drivers/md/dm-writecache.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2018 Red Hat. All rights reserved.
   4  *
   5  * This file is released under the GPL.
   6  */
   7
   8 #include <linux/device-mapper.h>
   9 #include <linux/module.h>
  10 #include <linux/init.h>
  11 #include <linux/vmalloc.h>
  12 #include <linux/kthread.h>
  13 #include <linux/dm-io.h>
  14 #include <linux/dm-kcopyd.h>
  15 #include <linux/dax.h>
  16 #include <linux/pfn_t.h>
  17 #include <linux/libnvdimm.h>
  18 #include <linux/delay.h>
  19 #include "dm-io-tracker.h"
  20
  21 #define DM_MSG_PREFIX "writecache"
  22
  23 #define HIGH_WATERMARK                  50
  24 #define LOW_WATERMARK                   45
  25 #define MAX_WRITEBACK_JOBS              0
  26 #define ENDIO_LATENCY                   16
  27 #define WRITEBACK_LATENCY               64
  28 #define AUTOCOMMIT_BLOCKS_SSD           65536
  29 #define AUTOCOMMIT_BLOCKS_PMEM          64
  30 #define AUTOCOMMIT_MSEC                 1000
  31 #define MAX_AGE_DIV                     16
  32 #define MAX_AGE_UNSPECIFIED             -1UL
  33 #define PAUSE_WRITEBACK                 (HZ * 3)
  34
  35 #define BITMAP_GRANULARITY      65536
  36 #if BITMAP_GRANULARITY < PAGE_SIZE
  37 #undef BITMAP_GRANULARITY
  38 #define BITMAP_GRANULARITY      PAGE_SIZE
  39 #endif
  40
  41 #if IS_ENABLED(CONFIG_ARCH_HAS_PMEM_API) && IS_ENABLED(CONFIG_FS_DAX)
  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, DAX_ACCESS,
 290                         &wc->memory_map, &pfn);
 291         if (da < 0) {
 292                 wc->memory_map = NULL;
 293                 r = da;
 294                 goto err2;
 295         }
 296         if (!pfn_t_has_page(pfn)) {
 297                 wc->memory_map = NULL;
 298                 r = -EOPNOTSUPP;
 299                 goto err2;
 300         }
 301         if (da != p) {
 302                 long i;
 303                 wc->memory_map = NULL;
 304                 pages = kvmalloc_array(p, sizeof(struct page *), GFP_KERNEL);
 305                 if (!pages) {
 306                         r = -ENOMEM;
 307                         goto err2;
 308                 }
 309                 i = 0;
 310                 do {
 311                         long daa;
 312                         daa = dax_direct_access(wc->ssd_dev->dax_dev, offset + i,
 313                                         p - i, DAX_ACCESS, NULL, &pfn);
 314                         if (daa <= 0) {
 315                                 r = daa ? daa : -EINVAL;
 316                                 goto err3;
 317                         }
 318                         if (!pfn_t_has_page(pfn)) {
 319                                 r = -EOPNOTSUPP;
 320                                 goto err3;
 321                         }
 322                         while (daa-- && i < p) {
 323                                 pages[i++] = pfn_t_to_page(pfn);
 324                                 pfn.val++;
 325                                 if (!(i & 15))
 326                                         cond_resched();
 327                         }
 328                 } while (i < p);
 329                 wc->memory_map = vmap(pages, p, VM_MAP, PAGE_KERNEL);
 330                 if (!wc->memory_map) {
 331                         r = -ENOMEM;
 332                         goto err3;
 333                 }
 334                 kvfree(pages);
 335                 wc->memory_vmapped = true;
 336         }
 337
 338         dax_read_unlock(id);
 339
 340         wc->memory_map += (size_t)wc->start_sector << SECTOR_SHIFT;
 341         wc->memory_map_size -= (size_t)wc->start_sector << SECTOR_SHIFT;
 342
 343         return 0;
 344 err3:
 345         kvfree(pages);
 346 err2:
 347         dax_read_unlock(id);
 348 err1:
 349         return r;
 350 }
 351 #else
 352 static int persistent_memory_claim(struct dm_writecache *wc)
 353 {
 354         return -EOPNOTSUPP;
 355 }
 356 #endif
 357
 358 static void persistent_memory_release(struct dm_writecache *wc)
 359 {
 360         if (wc->memory_vmapped)
 361                 vunmap(wc->memory_map - ((size_t)wc->start_sector << SECTOR_SHIFT));
 362 }
 363
 364 static struct page *persistent_memory_page(void *addr)
 365 {
 366         if (is_vmalloc_addr(addr))
 367                 return vmalloc_to_page(addr);
 368         else
 369                 return virt_to_page(addr);
 370 }
 371
 372 static unsigned persistent_memory_page_offset(void *addr)
 373 {
 374         return (unsigned long)addr & (PAGE_SIZE - 1);
 375 }
 376
 377 static void persistent_memory_flush_cache(void *ptr, size_t size)
 378 {
 379         if (is_vmalloc_addr(ptr))
 380                 flush_kernel_vmap_range(ptr, size);
 381 }
 382
 383 static void persistent_memory_invalidate_cache(void *ptr, size_t size)
 384 {
 385         if (is_vmalloc_addr(ptr))
 386                 invalidate_kernel_vmap_range(ptr, size);
 387 }
 388
 389 static struct wc_memory_superblock *sb(struct dm_writecache *wc)
 390 {
 391         return wc->memory_map;
 392 }
 393
 394 static struct wc_memory_entry *memory_entry(struct dm_writecache *wc, struct wc_entry *e)
 395 {
 396         return &sb(wc)->entries[e->index];
 397 }
 398
 399 static void *memory_data(struct dm_writecache *wc, struct wc_entry *e)
 400 {
 401         return (char *)wc->block_start + (e->index << wc->block_size_bits);
 402 }
 403
 404 static sector_t cache_sector(struct dm_writecache *wc, struct wc_entry *e)
 405 {
 406         return wc->start_sector + wc->metadata_sectors +
 407                 ((sector_t)e->index << (wc->block_size_bits - SECTOR_SHIFT));
 408 }
 409
 410 static uint64_t read_original_sector(struct dm_writecache *wc, struct wc_entry *e)
 411 {
 412 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 413         return e->original_sector;
 414 #else
 415         return le64_to_cpu(memory_entry(wc, e)->original_sector);
 416 #endif
 417 }
 418
 419 static uint64_t read_seq_count(struct dm_writecache *wc, struct wc_entry *e)
 420 {
 421 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 422         return e->seq_count;
 423 #else
 424         return le64_to_cpu(memory_entry(wc, e)->seq_count);
 425 #endif
 426 }
 427
 428 static void clear_seq_count(struct dm_writecache *wc, struct wc_entry *e)
 429 {
 430 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 431         e->seq_count = -1;
 432 #endif
 433         pmem_assign(memory_entry(wc, e)->seq_count, cpu_to_le64(-1));
 434 }
 435
 436 static void write_original_sector_seq_count(struct dm_writecache *wc, struct wc_entry *e,
 437                                             uint64_t original_sector, uint64_t seq_count)
 438 {
 439         struct wc_memory_entry me;
 440 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 441         e->original_sector = original_sector;
 442         e->seq_count = seq_count;
 443 #endif
 444         me.original_sector = cpu_to_le64(original_sector);
