST_HEAD(&e->lru);
1725                         if (!writecache_has_error(wc))
1726                                 writecache_free_entry(wc, e);
1727
1728                         BUG_ON(!wc->writeback_size);
1729                         wc->writeback_size--;
1730                         e++;
1731                 } while (--c->n_entries);
1732                 mempool_free(c, &wc->copy_pool);
1733         } while (!list_empty(list));
1734 }
1735
1736 static int writecache_endio_thread(void *data)
1737 {
1738         struct dm_writecache *wc = data;
1739
1740         while (1) {
1741                 struct list_head list;
1742
1743                 raw_spin_lock_irq(&wc->endio_list_lock);
1744                 if (!list_empty(&wc->endio_list))
1745                         goto pop_from_list;
1746                 set_current_state(TASK_INTERRUPTIBLE);
1747                 raw_spin_unlock_irq(&wc->endio_list_lock);
1748
1749                 if (unlikely(kthread_should_stop())) {
1750                         set_current_state(TASK_RUNNING);
1751                         break;
1752                 }
1753
1754                 schedule();
1755
1756                 continue;
1757
1758 pop_from_list:
1759                 list = wc->endio_list;
1760                 list.next->prev = list.prev->next = &list;
1761                 INIT_LIST_HEAD(&wc->endio_list);
1762                 raw_spin_unlock_irq(&wc->endio_list_lock);
1763
1764                 if (!WC_MODE_FUA(wc))
1765                         writecache_disk_flush(wc, wc->dev);
1766
1767                 wc_lock(wc);
1768
1769                 if (WC_MODE_PMEM(wc)) {
1770                         __writecache_endio_pmem(wc, &list);
1771                 } else {
1772                         __writecache_endio_ssd(wc, &list);
1773                         writecache_wait_for_ios(wc, READ);
1774                 }
1775
1776                 writecache_commit_flushed(wc, false);
1777
1778                 wc_unlock(wc);
1779         }
1780
1781         return 0;
1782 }
1783
1784 static bool wc_add_block(struct writeback_struct *wb, struct wc_entry *e)
1785 {
1786         struct dm_writecache *wc = wb->wc;
1787         unsigned block_size = wc->block_size;
1788         void *address = memory_data(wc, e);
1789
1790         persistent_memory_flush_cache(address, block_size);
1791
1792         if (unlikely(bio_end_sector(&wb->bio) >= wc->data_device_sectors))
1793                 return true;
1794
1795         return bio_add_page(&wb->bio, persistent_memory_page(address),
1796                             block_size, persistent_memory_page_offset(address)) != 0;
1797 }
1798
1799 struct writeback_list {
1800         struct list_head list;
1801         size_t size;
1802 };
1803
1804 static void __writeback_throttle(struct dm_writecache *wc, struct writeback_list *wbl)
1805 {
1806         if (unlikely(wc->max_writeback_jobs)) {
1807                 if (READ_ONCE(wc->writeback_size) - wbl->size >= wc->max_writeback_jobs) {
1808                         wc_lock(wc);
1809                         while (wc->writeback_size - wbl->size >= wc->max_writeback_jobs)
1810                                 writecache_wait_on_freelist(wc);
1811                         wc_unlock(wc);
1812                 }
1813         }
1814         cond_resched();
1815 }
1816
1817 static void __writecache_writeback_pmem(struct dm_writecache *wc, struct writeback_list *wbl)
1818 {
1819         struct wc_entry *e, *f;
1820         struct bio *bio;
1821         struct writeback_struct *wb;
1822         unsigned max_pages;
1823
1824         while (wbl->size) {
1825                 wbl->size--;
1826                 e = container_of(wbl->list.prev, struct wc_entry, lru);
1827                 list_del(&e->lru);
1828
1829                 max_pages = e->wc_list_contiguous;
1830
1831                 bio = bio_alloc_bioset(GFP_NOIO, max_pages, &wc->bio_set);
1832                 wb = container_of(bio, struct writeback_struct, bio);
1833                 wb->wc = wc;
1834                 bio->bi_end_io = writecache_writeback_endio;
1835                 bio_set_dev(bio, wc->dev->bdev);
1836                 bio->bi_iter.bi_sector = read_original_sector(wc, e);
1837                 if (max_pages <= WB_LIST_INLINE ||
1838                     unlikely(!(wb->wc_list = kmalloc_array(max_pages, sizeof(struct wc_entry *),
1839                                                            GFP_NOIO | __GFP_NORETRY |
1840                                                            __GFP_NOMEMALLOC | __GFP_NOWARN)))) {
1841                         wb->wc_list = wb->wc_list_inline;
1842                         max_pages = WB_LIST_INLINE;
1843                 }
1844
1845                 BUG_ON(!wc_add_block(wb, e));
1846
1847                 wb->wc_list[0] = e;
1848                 wb->wc_list_n = 1;
1849
1850                 while (wbl->size && wb->wc_list_n < max_pages) {
1851                         f = container_of(wbl->list.prev, struct wc_entry, lru);
1852                         if (read_original_sector(wc, f) !=
1853                             read_original_sector(wc, e) + (wc->block_size >> SECTOR_SHIFT))
1854                                 break;
1855                         if (!wc_add_block(wb, f))
1856                                 break;
1857                         wbl->size--;
1858                         list_del(&f->lru);
1859                         wb->wc_list[wb->wc_list_n++] = f;
1860                         e = f;
1861                 }
1862                 bio_set_op_attrs(bio, REQ_OP_WRITE, WC_MODE_FUA(wc) * REQ_FUA);
1863                 if (writecache_has_error(wc)) {
1864                         bio->bi_status = BLK_STS_IOERR;
1865                         bio_endio(bio);