 445         me.seq_count = cpu_to_le64(seq_count);
 446         pmem_assign(*memory_entry(wc, e), me);
 447 }
 448
 449 #define writecache_error(wc, err, msg, arg...)                          \
 450 do {                                                                    \
 451         if (!cmpxchg(&(wc)->error, 0, err))                             \
 452                 DMERR(msg, ##arg);                                      \
 453         wake_up(&(wc)->freelist_wait);                                  \
 454 } while (0)
 455
 456 #define writecache_has_error(wc)        (unlikely(READ_ONCE((wc)->error)))
 457
 458 static void writecache_flush_all_metadata(struct dm_writecache *wc)
 459 {
 460         if (!WC_MODE_PMEM(wc))
 461                 memset(wc->dirty_bitmap, -1, wc->dirty_bitmap_size);
 462 }
 463
 464 static void writecache_flush_region(struct dm_writecache *wc, void *ptr, size_t size)
 465 {
 466         if (!WC_MODE_PMEM(wc))
 467                 __set_bit(((char *)ptr - (char *)wc->memory_map) / BITMAP_GRANULARITY,
 468                           wc->dirty_bitmap);
 469 }
 470
 471 static void writecache_disk_flush(struct dm_writecache *wc, struct dm_dev *dev);
 472
 473 struct io_notify {
 474         struct dm_writecache *wc;
 475         struct completion c;
 476         atomic_t count;
 477 };
 478
 479 static void writecache_notify_io(unsigned long error, void *context)
 480 {
 481         struct io_notify *endio = context;
 482
 483         if (unlikely(error != 0))
 484                 writecache_error(endio->wc, -EIO, "error writing metadata");
 485         BUG_ON(atomic_read(&endio->count) <= 0);
 486         if (atomic_dec_and_test(&endio->count))
 487                 complete(&endio->c);
 488 }
 489
 490 static void writecache_wait_for_ios(struct dm_writecache *wc, int direction)
 491 {
 492         wait_event(wc->bio_in_progress_wait[direction],
 493                    !atomic_read(&wc->bio_in_progress[direction]));
 494 }
 495
 496 static void ssd_commit_flushed(struct dm_writecache *wc, bool wait_for_ios)
 497 {
 498         struct dm_io_region region;
 499         struct dm_io_request req;
 500         struct io_notify endio = {
 501                 wc,
 502                 COMPLETION_INITIALIZER_ONSTACK(endio.c),
 503                 ATOMIC_INIT(1),
 504         };
 505         unsigned bitmap_bits = wc->dirty_bitmap_size * 8;
 506         unsigned i = 0;
 507
 508         while (1) {
 509                 unsigned j;
 510                 i = find_next_bit(wc->dirty_bitmap, bitmap_bits, i);
 511                 if (unlikely(i == bitmap_bits))
 512                         break;
 513                 j = find_next_zero_bit(wc->dirty_bitmap, bitmap_bits, i);
 514
 515                 region.bdev = wc->ssd_dev->bdev;
 516                 region.sector = (sector_t)i * (BITMAP_GRANULARITY >> SECTOR_SHIFT);
 517                 region.count = (sector_t)(j - i) * (BITMAP_GRANULARITY >> SECTOR_SHIFT);
 518
 519                 if (unlikely(region.sector >= wc->metadata_sectors))
 520                         break;
 521                 if (unlikely(region.sector + region.count > wc->metadata_sectors))
 522                         region.count = wc->metadata_sectors - region.sector;
 523
 524                 region.sector += wc->start_sector;
 525                 atomic_inc(&endio.count);
 526                 req.bi_op = REQ_OP_WRITE;
 527                 req.bi_op_flags = REQ_SYNC;
 528                 req.mem.type = DM_IO_VMA;
 529                 req.mem.ptr.vma = (char *)wc->memory_map + (size_t)i * BITMAP_GRANULARITY;
 530                 req.client = wc->dm_io;
 531                 req.notify.fn = writecache_notify_io;
 532                 req.notify.context = &endio;
 533
 534                 /* writing via async dm-io (implied by notify.fn above) won't return an error */
 535                 (void) dm_io(&req, 1, &region, NULL);
 536                 i = j;
 537         }
 538
 539         writecache_notify_io(0, &endio);
 540         wait_for_completion_io(&endio.c);
 541
 542         if (wait_for_ios)
 543                 writecache_wait_for_ios(wc, WRITE);
 544
 545         writecache_disk_flush(wc, wc->ssd_dev);
 546
 547         memset(wc->dirty_bitmap, 0, wc->dirty_bitmap_size);
 548 }
 549
 550 static void ssd_commit_superblock(struct dm_writecache *wc)
 551 {
 552         int r;
 553         struct dm_io_region region;
 554         struct dm_io_request req;
 555
 556         region.bdev = wc->ssd_dev->bdev;
 557         region.sector = 0;
 558         region.count = max(4096U, wc->block_size) >> SECTOR_SHIFT;
 559
 560         if (unlikely(region.sector + region.count > wc->metadata_sectors))
 561                 region.count = wc->metadata_sectors - region.sector;
 562
 563         region.sector += wc->start_sector;
 564
 565         req.bi_op = REQ_OP_WRITE;
 566         req.bi_op_flags = REQ_SYNC | REQ_FUA;
 567         req.mem.type = DM_IO_VMA;
 568         req.mem.ptr.vma = (char *)wc->memory_map;
 569         req.client = wc->dm_io;
 570         req.notify.fn = NULL;
 571         req.notify.context = NULL;
 572
 573         r = dm_io(&req, 1, &region, NULL);
 574         if (unlikely(r))
 575                 writecache_error(wc, r, "error writing superblock");
 576 }
 577
 578 static void writecache_commit_flushed(struct dm_writecache *wc, bool wait_for_ios)
 579 {
 580         if (WC_MODE_PMEM(wc))
 581                 pmem_wmb();
 582         else
 583                 ssd_commit_flushed(wc, wait_for_ios);
 584 }
 585
 586 static void writecache_disk_flush(struct dm_writecache *wc, struct dm_dev *dev)
 587 {
 588         int r;
 589         struct dm_io_region region;
 590         struct dm_io_request req;
 591
 592         region.bdev = dev->bdev;
 593         region.sector = 0;
 594         region.count = 0;
 595         req.bi_op = REQ_OP_WRITE;
 596         req.bi_op_flags = REQ_PREFLUSH;
 597         req.mem.type = DM_IO_KMEM;
 598         req.mem.ptr.addr = NULL;
 599         req.client = wc->dm_io;
 600         req.notify.fn = NULL;
 601
 602         r = dm_io(&req, 1, &region, NULL);
 603         if (unlikely(r))
 604                 writecache_error(wc, r, "error flushing metadata: %d", r);
 605 }
 606
 607 #define WFE_RETURN_FOLLOWING    1
 608 #define WFE_LOWEST_SEQ          2
 609
 610 static struct wc_entry *writecache_find_entry(struct dm_writecache *wc,
 611                                               uint64_t block, int flags)
 612 {
 613         struct wc_entry *e;
 614         struct rb_node *node = wc->tree.rb_node;
 615
 616         if (unlikely(!node))
 617                 return NULL;
 618
 619         while (1) {
 620                 e = container_of(node, struct wc_entry, rb_node);
 621                 if (read_original_sector(wc, e) == block)
 622                         break;
 623
 624                 node = (read_original_sector(wc, e) >= block ?