1866                 } else if (unlikely(!bio_sectors(bio))) {
1867                         bio->bi_status = BLK_STS_OK;
1868                         bio_endio(bio);
1869                 } else {
1870                         submit_bio(bio);
1871                 }
1872
1873                 __writeback_throttle(wc, wbl);
1874         }
1875 }
1876
1877 static void __writecache_writeback_ssd(struct dm_writecache *wc, struct writeback_list *wbl)
1878 {
1879         struct wc_entry *e, *f;
1880         struct dm_io_region from, to;
1881         struct copy_struct *c;
1882
1883         while (wbl->size) {
1884                 unsigned n_sectors;
1885
1886                 wbl->size--;
1887                 e = container_of(wbl->list.prev, struct wc_entry, lru);
1888                 list_del(&e->lru);
1889
1890                 n_sectors = e->wc_list_contiguous << (wc->block_size_bits - SECTOR_SHIFT);
1891
1892                 from.bdev = wc->ssd_dev->bdev;
1893                 from.sector = cache_sector(wc, e);
1894                 from.count = n_sectors;
1895                 to.bdev = wc->dev->bdev;
1896                 to.sector = read_original_sector(wc, e);
1897                 to.count = n_sectors;
1898
1899                 c = mempool_alloc(&wc->copy_pool, GFP_NOIO);
1900                 c->wc = wc;
1901                 c->e = e;
1902                 c->n_entries = e->wc_list_contiguous;
1903
1904                 while ((n_sectors -= wc->block_size >> SECTOR_SHIFT)) {
1905                         wbl->size--;
1906                         f = container_of(wbl->list.prev, struct wc_entry, lru);
1907                         BUG_ON(f != e + 1);
1908                         list_del(&f->lru);
1909                         e = f;
1910                 }
1911
1912                 if (unlikely(to.sector + to.count > wc->data_device_sectors)) {
1913                         if (to.sector >= wc->data_device_sectors) {
1914                                 writecache_copy_endio(0, 0, c);
1915                                 continue;
1916                         }
1917                         from.count = to.count = wc->data_device_sectors - to.sector;
1918                 }
1919
1920                 dm_kcopyd_copy(wc->dm_kcopyd, &from, 1, &to, 0, writecache_copy_endio, c);
1921
1922                 __writeback_throttle(wc, wbl);
1923         }
1924 }
1925
1926 static void writecache_writeback(struct work_struct *work)
1927 {
1928         struct dm_writecache *wc = container_of(work, struct dm_writecache, writeback_work);
1929         struct blk_plug plug;
1930         struct wc_entry *f, *g, *e = NULL;
1931         struct rb_node *node, *next_node;
1932         struct list_head skipped;
1933         struct writeback_list wbl;
1934         unsigned long n_walked;
1935
1936         if (!WC_MODE_PMEM(wc)) {
1937                 /* Wait for any active kcopyd work on behalf of ssd writeback */
1938                 dm_kcopyd_client_flush(wc->dm_kcopyd);
1939         }
1940
1941         if (likely(wc->pause != 0)) {
1942                 while (1) {
1943                         unsigned long idle;
1944                         if (unlikely(wc->cleaner) || unlikely(wc->writeback_all) ||
1945                             unlikely(dm_suspended(wc->ti)))
1946                                 break;
1947                         idle = dm_iot_idle_time(&wc->iot);
1948                         if (idle >= wc->pause)
1949                                 break;
1950                         idle = wc->pause - idle;
1951                         if (idle > HZ)
1952                                 idle = HZ;
1953                         schedule_timeout_idle(idle);
1954                 }
1955         }
1956
1957         wc_lock(wc);
1958 restart:
1959         if (writecache_has_error(wc)) {
1960                 wc_unlock(wc);
1961                 return;
1962         }
1963
1964         if (unlikely(wc->writeback_all)) {
1965                 if (writecache_wait_for_writeback(wc))
1966                         goto restart;
1967         }
1968
1969         if (wc->overwrote_committed) {
1970                 writecache_wait_for_ios(wc, WRITE);
1971         }
1972
1973         n_walked = 0;
1974         INIT_LIST_HEAD(&skipped);
1975         INIT_LIST_HEAD(&wbl.list);
1976         wbl.size = 0;
1977         while (!list_empty(&wc->lru) &&
1978                (wc->writeback_all ||
1979                 wc->freelist_size + wc->writeback_size <= wc->freelist_low_watermark ||
1980                 (jiffies - container_of(wc->lru.prev, struct wc_entry, lru)->age >=
1981                  wc->max_age - wc->max_age / MAX_AGE_DIV))) {
1982
1983                 n_walked++;
1984                 if (unlikely(n_walked > WRITEBACK_LATENCY) &&
1985                     likely(!wc->writeback_all)) {
1986                         if (likely(!dm_suspended(wc->ti)))
1987                                 queue_work(wc->writeback_wq, &wc->writeback_work);
1988                         break;
1989                 }
1990
1991                 if (unlikely(wc->writeback_all)) {
1992                         if (unlikely(!e)) {
1993                                 writecache_flush(wc);
1994                                 e = container_of(rb_first(&wc->tree), struct wc_entry, rb_node);
1995                         } else
1996                                 e = g;
1997                 } else
1998                         e = container_of(wc->lru.prev, struct wc_entry, lru);
1999                 BUG_ON(e->write_in_progress);
2000                 if (unlikely(!writecache_entry_is_committed(wc, e))) {
2001                         writecache_flush(wc);
2002                 }
2003                 node = rb_prev(&e->rb_node);
2004                 if (node) {