 625                         e->rb_node.rb_left : e->rb_node.rb_right);
 626                 if (unlikely(!node)) {
 627                         if (!(flags & WFE_RETURN_FOLLOWING))
 628                                 return NULL;
 629                         if (read_original_sector(wc, e) >= block) {
 630                                 return e;
 631                         } else {
 632                                 node = rb_next(&e->rb_node);
 633                                 if (unlikely(!node))
 634                                         return NULL;
 635                                 e = container_of(node, struct wc_entry, rb_node);
 636                                 return e;
 637                         }
 638                 }
 639         }
 640
 641         while (1) {
 642                 struct wc_entry *e2;
 643                 if (flags & WFE_LOWEST_SEQ)
 644                         node = rb_prev(&e->rb_node);
 645                 else
 646                         node = rb_next(&e->rb_node);
 647                 if (unlikely(!node))
 648                         return e;
 649                 e2 = container_of(node, struct wc_entry, rb_node);
 650                 if (read_original_sector(wc, e2) != block)
 651                         return e;
 652                 e = e2;
 653         }
 654 }
 655
 656 static void writecache_insert_entry(struct dm_writecache *wc, struct wc_entry *ins)
 657 {
 658         struct wc_entry *e;
 659         struct rb_node **node = &wc->tree.rb_node, *parent = NULL;
 660
 661         while (*node) {
 662                 e = container_of(*node, struct wc_entry, rb_node);
 663                 parent = &e->rb_node;
 664                 if (read_original_sector(wc, e) > read_original_sector(wc, ins))
 665                         node = &parent->rb_left;
 666                 else
 667                         node = &parent->rb_right;
 668         }
 669         rb_link_node(&ins->rb_node, parent, node);
 670         rb_insert_color(&ins->rb_node, &wc->tree);
 671         list_add(&ins->lru, &wc->lru);
 672         ins->age = jiffies;
 673 }
 674
 675 static void writecache_unlink(struct dm_writecache *wc, struct wc_entry *e)
 676 {
 677         list_del(&e->lru);
 678         rb_erase(&e->rb_node, &wc->tree);
 679 }
 680
 681 static void writecache_add_to_freelist(struct dm_writecache *wc, struct wc_entry *e)
 682 {
 683         if (WC_MODE_SORT_FREELIST(wc)) {
 684                 struct rb_node **node = &wc->freetree.rb_node, *parent = NULL;
 685                 if (unlikely(!*node))
 686                         wc->current_free = e;
 687                 while (*node) {
 688                         parent = *node;
 689                         if (&e->rb_node < *node)
 690                                 node = &parent->rb_left;
 691                         else
 692                                 node = &parent->rb_right;
 693                 }
 694                 rb_link_node(&e->rb_node, parent, node);
 695                 rb_insert_color(&e->rb_node, &wc->freetree);
 696         } else {
 697                 list_add_tail(&e->lru, &wc->freelist);
 698         }
 699         wc->freelist_size++;
 700 }
 701
 702 static inline void writecache_verify_watermark(struct dm_writecache *wc)
 703 {
 704         if (unlikely(wc->freelist_size + wc->writeback_size <= wc->freelist_high_watermark))
 705                 queue_work(wc->writeback_wq, &wc->writeback_work);
 706 }
 707
 708 static void writecache_max_age_timer(struct timer_list *t)
 709 {
 710         struct dm_writecache *wc = from_timer(wc, t, max_age_timer);
 711
 712         if (!dm_suspended(wc->ti) && !writecache_has_error(wc)) {
 713                 queue_work(wc->writeback_wq, &wc->writeback_work);
 714                 mod_timer(&wc->max_age_timer, jiffies + wc->max_age / MAX_AGE_DIV);
 715         }
 716 }
 717
 718 static struct wc_entry *writecache_pop_from_freelist(struct dm_writecache *wc, sector_t expected_sector)
 719 {
 720         struct wc_entry *e;
 721
 722         if (WC_MODE_SORT_FREELIST(wc)) {
 723                 struct rb_node *next;
 724                 if (unlikely(!wc->current_free))
 725                         return NULL;
 726                 e = wc->current_free;
 727                 if (expected_sector != (sector_t)-1 && unlikely(cache_sector(wc, e) != expected_sector))
 728                         return NULL;
 729                 next = rb_next(&e->rb_node);
 730                 rb_erase(&e->rb_node, &wc->freetree);
 731                 if (unlikely(!next))
 732                         next = rb_first(&wc->freetree);
 733                 wc->current_free = next ? container_of(next, struct wc_entry, rb_node) : NULL;
 734         } else {
 735                 if (unlikely(list_empty(&wc->freelist)))
 736                         return NULL;
 737                 e = container_of(wc->freelist.next, struct wc_entry, lru);
 738                 if (expected_sector != (sector_t)-1 && unlikely(cache_sector(wc, e) != expected_sector))
 739                         return NULL;
 740                 list_del(&e->lru);
 741         }
 742         wc->freelist_size--;
 743
 744         writecache_verify_watermark(wc);
 745
 746         return e;
 747 }
 748
 749 static void writecache_free_entry(struct dm_writecache *wc, struct wc_entry *e)
 750 {
 751         writecache_unlink(wc, e);
 752         writecache_add_to_freelist(wc, e);
 753         clear_seq_count(wc, e);
 754         writecache_flush_region(wc, memory_entry(wc, e), sizeof(struct wc_memory_entry));
 755         if (unlikely(waitqueue_active(&wc->freelist_wait)))
 756                 wake_up(&wc->freelist_wait);
 757 }
 758
 759 static void writecache_wait_on_freelist(struct dm_writecache *wc)
 760 {
 761         DEFINE_WAIT(wait);
 762
 763         prepare_to_wait(&wc->freelist_wait, &wait, TASK_UNINTERRUPTIBLE);
 764         wc_unlock(wc);
 765         io_schedule();
 766         finish_wait(&wc->freelist_wait, &wait);
 767         wc_lock(wc);
 768 }
 769
 770 static void writecache_poison_lists(struct dm_writecache *wc)
 771 {
 772         /*
 773          * Catch incorrect access to these values while the device is suspended.
 774          */
 775         memset(&wc->tree, -1, sizeof wc->tree);
 776         wc->lru.next = LIST_POISON1;
 777         wc->lru.prev = LIST_POISON2;
 778         wc->freelist.next = LIST_POISON1;
 779         wc->freelist.prev = LIST_POISON2;
 780 }
 781
 782 static void writecache_flush_entry(struct dm_writecache *wc, struct wc_entry *e)
 783 {
 784         writecache_flush_region(wc, memory_entry(wc, e), sizeof(struct wc_memory_entry));
 785         if (WC_MODE_PMEM(wc))
 786                 writecache_flush_region(wc, memory_data(wc, e), wc->block_size);
 787 }
 788
 789 static bool writecache_entry_is_committed(struct dm_writecache *wc, struct wc_entry *e)
 790 {
 791         return read_seq_count(wc, e) < wc->seq_count;
 792 }
 793
 794 static void writecache_flush(struct dm_writecache *wc)
 795 {
 796         struct wc_entry *e, *e2;
 797         bool need_flush_after_free;
 798
 799         wc->uncommitted_blocks = 0;
 800         del_timer(&wc->autocommit_timer);
 801
 802         if (list_empty(&wc->lru))
 803                 return;
 804
 805         e = container_of(wc->lru.next, struct wc_entry, lru);
 806         if (writecache_entry_is_committed(wc, e)) {
 807                 if (wc->overwrote_committed) {
 808                         writecache_wait_for_ios(wc, WRITE);
 809                         writecache_disk_flush(wc, wc->ssd_dev);
 810                         wc->overwrote_committed = false;
 811                 }
 812                 return;
 813         }
 814         while (1) {
 815                 writecache_flush_entry(wc, e);
 816                 if (unlikely(e->lru.next == &wc->lru))
 817                         break;
 818                 e2 = container_of(e->lru.next, struct wc_entry, lru);
 819                 if (writecache_entry_is_committed(wc, e2))
 820                         break;
 821                 e = e2;
 822                 cond_resched();
 823         }
 824         writecache_commit_flushed(wc, true);
 825
 826         wc->seq_count++;
 827         pmem_assign(sb(wc)->seq_count, cpu_to_le64(wc->seq_count));
 828         if (WC_MODE_PMEM(wc))
 829                 writecache_commit_flushed(wc, false);
 830         else
 831                 ssd_commit_superblock(wc);
 832
 833         wc->overwrote_committed = false;
 834
 835         need_flush_after_free = false;
 836         while (1) {
 837                 /* Free another committed entry with lower seq-count */
 838                 struct rb_node *rb_node = rb_prev(&e->rb_node);
 839
 840                 if (rb_node) {
 841                         e2 = container_of(rb_node, struct wc_entry, rb_node);
 842                         if (read_original_sector(wc, e2) == read_original_sector(wc, e) &&
 843                             likely(!e2->write_in_progress)) {
 844                                 writecache_free_entry(wc, e2);
 845                                 need_flush_after_free = true;
 846                         }
 847                 }
 848                 if (unlikely(e->lru.prev == &wc->lru))
 849                         break;
 850                 e = container_of(e->lru.prev, struct wc_entry, lru);
 851                 cond_resched();
 852         }
 853
 854         if (need_flush_after_free)
 855                 writecache_commit_flushed(wc, false);
 856 }
 857
 858 static void writecache_flush_work(struct work_struct *work)
 859 {
 860         struct dm_writecache *wc = container_of(work, struct dm_writecache, flush_work);
 861
 862         wc_lock(wc);
 863         writecache_flush(wc);
 864         wc_unlock(wc);
 865 }
 866
 867 static void writecache_autocommit_timer(struct timer_list *t)
 868 {
 869         struct dm_writecache *wc = from_timer(wc, t, autocommit_timer);
 870         if (!writecache_has_error(wc))
 871                 queue_work(wc->writeback_wq, &wc->flush_work);
 872 }
 873
 874 static void writecache_schedule_autocommit(struct dm_writecache *wc)
 875 {
 876         if (!timer_pending(&wc->autocommit_timer))
 877                 mod_timer(&wc->autocommit_timer, jiffies + wc->autocommit_jiffies);
 878 }
 879
 880 static void writecache_discard(struct dm_writecache *wc, sector_t start, sector_t end)
 881 {
 882         struct wc_entry *e;
 883         bool discarded_something = false;
 884
 885         e = writecache_find_entry(wc, start, WFE_RETURN_FOLLOWING | WFE_LOWEST_SEQ);
 886         if (unlikely(!e))
 887                 return;
 888
 889         while (read_original_sector(wc, e) < end) {
 890                 struct rb_node *node = rb_next(&e->rb_node);
 891
 892                 if (likely(!e->write_in_progress)) {
 893                         if (!discarded_something) {
 894                                 if (!WC_MODE_PMEM(wc)) {
 895                                         writecache_wait_for_ios(wc, READ);
 896                                         writecache_wait_for_ios(wc, WRITE);
 897                                 }
 898                                 discarded_something = true;
 899                         }
 900                         if (!writecache_entry_is_committed(wc, e))
 901                                 wc->uncommitted_blocks--;
 902                         writecache_free_entry(wc, e);
 903                 }
 904
 905                 if (unlikely(!node))
 906                         break;
 907
 908                 e = container_of(node, struct wc_entry, rb_node);
 909         }
 910
 911         if (discarded_something)
 912                 writecache_commit_flushed(wc, false);
 913 }
 914
 915 static bool writecache_wait_for_writeback(struct dm_writecache *wc)
 916 {
 917         if (wc->writeback_size) {
 918                 writecache_wait_on_freelist(wc);
 919                 return true;
 920         }
 921         return false;
 922 }
 923