2005                         f = container_of(node, struct wc_entry, rb_node);
2006                         if (unlikely(read_original_sector(wc, f) ==
2007                                      read_original_sector(wc, e))) {
2008                                 BUG_ON(!f->write_in_progress);
2009                                 list_move(&e->lru, &skipped);
2010                                 cond_resched();
2011                                 continue;
2012                         }
2013                 }
2014                 wc->writeback_size++;
2015                 list_move(&e->lru, &wbl.list);
2016                 wbl.size++;
2017                 e->write_in_progress = true;
2018                 e->wc_list_contiguous = 1;
2019
2020                 f = e;
2021
2022                 while (1) {
2023                         next_node = rb_next(&f->rb_node);
2024                         if (unlikely(!next_node))
2025                                 break;
2026                         g = container_of(next_node, struct wc_entry, rb_node);
2027                         if (unlikely(read_original_sector(wc, g) ==
2028                             read_original_sector(wc, f))) {
2029                                 f = g;
2030                                 continue;
2031                         }
2032                         if (read_original_sector(wc, g) !=
2033                             read_original_sector(wc, f) + (wc->block_size >> SECTOR_SHIFT))
2034                                 break;
2035                         if (unlikely(g->write_in_progress))
2036                                 break;
2037                         if (unlikely(!writecache_entry_is_committed(wc, g)))
2038                                 break;
2039
2040                         if (!WC_MODE_PMEM(wc)) {
2041                                 if (g != f + 1)
2042                                         break;
2043                         }
2044
2045                         n_walked++;
2046                         //if (unlikely(n_walked > WRITEBACK_LATENCY) && likely(!wc->writeback_all))
2047                         //      break;
2048
2049                         wc->writeback_size++;
2050                         list_move(&g->lru, &wbl.list);
2051                         wbl.size++;
2052                         g->write_in_progress = true;
2053                         g->wc_list_contiguous = BIO_MAX_VECS;
2054                         f = g;
2055                         e->wc_list_contiguous++;
2056                         if (unlikely(e->wc_list_contiguous == BIO_MAX_VECS)) {
2057                                 if (unlikely(wc->writeback_all)) {
2058                                         next_node = rb_next(&f->rb_node);
2059                                         if (likely(next_node))
2060                                                 g = container_of(next_node, struct wc_entry, rb_node);
2061                                 }
2062                                 break;
2063                         }
2064                 }
2065                 cond_resched();
2066         }
2067
2068         if (!list_empty(&skipped)) {
2069                 list_splice_tail(&skipped, &wc->lru);
2070                 /*
2071                  * If we didn't do any progress, we must wait until some
2072                  * writeback finishes to avoid burning CPU in a loop
2073                  */
2074                 if (unlikely(!wbl.size))
2075                         writecache_wait_for_writeback(wc);
2076         }
2077
2078         wc_unlock(wc);
2079
2080         blk_start_plug(&plug);
2081
2082         if (WC_MODE_PMEM(wc))
2083                 __writecache_writeback_pmem(wc, &wbl);
2084         else
2085                 __writecache_writeback_ssd(wc, &wbl);
2086
2087         blk_finish_plug(&plug);
2088
2089         if (unlikely(wc->writeback_all)) {
2090                 wc_lock(wc);
2091                 while (writecache_wait_for_writeback(wc));
2092                 wc_unlock(wc);
2093         }
2094 }
2095
2096 static int calculate_memory_size(uint64_t device_size, unsigned block_size,
2097                                  size_t *n_blocks_p, size_t *n_metadata_blocks_p)
2098 {
2099         uint64_t n_blocks, offset;
2100         struct wc_entry e;
2101
2102         n_blocks = device_size;
2103         do_div(n_blocks, block_size + sizeof(struct wc_memory_entry));
2104
2105         while (1) {
2106                 if (!n_blocks)
2107                         return -ENOSPC;
2108                 /* Verify the following entries[n_blocks] won't overflow */
2109                 if (n_blocks >= ((size_t)-sizeof(struct wc_memory_superblock) /
2110                                  sizeof(struct wc_memory_entry)))
2111                         return -EFBIG;
2112                 offset = offsetof(struct wc_memory_superblock, entries[n_blocks]);
2113                 offset = (offset + block_size - 1) & ~(uint64_t)(block_size - 1);
2114                 if (offset + n_blocks * block_size <= device_size)
2115                         break;
2116                 n_blocks--;
2117         }
2118
2119         /* check if the bit field overflows */
2120         e.index = n_blocks;
2121         if (e.index != n_blocks)
2122                 return -EFBIG;
2123
2124         if (n_blocks_p)
2125                 *n_blocks_p = n_blocks;
2126         if (n_metadata_blocks_p)
2127                 *n_metadata_blocks_p = offset >> __ffs(block_size);
2128         return 0;
2129 }
2130
2131 static int init_memory(struct dm_writecache *wc)
2132 {
2133         size_t b;
2134         int r;
2135
2136         r = calculate_memory_size(wc->memory_map_size, wc->block_size, &wc->n_blocks, NULL);
2137         if (r)
2138                 return r;
2139
2140         r = writecache_alloc_entries(wc);
2141         if (r)
2142                 return r;
2143
2144         for (b = 0; b < ARRAY_SIZE(sb(wc)->padding); b++)
2145                 pmem_assign(sb(wc)->padding[b], cpu_to_le64(0));
2146         pmem_assign(sb(wc)->version, cpu_to_le32(MEMORY_SUPERBLOCK_VERSION));
2147         pmem_assign(sb(wc)->block_size, cpu_to_le32(wc->block_size));
2148         pmem_assign(sb(wc)->n_blocks, cpu_to_le64(wc->n_blocks));
2149         pmem_assign(sb(wc)->seq_count, cpu_to_le64(0));
2150
2151         for (b = 0; b < wc->n_blocks; b++) {
2152                 write_original_sector_seq_count(wc, &wc->entries[b], -1, -1);
2153                 cond_resched();
2154         }
2155
2156         writecache_flush_all_metadata(wc);
2157         writecache_commit_flushed(wc, false);
2158         pmem_assign(sb(wc)->magic, cpu_to_le32(MEMORY_SUPERBLOCK_MAGIC));
2159         writecache_flush_region(wc, &sb(wc)->magic, sizeof sb(wc)->magic);
2160         writecache_commit_flushed(wc, false);
2161
2162         return 0;
2163 }
2164
2165 static void writecache_dtr(struct dm_target *ti)
2166 {
2167         struct dm_writecache *wc = ti->private;
2168
2169         if (!wc)
2170                 return;
2171
2172         if (wc->endio_thread)
2173                 kthread_stop(wc->endio_thread);
2174
2175         if (wc->flush_thread)
2176                 kthread_stop(wc->flush_thread);
2177
2178         bioset_exit(&wc->bio_set);
2179
2180         mempool_exit(&wc->copy_pool);
2181
2182         if (wc->writeback_wq)
2183                 destroy_workqueue(wc->writeback_wq);
2184
2185         if (wc->dev)
2186                 dm_put_device(ti, wc->dev);
2187
2188         if (wc->ssd_dev)
2189                 dm_put_device(ti, wc->ssd_dev);
2190
2191         vfree(wc->entries);
2192
2193         if (wc->memory_map) {
2194                 if (WC_MODE_PMEM(wc))
2195                         persistent_memory_release(wc);
2196                 else
2197                         vfree(wc->memory_map);
2198         }
2199
2200         if (wc->dm_kcopyd)
2201                 dm_kcopyd_client_destroy(wc->dm_kcopyd);
2202
2203         if (wc->dm_io)
2204                 dm_io_client_destroy(wc->dm_io);
2205
2206         vfree(wc->dirty_bitmap);
2207
2208         kfree(wc);
2209 }
2210
2211 static int writecache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2212 {
2213         struct dm_writecache *wc;
2214         struct dm_arg_set as;
2215         const char *string;
2216         unsigned opt_params;
2217         size_t offset, data_size;
2218         int i, r;
2219         char dummy;
2220         int high_wm_percent = HIGH_WATERMARK;
2221         int low_wm_percent = LOW_WATERMARK;
2222         uint64_t x;
2223         struct wc_memory_superblock s;
2224
2225         static struct dm_arg _args[] = {
2226                 {0, 18, "Invalid number of feature args"},
2227         };
2228
2229         as.argc = argc;
2230         as.argv = argv;
2231
2232         wc = kzalloc(sizeof(struct dm_writecache), GFP_KERNEL);
2233         if (!wc) {
2234                 ti->error = "Cannot allocate writecache structure";
2235                 r = -ENOMEM;
2236                 goto bad;
2237         }
2238         ti->private = wc;
2239         wc->ti = ti;
2240
2241         mutex_init(&wc->lock);
2242         wc->max_age = MAX_AGE_UNSPECIFIED;
2243         writecache_poison_lists(wc);
2244         init_waitqueue_head(&wc->freelist_wait);
2245         timer_setup(&wc->autocommit_timer, writecache_autocommit_timer, 0);
2246         timer_setup(&wc->max_age_timer, writecache_max_age_timer, 0);
2247
2248         for (i = 0; i < 2; i++) {
2249                 atomic_set(&wc->bio_in_progress[i], 0);
2250                 init_waitqueue_head(&wc->bio_in_progress_wait[i]);
2251         }
2252
2253         wc->dm_io = dm_io_client_create();
2254         if (IS_ERR(wc->dm_io)) {
2255                 r = PTR_ERR(wc->dm_io);
2256                 ti->error = "Unable to allocate dm-io client";
2257                 wc->dm_io = NULL;
2258                 goto bad;
2259         }
2260
2261         wc->writeback_wq = alloc_workqueue("writecache-writeback", WQ_MEM_RECLAIM, 1);
2262         if (!wc->writeback_wq) {
2263                 r = -ENOMEM;
2264                 ti->error = "Could not allocate writeback workqueue";
2265                 goto bad;
2266         }
2267         INIT_WORK(&wc->writeback_work, writecache_writeback);
2268         INIT_WORK(&wc->flush_work, writecache_flush_work);
2269
2270         dm_iot_init(&wc->iot);
2271
2272         raw_spin_lock_init(&wc->endio_list_lock);
2273         INIT_LIST_HEAD(&wc->endio_list);
2274         wc->endio_thread = kthread_create(writecache_endio_thread, wc, "writecache_endio");
2275         if (IS_ERR(wc->endio_thread)) {
2276                 r = PTR_ERR(wc->endio_thread);
2277                 wc->endio_thread = NULL;
2278                 ti->error = "Couldn't spawn endio thread";
2279                 goto bad;
2280         }
2281         wake_up_process(wc->endio_thread);
2282
2283         /*
2284          * Parse the mode (pmem or ssd)
2285          */
2286         string = dm_shift_arg(&as);
2287         if (!string)
2288                 goto bad_arguments;
2289
2290         if (!strcasecmp(string, "s")) {
2291                 wc->pmem_mode = false;
2292         } else if (!strcasecmp(string, "p")) {
2293 #ifdef DM_WRITECACHE_HAS_PMEM
2294                 wc->pmem_mode = true;
2295                 wc->writeback_fua = true;
2296 #else
2297                 /*
2298                  * If the architecture doesn't support persistent memory or
2299                  * the kernel doesn't support any DAX drivers, this driver can
2300                  * only be used in SSD-only mode.