 924 static void writecache_suspend(struct dm_target *ti)
 925 {
 926         struct dm_writecache *wc = ti->private;
 927         bool flush_on_suspend;
 928
 929         del_timer_sync(&wc->autocommit_timer);
 930         del_timer_sync(&wc->max_age_timer);
 931
 932         wc_lock(wc);
 933         writecache_flush(wc);
 934         flush_on_suspend = wc->flush_on_suspend;
 935         if (flush_on_suspend) {
 936                 wc->flush_on_suspend = false;
 937                 wc->writeback_all++;
 938                 queue_work(wc->writeback_wq, &wc->writeback_work);
 939         }
 940         wc_unlock(wc);
 941
 942         drain_workqueue(wc->writeback_wq);
 943
 944         wc_lock(wc);
 945         if (flush_on_suspend)
 946                 wc->writeback_all--;
 947         while (writecache_wait_for_writeback(wc));
 948
 949         if (WC_MODE_PMEM(wc))
 950                 persistent_memory_flush_cache(wc->memory_map, wc->memory_map_size);
 951
 952         writecache_poison_lists(wc);
 953
 954         wc_unlock(wc);
 955 }
 956
 957 static int writecache_alloc_entries(struct dm_writecache *wc)
 958 {
 959         size_t b;
 960
 961         if (wc->entries)
 962                 return 0;
 963         wc->entries = vmalloc(array_size(sizeof(struct wc_entry), wc->n_blocks));
 964         if (!wc->entries)
 965                 return -ENOMEM;
 966         for (b = 0; b < wc->n_blocks; b++) {
 967                 struct wc_entry *e = &wc->entries[b];
 968                 e->index = b;
 969                 e->write_in_progress = false;
 970                 cond_resched();
 971         }
 972
 973         return 0;
 974 }
 975
 976 static int writecache_read_metadata(struct dm_writecache *wc, sector_t n_sectors)
 977 {
 978         struct dm_io_region region;
 979         struct dm_io_request req;
 980
 981         region.bdev = wc->ssd_dev->bdev;
 982         region.sector = wc->start_sector;
 983         region.count = n_sectors;
 984         req.bi_op = REQ_OP_READ;
 985         req.bi_op_flags = REQ_SYNC;
 986         req.mem.type = DM_IO_VMA;
 987         req.mem.ptr.vma = (char *)wc->memory_map;
 988         req.client = wc->dm_io;
 989         req.notify.fn = NULL;
 990
 991         return dm_io(&req, 1, &region, NULL);
 992 }
 993
 994 static void writecache_resume(struct dm_target *ti)
 995 {
 996         struct dm_writecache *wc = ti->private;
 997         size_t b;
 998         bool need_flush = false;
 999         __le64 sb_seq_count;
1000         int r;
1001
1002         wc_lock(wc);
1003
1004         wc->data_device_sectors = bdev_nr_sectors(wc->dev->bdev);
1005
1006         if (WC_MODE_PMEM(wc)) {
1007                 persistent_memory_invalidate_cache(wc->memory_map, wc->memory_map_size);
1008         } else {
1009                 r = writecache_read_metadata(wc, wc->metadata_sectors);
1010                 if (r) {
1011                         size_t sb_entries_offset;
1012                         writecache_error(wc, r, "unable to read metadata: %d", r);
1013                         sb_entries_offset = offsetof(struct wc_memory_superblock, entries);
1014                         memset((char *)wc->memory_map + sb_entries_offset, -1,
1015                                (wc->metadata_sectors << SECTOR_SHIFT) - sb_entries_offset);
1016                 }
1017         }
1018
1019         wc->tree = RB_ROOT;
1020         INIT_LIST_HEAD(&wc->lru);
1021         if (WC_MODE_SORT_FREELIST(wc)) {
1022                 wc->freetree = RB_ROOT;
1023                 wc->current_free = NULL;
1024         } else {
1025                 INIT_LIST_HEAD(&wc->freelist);
1026         }
1027         wc->freelist_size = 0;
1028
1029         r = copy_mc_to_kernel(&sb_seq_count, &sb(wc)->seq_count,
1030                               sizeof(uint64_t));
1031         if (r) {
1032                 writecache_error(wc, r, "hardware memory error when reading superblock: %d", r);
1033                 sb_seq_count = cpu_to_le64(0);
1034         }
1035         wc->seq_count = le64_to_cpu(sb_seq_count);
1036
1037 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
1038         for (b = 0; b < wc->n_blocks; b++) {
1039                 struct wc_entry *e = &wc->entries[b];
1040                 struct wc_memory_entry wme;
1041                 if (writecache_has_error(wc)) {
1042                         e->original_sector = -1;
1043                         e->seq_count = -1;
1044                         continue;
1045                 }
1046                 r = copy_mc_to_kernel(&wme, memory_entry(wc, e),
1047                                       sizeof(struct wc_memory_entry));
1048                 if (r) {
1049                         writecache_error(wc, r, "hardware memory error when reading metadata entry %lu: %d",
1050                                          (unsigned long)b, r);
1051                         e->original_sector = -1;
1052                         e->seq_count = -1;
1053                 } else {
1054                         e->original_sector = le64_to_cpu(wme.original_sector);
1055                         e->seq_count = le64_to_cpu(wme.seq_count);
1056                 }
1057                 cond_resched();
1058         }
1059 #endif
1060         for (b = 0; b < wc->n_blocks; b++) {
1061                 struct wc_entry *e = &wc->entries[b];
1062                 if (!writecache_entry_is_committed(wc, e)) {
1063                         if (read_seq_count(wc, e) != -1) {
1064 erase_this:
1065                                 clear_seq_count(wc, e);
1066                                 need_flush = true;
1067                         }
1068                         writecache_add_to_freelist(wc, e);
1069                 } else {
1070                         struct wc_entry *old;
1071
1072                         old = writecache_find_entry(wc, read_original_sector(wc, e), 0);