2301                  */
2302                 r = -EOPNOTSUPP;
2303                 ti->error = "Persistent memory or DAX not supported on this system";
2304                 goto bad;
2305 #endif
2306         } else {
2307                 goto bad_arguments;
2308         }
2309
2310         if (WC_MODE_PMEM(wc)) {
2311                 r = bioset_init(&wc->bio_set, BIO_POOL_SIZE,
2312                                 offsetof(struct writeback_struct, bio),
2313                                 BIOSET_NEED_BVECS);
2314                 if (r) {
2315                         ti->error = "Could not allocate bio set";
2316                         goto bad;
2317                 }
2318         } else {
2319                 wc->pause = PAUSE_WRITEBACK;
2320                 r = mempool_init_kmalloc_pool(&wc->copy_pool, 1, sizeof(struct copy_struct));
2321                 if (r) {
2322                         ti->error = "Could not allocate mempool";
2323                         goto bad;
2324                 }
2325         }
2326
2327         /*
2328          * Parse the origin data device
2329          */
2330         string = dm_shift_arg(&as);
2331         if (!string)
2332                 goto bad_arguments;
2333         r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->dev);
2334         if (r) {
2335                 ti->error = "Origin data device lookup failed";
2336                 goto bad;
2337         }
2338
2339         /*
2340          * Parse cache data device (be it pmem or ssd)
2341          */
2342         string = dm_shift_arg(&as);
2343         if (!string)
2344                 goto bad_arguments;
2345
2346         r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->ssd_dev);
2347         if (r) {
2348                 ti->error = "Cache data device lookup failed";
2349                 goto bad;
2350         }
2351         wc->memory_map_size = i_size_read(wc->ssd_dev->bdev->bd_inode);
2352
2353         /*
2354          * Parse the cache block size
2355          */
2356         string = dm_shift_arg(&as);
2357         if (!string)
2358                 goto bad_arguments;
2359         if (sscanf(string, "%u%c", &wc->block_size, &dummy) != 1 ||
2360             wc->block_size < 512 || wc->block_size > PAGE_SIZE ||
2361             (wc->block_size & (wc->block_size - 1))) {
2362                 r = -EINVAL;
2363                 ti->error = "Invalid block size";
2364                 goto bad;
2365         }
2366         if (wc->block_size < bdev_logical_block_size(wc->dev->bdev) ||
2367             wc->block_size < bdev_logical_block_size(wc->ssd_dev->bdev)) {
2368                 r = -EINVAL;
2369                 ti->error = "Block size is smaller than device logical block size";
2370                 goto bad;
2371         }
2372         wc->block_size_bits = __ffs(wc->block_size);
2373
2374         wc->max_writeback_jobs = MAX_WRITEBACK_JOBS;
2375         wc->autocommit_blocks = !WC_MODE_PMEM(wc) ? AUTOCOMMIT_BLOCKS_SSD : AUTOCOMMIT_BLOCKS_PMEM;
2376         wc->autocommit_jiffies = msecs_to_jiffies(AUTOCOMMIT_MSEC);
2377
2378         /*
2379          * Parse optional arguments
2380          */
2381         r = dm_read_arg_group(_args, &as, &opt_params, &ti->error);
2382         if (r)
2383                 goto bad;
2384
2385         while (opt_params) {
2386                 string = dm_shift_arg(&as), opt_params--;
2387                 if (!strcasecmp(string, "start_sector") && opt_params >= 1) {
2388                         unsigned long long start_sector;
2389                         string = dm_shift_arg(&as), opt_params--;
2390                         if (sscanf(string, "%llu%c", &start_sector, &dummy) != 1)
2391                                 goto invalid_optional;
2392                         wc->start_sector = start_sector;
2393                         wc->start_sector_set = true;
2394                         if (wc->start_sector != start_sector ||
2395                             wc->start_sector >= wc->memory_map_size >> SECTOR_SHIFT)
2396                                 goto invalid_optional;
2397                 } else if (!strcasecmp(string, "high_watermark") && opt_params >= 1) {
2398                         string = dm_shift_arg(&as), opt_params--;
2399                         if (sscanf(string, "%d%c", &high_wm_percent, &dummy) != 1)
2400                                 goto invalid_optional;
2401                         if (high_wm_percent < 0 || high_wm_percent > 100)
2402                                 goto invalid_optional;
2403                         wc->high_wm_percent_value = high_wm_percent;
2404                         wc->high_wm_percent_set = true;
2405                 } else if (!strcasecmp(string, "low_watermark") && opt_params >= 1) {
2406                         string = dm_shift_arg(&as), opt_params--;
2407                         if (sscanf(string, "%d%c", &low_wm_percent, &dummy) != 1)
2408                                 goto invalid_optional;
2409                         if (low_wm_percent < 0 || low_wm_percent > 100)
2410                                 goto invalid_optional;
2411                         wc->low_wm_percent_value = low_wm_percent;
2412                         wc->low_wm_percent_set = true;
2413                 } else if (!strcasecmp(string, "writeback_jobs") && opt_params >= 1) {
2414                         string = dm_shift_arg(&as), opt_params--;
2415                         if (sscanf(string, "%u%c", &wc->max_writeback_jobs, &dummy) != 1)
2416                                 goto invalid_optional;
2417                         wc->max_writeback_jobs_set = true;
2418                 } else if (!strcasecmp(string, "autocommit_blocks") && opt_params >= 1) {
2419                         string = dm_shift_arg(&as), opt_params--;
2420                         if (sscanf(string, "%u%c", &wc->autocommit_blocks, &dummy) != 1)
2421                                 goto invalid_optional;
2422                         wc->autocommit_blocks_set = true;
2423                 } else if (!strcasecmp(string, "autocommit_time") && opt_params >= 1) {
2424                         unsigned autocommit_msecs;
2425                         string = dm_shift_arg(&as), opt_params--;
2426                         if (sscanf(string, "%u%c", &autocommit_msecs, &dummy) != 1)
2427                                 goto invalid_optional;
2428                         if (autocommit_msecs > 3600000)
2429                                 goto invalid_optional;
2430                         wc->autocommit_jiffies = msecs_to_jiffies(autocommit_msecs);
2431                         wc->autocommit_time_value = autocommit_msecs;
2432                         wc->autocommit_time_set = true;
2433                 } else if (!strcasecmp(string, "max_age") && opt_params >= 1) {
2434                         unsigned max_age_msecs;
2435                         string = dm_shift_arg(&as), opt_params--;
2436                         if (sscanf(string, "%u%c", &max_age_msecs, &dummy) != 1)
2437                                 goto invalid_optional;
2438                         if (max_age_msecs > 86400000)
2439                                 goto invalid_optional;
2440                         wc->max_age = msecs_to_jiffies(max_age_msecs);
2441                         wc->max_age_set = true;
2442                         wc->max_age_value = max_age_msecs;
2443                 } else if (!strcasecmp(string, "cleaner")) {
2444                         wc->cleaner_set = true;
2445                         wc->cleaner = true;
2446                 } else if (!strcasecmp(string, "fua")) {
2447                         if (WC_MODE_PMEM(wc)) {
2448                                 wc->writeback_fua = true;
2449                                 wc->writeback_fua_set = true;
2450                         } else goto invalid_optional;
2451                 } else if (!strcasecmp(string, "nofua")) {
2452                         if (WC_MODE_PMEM(wc)) {
2453                                 wc->writeback_fua = false;
2454                                 wc->writeback_fua_set = true;
2455                         } else goto invalid_optional;
2456                 } else if (!strcasecmp(string, "metadata_only")) {
2457                         wc->metadata_only = true;
2458                 } else if (!strcasecmp(string, "pause_writeback") && opt_params >= 1) {
2459                         unsigned pause_msecs;
2460                         if (WC_MODE_PMEM(wc))
2461                                 goto invalid_optional;
2462                         string = dm_shift_arg(&as), opt_params--;
2463                         if (sscanf(string, "%u%c", &pause_msecs, &dummy) != 1)
2464                                 goto invalid_optional;
2465                         if (pause_msecs > 60000)
2466                                 goto invalid_optional;
2467                         wc->pause = msecs_to_jiffies(pause_msecs);
2468                         wc->pause_set = true;
2469                         wc->pause_value = pause_msecs;
2470                 } else {
2471 invalid_optional:
2472                         r = -EINVAL;
2473                         ti->error = "Invalid optional argument";
2474                         goto bad;
2475                 }
2476         }
2477
2478         if (high_wm_percent < low_wm_percent) {
2479                 r = -EINVAL;
2480                 ti->error = "High watermark must be greater than or equal to low watermark";
2481                 goto bad;
2482         }
2483
2484         if (WC_MODE_PMEM(wc)) {
2485                 if (!dax_synchronous(wc->ssd_dev->dax_dev)) {
2486                         r = -EOPNOTSUPP;
2487                         ti->error = "Asynchronous persistent memory not supported as pmem cache";
2488                         goto bad;
2489                 }
2490
2491                 r = persistent_memory_claim(wc);
2492                 if (r) {
2493                         ti->error = "Unable to map persistent memory for cache";
2494                         goto bad;
2495                 }
2496         } else {
2497                 size_t n_blocks, n_metadata_blocks;
2498                 uint64_t n_bitmap_bits;
2499
2500                 wc->memory_map_size -= (uint64_t)wc->start_sector << SECTOR_SHIFT;
2501
2502                 bio_list_init(&wc->flush_list);
2503                 wc->flush_thread = kthread_create(writecache_flush_thread, wc, "dm_writecache_flush");
2504                 if (IS_ERR(wc->flush_thread)) {
2505                         r = PTR_ERR(wc->flush_thread);
2506                         wc->flush_thread = NULL;
2507                         ti->error = "Couldn't spawn flush thread";
2508                         goto bad;
2509                 }
2510                 wake_up_process(wc->flush_thread);
2511
2512                 r = calculate_memory_size(wc->memory_map_size, wc->block_size,
2513                                           &n_blocks, &n_metadata_blocks);
2514                 if (r) {
2515                         ti->error = "Invalid device size";
2516                         goto bad;
2517                 }
2518
2519                 n_bitmap_bits = (((uint64_t)n_metadata_blocks << wc->block_size_bits) +