1073                         if (!old) {
1074                                 writecache_insert_entry(wc, e);
1075                         } else {
1076                                 if (read_seq_count(wc, old) == read_seq_count(wc, e)) {
1077                                         writecache_error(wc, -EINVAL,
1078                                                  "two identical entries, position %llu, sector %llu, sequence %llu",
1079                                                  (unsigned long long)b, (unsigned long long)read_original_sector(wc, e),
1080                                                  (unsigned long long)read_seq_count(wc, e));
1081                                 }
1082                                 if (read_seq_count(wc, old) > read_seq_count(wc, e)) {
1083                                         goto erase_this;
1084                                 } else {
1085                                         writecache_free_entry(wc, old);
1086                                         writecache_insert_entry(wc, e);
1087                                         need_flush = true;
1088                                 }
1089                         }
1090                 }
1091                 cond_resched();
1092         }
1093
1094         if (need_flush) {
1095                 writecache_flush_all_metadata(wc);
1096                 writecache_commit_flushed(wc, false);
1097         }
1098
1099         writecache_verify_watermark(wc);
1100
1101         if (wc->max_age != MAX_AGE_UNSPECIFIED)
1102                 mod_timer(&wc->max_age_timer, jiffies + wc->max_age / MAX_AGE_DIV);
1103
1104         wc_unlock(wc);
1105 }
1106
1107 static int process_flush_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1108 {
1109         if (argc != 1)
1110                 return -EINVAL;
1111
1112         wc_lock(wc);
1113         if (dm_suspended(wc->ti)) {
1114                 wc_unlock(wc);
1115                 return -EBUSY;
1116         }
1117         if (writecache_has_error(wc)) {
1118                 wc_unlock(wc);
1119                 return -EIO;
1120         }
1121
1122         writecache_flush(wc);
1123         wc->writeback_all++;
1124         queue_work(wc->writeback_wq, &wc->writeback_work);
1125         wc_unlock(wc);
1126
1127         flush_workqueue(wc->writeback_wq);
1128
1129         wc_lock(wc);
1130         wc->writeback_all--;
1131         if (writecache_has_error(wc)) {
1132                 wc_unlock(wc);
1133                 return -EIO;
1134         }
1135         wc_unlock(wc);
1136
1137         return 0;
1138 }
1139
1140 static int process_flush_on_suspend_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1141 {
1142         if (argc != 1)
1143                 return -EINVAL;
1144
1145         wc_lock(wc);
1146         wc->flush_on_suspend = true;
1147         wc_unlock(wc);
1148
1149         return 0;
1150 }
1151
1152 static void activate_cleaner(struct dm_writecache *wc)
1153 {
1154         wc->flush_on_suspend = true;
1155         wc->cleaner = true;
1156         wc->freelist_high_watermark = wc->n_blocks;
1157         wc->freelist_low_watermark = wc->n_blocks;
1158 }
1159
1160 static int process_cleaner_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1161 {
1162         if (argc != 1)
1163                 return -EINVAL;
1164
1165         wc_lock(wc);
1166         activate_cleaner(wc);
1167         if (!dm_suspended(wc->ti))
1168                 writecache_verify_watermark(wc);
1169         wc_unlock(wc);
1170
1171         return 0;
1172 }
1173
1174 static int process_clear_stats_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1175 {
1176         if (argc != 1)
1177                 return -EINVAL;
1178
1179         wc_lock(wc);
1180         memset(&wc->stats, 0, sizeof wc->stats);
1181         wc_unlock(wc);
1182
1183         return 0;
1184 }
1185
1186 static int writecache_message(struct dm_target *ti, unsigned argc, char **argv,
1187                               char *result, unsigned maxlen)
1188 {
1189         int r = -EINVAL;
1190         struct dm_writecache *wc = ti->private;
1191
1192         if (!strcasecmp(argv[0], "flush"))
1193                 r = process_flush_mesg(argc, argv, wc);
1194         else if (!strcasecmp(argv[0], "flush_on_suspend"))
1195                 r = process_flush_on_suspend_mesg(argc, argv, wc);
1196         else if (!strcasecmp(argv[0], "cleaner"))
1197                 r = process_cleaner_mesg(argc, argv, wc);
1198         else if (!strcasecmp(argv[0], "clear_stats"))
1199                 r = process_clear_stats_mesg(argc, argv, wc);
1200         else
1201                 DMERR("unrecognised message received: %s", argv[0]);
1202
1203         return r;
1204 }
1205
1206 static void memcpy_flushcache_optimized(void *dest, void *source, size_t size)
1207 {
1208         /*
1209          * clflushopt performs better with block size 1024, 2048, 4096
1210          * non-temporal stores perform better with block size 512
1211          *
1212          * block size   512             1024            2048            4096
1213          * movnti       496 MB/s        642 MB/s        725 MB/s        744 MB/s
1214          * clflushopt   373 MB/s        688 MB/s        1.1 GB/s        1.2 GB/s
1215          *
1216          * We see that movnti performs better for 512-byte blocks, and
1217          * clflushopt performs better for 1024-byte and larger blocks. So, we
1218          * prefer clflushopt for sizes >= 768.
1219          *
1220          * NOTE: this happens to be the case now (with dm-writecache's single
1221          * threaded model) but re-evaluate this once memcpy_flushcache() is
1222          * enabled to use movdir64b which might invalidate this performance
1223          * advantage seen with cache-allocating-writes plus flushing.