2520                                  BITMAP_GRANULARITY - 1) / BITMAP_GRANULARITY;
2521                 /* this is limitation of test_bit functions */
2522                 if (n_bitmap_bits > 1U << 31) {
2523                         r = -EFBIG;
2524                         ti->error = "Invalid device size";
2525                         goto bad;
2526                 }
2527
2528                 wc->memory_map = vmalloc(n_metadata_blocks << wc->block_size_bits);
2529                 if (!wc->memory_map) {
2530                         r = -ENOMEM;
2531                         ti->error = "Unable to allocate memory for metadata";
2532                         goto bad;
2533                 }
2534
2535                 wc->dm_kcopyd = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2536                 if (IS_ERR(wc->dm_kcopyd)) {
2537                         r = PTR_ERR(wc->dm_kcopyd);
2538                         ti->error = "Unable to allocate dm-kcopyd client";
2539                         wc->dm_kcopyd = NULL;
2540                         goto bad;
2541                 }
2542
2543                 wc->metadata_sectors = n_metadata_blocks << (wc->block_size_bits - SECTOR_SHIFT);
2544                 wc->dirty_bitmap_size = (n_bitmap_bits + BITS_PER_LONG - 1) /
2545                         BITS_PER_LONG * sizeof(unsigned long);
2546                 wc->dirty_bitmap = vzalloc(wc->dirty_bitmap_size);
2547                 if (!wc->dirty_bitmap) {
2548                         r = -ENOMEM;
2549                         ti->error = "Unable to allocate dirty bitmap";
2550                         goto bad;
2551                 }
2552
2553                 r = writecache_read_metadata(wc, wc->block_size >> SECTOR_SHIFT);
2554                 if (r) {
2555                         ti->error = "Unable to read first block of metadata";
2556                         goto bad;
2557                 }
2558         }
2559
2560         r = copy_mc_to_kernel(&s, sb(wc), sizeof(struct wc_memory_superblock));
2561         if (r) {
2562                 ti->error = "Hardware memory error when reading superblock";
2563                 goto bad;
2564         }
2565         if (!le32_to_cpu(s.magic) && !le32_to_cpu(s.version)) {
2566                 r = init_memory(wc);
2567                 if (r) {
2568                         ti->error = "Unable to initialize device";
2569                         goto bad;
2570                 }
2571                 r = copy_mc_to_kernel(&s, sb(wc),
2572                                       sizeof(struct wc_memory_superblock));
2573                 if (r) {
2574                         ti->error = "Hardware memory error when reading superblock";
2575                         goto bad;
2576                 }
2577         }
2578
2579         if (le32_to_cpu(s.magic) != MEMORY_SUPERBLOCK_MAGIC) {
2580                 ti->error = "Invalid magic in the superblock";
2581                 r = -EINVAL;
2582                 goto bad;
2583         }
2584
2585         if (le32_to_cpu(s.version) != MEMORY_SUPERBLOCK_VERSION) {
2586                 ti->error = "Invalid version in the superblock";
2587                 r = -EINVAL;
2588                 goto bad;
2589         }
2590
2591         if (le32_to_cpu(s.block_size) != wc->block_size) {
2592                 ti->error = "Block size does not match superblock";
2593                 r = -EINVAL;
2594                 goto bad;
2595         }
2596
2597         wc->n_blocks = le64_to_cpu(s.n_blocks);
2598
2599         offset = wc->n_blocks * sizeof(struct wc_memory_entry);
2600         if (offset / sizeof(struct wc_memory_entry) != le64_to_cpu(sb(wc)->n_blocks)) {
2601 overflow:
2602                 ti->error = "Overflow in size calculation";
2603                 r = -EINVAL;
2604                 goto bad;
2605         }
2606         offset += sizeof(struct wc_memory_superblock);
2607         if (offset < sizeof(struct wc_memory_superblock))
2608                 goto overflow;
2609         offset = (offset + wc->block_size - 1) & ~(size_t)(wc->block_size - 1);
2610         data_size = wc->n_blocks * (size_t)wc->block_size;
2611         if (!offset || (data_size / wc->block_size != wc->n_blocks) ||
2612             (offset + data_size < offset))
2613                 goto overflow;
2614         if (offset + data_size > wc->memory_map_size) {
2615                 ti->error = "Memory area is too small";
2616                 r = -EINVAL;
2617                 goto bad;
2618         }
2619
2620         wc->metadata_sectors = offset >> SECTOR_SHIFT;
2621         wc->block_start = (char *)sb(wc) + offset;
2622
2623         x = (uint64_t)wc->n_blocks * (100 - high_wm_percent);
2624         x += 50;
2625         do_div(x, 100);
2626         wc->freelist_high_watermark = x;
2627         x = (uint64_t)wc->n_blocks * (100 - low_wm_percent);
2628         x += 50;
2629         do_div(x, 100);
2630         wc->freelist_low_watermark = x;
2631
2632         if (wc->cleaner)
2633                 activate_cleaner(wc);
2634
2635         r = writecache_alloc_entries(wc);
2636         if (r) {
2637                 ti->error = "Cannot allocate memory";
2638                 goto bad;
2639         }
2640
2641         ti->num_flush_bios = WC_MODE_PMEM(wc) ? 1 : 2;
2642         ti->flush_supported = true;
2643         ti->num_discard_bios = 1;
2644
2645         if (WC_MODE_PMEM(wc))
2646                 persistent_memory_flush_cache(wc->memory_map, wc->memory_map_size);
2647
2648         return 0;
2649
2650 bad_arguments:
2651         r = -EINVAL;