1224          */
1225 #ifdef CONFIG_X86
1226         if (static_cpu_has(X86_FEATURE_CLFLUSHOPT) &&
1227             likely(boot_cpu_data.x86_clflush_size == 64) &&
1228             likely(size >= 768)) {
1229                 do {
1230                         memcpy((void *)dest, (void *)source, 64);
1231                         clflushopt((void *)dest);
1232                         dest += 64;
1233                         source += 64;
1234                         size -= 64;
1235                 } while (size >= 64);
1236                 return;
1237         }
1238 #endif
1239         memcpy_flushcache(dest, source, size);
1240 }
1241
1242 static void bio_copy_block(struct dm_writecache *wc, struct bio *bio, void *data)
1243 {
1244         void *buf;
1245         unsigned size;
1246         int rw = bio_data_dir(bio);
1247         unsigned remaining_size = wc->block_size;
1248
1249         do {
1250                 struct bio_vec bv = bio_iter_iovec(bio, bio->bi_iter);
1251                 buf = bvec_kmap_local(&bv);
1252                 size = bv.bv_len;
1253                 if (unlikely(size > remaining_size))
1254                         size = remaining_size;
1255
1256                 if (rw == READ) {
1257                         int r;
1258                         r = copy_mc_to_kernel(buf, data, size);
1259                         flush_dcache_page(bio_page(bio));
1260                         if (unlikely(r)) {
1261                                 writecache_error(wc, r, "hardware memory error when reading data: %d", r);
1262                                 bio->bi_status = BLK_STS_IOERR;
1263                         }
1264                 } else {
1265                         flush_dcache_page(bio_page(bio));
1266                         memcpy_flushcache_optimized(data, buf, size);
1267                 }
1268
1269                 kunmap_local(buf);
1270
1271                 data = (char *)data + size;
1272                 remaining_size -= size;
1273                 bio_advance(bio, size);
1274         } while (unlikely(remaining_size));
1275 }
1276
1277 static int writecache_flush_thread(void *data)
1278 {
1279         struct dm_writecache *wc = data;
1280
1281         while (1) {
1282                 struct bio *bio;
1283
1284                 wc_lock(wc);
1285                 bio = bio_list_pop(&wc->flush_list);
1286                 if (!bio) {
1287                         set_current_state(TASK_INTERRUPTIBLE);
1288                         wc_unlock(wc);
1289
1290                         if (unlikely(kthread_should_stop())) {
1291                                 set_current_state(TASK_RUNNING);
1292                                 break;
1293                         }
1294
1295                         schedule();
1296                         continue;
1297                 }
1298
1299                 if (bio_op(bio) == REQ_OP_DISCARD) {
1300                         writecache_discard(wc, bio->bi_iter.bi_sector,
1301                                            bio_end_sector(bio));
1302                         wc_unlock(wc);
1303                         bio_set_dev(bio, wc->dev->bdev);
1304                         submit_bio_noacct(bio);
1305                 } else {
1306                         writecache_flush(wc);
1307                         wc_unlock(wc);
1308                         if (writecache_has_error(wc))
1309                                 bio->bi_status = BLK_STS_IOERR;
1310                         bio_endio(bio);
1311                 }
1312         }
1313
1314         return 0;
1315 }
1316
1317 static void writecache_offload_bio(struct dm_writecache *wc, struct bio *bio)
1318 {
1319         if (bio_list_empty(&wc->flush_list))
1320                 wake_up_process(wc->flush_thread);
1321         bio_list_add(&wc->flush_list, bio);
1322 }
1323
1324 enum wc_map_op {
1325         WC_MAP_SUBMIT,
1326         WC_MAP_REMAP,
1327         WC_MAP_REMAP_ORIGIN,
1328         WC_MAP_RETURN,
1329         WC_MAP_ERROR,
1330 };
1331
1332 static enum wc_map_op writecache_map_remap_origin(struct dm_writecache *wc, struct bio *bio,
1333                                                   struct wc_entry *e)
1334 {
1335         if (e) {
1336                 sector_t next_boundary =
1337                         read_original_sector(wc, e) - bio->bi_iter.bi_sector;
1338                 if (next_boundary < bio->bi_iter.bi_size >> SECTOR_SHIFT)
1339                         dm_accept_partial_bio(bio, next_boundary);
1340         }
1341
1342         return WC_MAP_REMAP_ORIGIN;
1343 }
1344
1345 static enum wc_map_op writecache_map_read(struct dm_writecache *wc, struct bio *bio)
1346 {
1347         enum wc_map_op map_op;
1348         struct wc_entry *e;
1349
1350 read_next_block:
1351         wc->stats.reads++;
1352         e = writecache_find_entry(wc, bio->bi_iter.bi_sector, WFE_RETURN_FOLLOWING);
1353         if (e && read_original_sector(wc, e) == bio->bi_iter.bi_sector) {
1354                 wc->stats.read_hits++;
1355                 if (WC_MODE_PMEM(wc)) {
1356                         bio_copy_block(wc, bio, memory_data(wc, e));
1357                         if (bio->bi_iter.bi_size)
1358                                 goto read_next_block;
1359                         map_op = WC_MAP_SUBMIT;
1360                 } else {
1361                         dm_accept_partial_bio(bio, wc->block_size >> SECTOR_SHIFT);
1362                         bio_set_dev(bio, wc->ssd_dev->bdev);
1363                         bio->bi_iter.bi_sector = cache_sector(wc, e);
1364                         if (!writecache_entry_is_committed(wc, e))
1365                                 writecache_wait_for_ios(wc, WRITE);
1366                         map_op = WC_MAP_REMAP;
1367                 }
1368         } else {
1369                 map_op = writecache_map_remap_origin(wc, bio, e);
1370         }
1371
1372         return map_op;
1373 }
1374
1375 static enum wc_map_op writecache_bio_copy_ssd(struct dm_writecache *wc, struct bio *bio,
1376                                               struct wc_entry *e, bool search_used)
1377 {
1378         unsigned bio_size = wc->block_size;
1379         sector_t start_cache_sec = cache_sector(wc, e);
1380         sector_t current_cache_sec = start_cache_sec + (bio_size >> SECTOR_SHIFT);
1381
1382         while (bio_size < bio->bi_iter.bi_size) {
1383                 if (!search_used) {
1384                         struct wc_entry *f = writecache_pop_from_freelist(wc, current_cache_sec);
1385                         if (!f)
1386                                 break;
1387                         write_original_sector_seq_count(wc, f, bio->bi_iter.bi_sector +
1388                                                         (bio_size >> SECTOR_SHIFT), wc->seq_count);
1389                         writecache_insert_entry(wc, f);
1390                         wc->uncommitted_blocks++;
1391                 } else {
1392                         struct wc_entry *f;
1393                         struct rb_node *next = rb_next(&e->rb_node);
1394                         if (!next)
1395                                 break;
1396                         f = container_of(next, struct wc_entry, rb_node);
1397                         if (f != e + 1)
1398                                 break;
1399                         if (read_original_sector(wc, f) !=
1400                             read_original_sector(wc, e) + (wc->block_size >> SECTOR_SHIFT))
1401                                 break;
1402                         if (unlikely(f->write_in_progress))
1403                                 break;
1404                         if (writecache_entry_is_committed(wc, f))
1405                                 wc->overwrote_committed = true;
1406                         e = f;
1407                 }