2652         ti->error = "Bad arguments";
2653 bad:
2654         writecache_dtr(ti);
2655         return r;
2656 }
2657
2658 static void writecache_status(struct dm_target *ti, status_type_t type,
2659                               unsigned status_flags, char *result, unsigned maxlen)
2660 {
2661         struct dm_writecache *wc = ti->private;
2662         unsigned extra_args;
2663         unsigned sz = 0;
2664
2665         switch (type) {
2666         case STATUSTYPE_INFO:
2667                 DMEMIT("%ld %llu %llu %llu %llu %llu %llu %llu %llu %llu %llu %llu %llu %llu",
2668                        writecache_has_error(wc),
2669                        (unsigned long long)wc->n_blocks, (unsigned long long)wc->freelist_size,
2670                        (unsigned long long)wc->writeback_size,
2671                        wc->stats.reads,
2672                        wc->stats.read_hits,
2673                        wc->stats.writes,
2674                        wc->stats.write_hits_uncommitted,
2675                        wc->stats.write_hits_committed,
2676                        wc->stats.writes_around,
2677                        wc->stats.writes_allocate,
2678                        wc->stats.writes_blocked_on_freelist,
2679                        wc->stats.flushes,
2680                        wc->stats.discards);
2681                 break;
2682         case STATUSTYPE_TABLE:
2683                 DMEMIT("%c %s %s %u ", WC_MODE_PMEM(wc) ? 'p' : 's',
2684                                 wc->dev->name, wc->ssd_dev->name, wc->block_size);
2685                 extra_args = 0;
2686                 if (wc->start_sector_set)
2687                         extra_args += 2;
2688                 if (wc->high_wm_percent_set)
2689                         extra_args += 2;
2690                 if (wc->low_wm_percent_set)
2691                         extra_args += 2;
2692                 if (wc->max_writeback_jobs_set)
2693                         extra_args += 2;
2694                 if (wc->autocommit_blocks_set)
2695                         extra_args += 2;
2696                 if (wc->autocommit_time_set)
2697                         extra_args += 2;
2698                 if (wc->max_age_set)
2699                         extra_args += 2;
2700                 if (wc->cleaner_set)
2701                         extra_args++;
2702                 if (wc->writeback_fua_set)
2703                         extra_args++;
2704                 if (wc->metadata_only)
2705                         extra_args++;
2706                 if (wc->pause_set)
2707                         extra_args += 2;
2708
2709                 DMEMIT("%u", extra_args);
2710                 if (wc->start_sector_set)
2711                         DMEMIT(" start_sector %llu", (unsigned long long)wc->start_sector);
2712                 if (wc->high_wm_percent_set)
2713                         DMEMIT(" high_watermark %u", wc->high_wm_percent_value);
2714                 if (wc->low_wm_percent_set)
2715                         DMEMIT(" low_watermark %u", wc->low_wm_percent_value);
2716                 if (wc->max_writeback_jobs_set)
2717                         DMEMIT(" writeback_jobs %u", wc->max_writeback_jobs);
2718                 if (wc->autocommit_blocks_set)
2719                         DMEMIT(" autocommit_blocks %u", wc->autocommit_blocks);
2720                 if (wc->autocommit_time_set)
2721                         DMEMIT(" autocommit_time %u", wc->autocommit_time_value);
2722                 if (wc->max_age_set)
2723                         DMEMIT(" max_age %u", wc->max_age_value);
2724                 if (wc->cleaner_set)
2725                         DMEMIT(" cleaner");
2726                 if (wc->writeback_fua_set)
2727                         DMEMIT(" %sfua", wc->writeback_fua ? "" : "no");
2728                 if (wc->metadata_only)
2729                         DMEMIT(" metadata_only");
2730                 if (wc->pause_set)
2731                         DMEMIT(" pause_writeback %u", wc->pause_value);
2732                 break;
2733         case STATUSTYPE_IMA:
2734                 *result = '\0';
2735                 break;
2736         }
2737 }
2738
2739 static struct target_type writecache_target = {
2740         .name                   = "writecache",
2741         .version                = {1, 6, 0},
2742         .module                 = THIS_MODULE,
2743         .ctr                    = writecache_ctr,
2744         .dtr                    = writecache_dtr,
2745         .status                 = writecache_status,
2746         .postsuspend            = writecache_suspend,
2747         .resume                 = writecache_resume,
2748         .message                = writecache_message,
2749         .map                    = writecache_map,
2750         .end_io                 = writecache_end_io,
2751         .iterate_devices        = writecache_iterate_devices,
2752         .io_hints               = writecache_io_hints,
2753 };
2754
2755 static int __init dm_writecache_init(void)
2756 {
2757         int r;
2758
2759         r = dm_register_target(&writecache_target);
2760         if (r < 0) {
2761                 DMERR("register failed %d", r);
2762                 return r;
2763         }
2764
2765         return 0;
2766 }
2767
2768 static void __exit dm_writecache_exit(void)
2769 {
2770         dm_unregister_target(&writecache_target);
2771 }
2772
2773 module_init(dm_writecache_init);
2774 module_exit(dm_writecache_exit);
2775
2776 MODULE_DESCRIPTION(DM_NAME " writecache target");
2777 MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
2778 MODULE_LICENSE("GPL");