1408                 bio_size += wc->block_size;
1409                 current_cache_sec += wc->block_size >> SECTOR_SHIFT;
1410         }
1411
1412         bio_set_dev(bio, wc->ssd_dev->bdev);
1413         bio->bi_iter.bi_sector = start_cache_sec;
1414         dm_accept_partial_bio(bio, bio_size >> SECTOR_SHIFT);
1415
1416         if (unlikely(wc->uncommitted_blocks >= wc->autocommit_blocks)) {
1417                 wc->uncommitted_blocks = 0;
1418                 queue_work(wc->writeback_wq, &wc->flush_work);
1419         } else {
1420                 writecache_schedule_autocommit(wc);
1421         }
1422
1423         return WC_MAP_REMAP;
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                 wc->stats.writes++;
1434                 if (writecache_has_error(wc))
1435                         return WC_MAP_ERROR;
1436                 e = writecache_find_entry(wc, bio->bi_iter.bi_sector, 0);
1437                 if (e) {
1438                         if (!writecache_entry_is_committed(wc, e)) {
1439                                 wc->stats.write_hits_uncommitted++;
1440                                 search_used = true;
1441                                 goto bio_copy;
1442                         }
1443                         wc->stats.write_hits_committed++;
1444                         if (!WC_MODE_PMEM(wc) && !e->write_in_progress) {
1445                                 wc->overwrote_committed = true;
1446                                 search_used = true;
1447                                 goto bio_copy;
1448                         }
1449                         found_entry = true;
1450                 } else {
1451                         if (unlikely(wc->cleaner) ||
1452                             (wc->metadata_only && !(bio->bi_opf & REQ_META)))
1453                                 goto direct_write;
1454                 }
1455                 e = writecache_pop_from_freelist(wc, (sector_t)-1);
1456                 if (unlikely(!e)) {
1457                         if (!WC_MODE_PMEM(wc) && !found_entry) {
1458 direct_write:
1459                                 wc->stats.writes_around++;
1460                                 e = writecache_find_entry(wc, bio->bi_iter.bi_sector, WFE_RETURN_FOLLOWING);
1461                                 return writecache_map_remap_origin(wc, bio, e);
1462                         }
1463                         wc->stats.writes_blocked_on_freelist++;
1464                         writecache_wait_on_freelist(wc);
1465                         continue;
1466                 }
1467                 write_original_sector_seq_count(wc, e, bio->bi_iter.bi_sector, wc->seq_count);
1468                 writecache_insert_entry(wc, e);
1469                 wc->uncommitted_blocks++;
1470                 wc->stats.writes_allocate++;
1471 bio_copy:
1472                 if (WC_MODE_PMEM(wc))
1473                         bio_copy_block(wc, bio, memory_data(wc, e));
1474                 else
1475                         return writecache_bio_copy_ssd(wc, bio, e, search_used);
1476         } while (bio->bi_iter.bi_size);
1477
1478         if (unlikely(bio->bi_opf & REQ_FUA || wc->uncommitted_blocks >= wc->autocommit_blocks))
1479                 writecache_flush(wc);
1480         else
1481                 writecache_schedule_autocommit(wc);
1482
1483         return WC_MAP_SUBMIT;
1484 }
1485
1486 static enum wc_map_op writecache_map_flush(struct dm_writecache *wc, struct bio *bio)
1487 {
1488         if (writecache_has_error(wc))
1489                 return WC_MAP_ERROR;
1490
1491         if (WC_MODE_PMEM(wc)) {
1492                 wc->stats.flushes++;
1493                 writecache_flush(wc);
1494                 if (writecache_has_error(wc))
1495                         return WC_MAP_ERROR;
1496                 else if (unlikely(wc->cleaner) || unlikely(wc->metadata_only))
1497                         return WC_MAP_REMAP_ORIGIN;
1498                 return WC_MAP_SUBMIT;
1499         }
1500         /* SSD: */
1501         if (dm_bio_get_target_bio_nr(bio))
1502                 return WC_MAP_REMAP_ORIGIN;
1503         wc->stats.flushes++;
1504         writecache_offload_bio(wc, bio);
1505         return WC_MAP_RETURN;
1506 }
1507
1508 static enum wc_map_op writecache_map_discard(struct dm_writecache *wc, struct bio *bio)
1509 {
1510         wc->stats.discards++;
1511
1512         if (writecache_has_error(wc))
1513                 return WC_MAP_ERROR;
1514
1515         if (WC_MODE_PMEM(wc)) {
1516                 writecache_discard(wc, bio->bi_iter.bi_sector, bio_end_sector(bio));
1517                 return WC_MAP_REMAP_ORIGIN;
1518         }
1519         /* SSD: */
1520         writecache_offload_bio(wc, bio);
1521         return WC_MAP_RETURN;
1522 }
1523
1524 static int writecache_map(struct dm_target *ti, struct bio *bio)
1525 {
1526         struct dm_writecache *wc = ti->private;
1527         enum wc_map_op map_op;
1528
1529         bio->bi_private = NULL;
1530
1531         wc_lock(wc);
1532
1533         if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1534                 map_op = writecache_map_flush(wc, bio);
1535                 goto done;
1536         }
1537
1538         bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
1539
1540         if (unlikely((((unsigned)bio->bi_iter.bi_sector | bio_sectors(bio)) &
1541                                 (wc->block_size / 512 - 1)) != 0)) {
1542                 DMERR("I/O is not aligned, sector %llu, size %u, block size %u",
1543                       (unsigned long long)bio->bi_iter.bi_sector,
1544                       bio->bi_iter.bi_size, wc->block_size);
1545                 map_op = WC_MAP_ERROR;
1546                 goto done;
1547         }
1548
1549         if (unlikely(bio_op(bio) == REQ_OP_DISCARD)) {
1550                 map_op = writecache_map_discard(wc, bio);
1551                 goto done;
1552         }
1553
1554         if (bio_data_dir(bio) == READ)
1555                 map_op = writecache_map_read(wc, bio);
1556         else
1557                 map_op = writecache_map_write(wc, bio);
1558 done:
1559         switch (map_op) {
1560         case WC_MAP_REMAP_ORIGIN:
1561                 if (likely(wc->pause != 0)) {
1562                         if (bio_op(bio) == REQ_OP_WRITE) {
1563                                 dm_iot_io_begin(&wc->iot, 1);
1564                                 bio->bi_private = (void *)2;
1565                         }
1566                 }
1567                 bio_set_dev(bio, wc->dev->bdev);
1568                 wc_unlock(wc);
1569                 return DM_MAPIO_REMAPPED;
1570
1571         case WC_MAP_REMAP:
1572                 /* make sure that writecache_end_io decrements bio_in_progress: */
1573                 bio->bi_private = (void *)1;
1574                 atomic_inc(&wc->bio_in_progress[bio_data_dir(bio)]);
1575                 wc_unlock(wc);
1576                 return DM_MAPIO_REMAPPED;
1577
1578         case WC_MAP_SUBMIT:
1579                 wc_unlock(wc);
1580                 bio_endio(bio);
1581                 return DM_MAPIO_SUBMITTED;
1582
1583         case WC_MAP_RETURN:
1584                 wc_unlock(wc);
1585                 return DM_MAPIO_SUBMITTED;
1586
1587         case WC_MAP_ERROR:
1588                 wc_unlock(wc);
1589                 bio_io_error(bio);
1590                 return DM_MAPIO_SUBMITTED;
1591
1592         default:
1593                 BUG();
1594                 return -1;
1595         }
1596 }
1597
1598 static int writecache_end_io(struct dm_target *ti, struct bio *bio, blk_status_t *status)
1599 {
1600         struct dm_writecache *wc = ti->private;
1601
1602         if (bio->bi_private == (void *)1) {
1603                 int dir = bio_data_dir(bio);
1604                 if (atomic_dec_and_test(&wc->bio_in_progress[dir]))
1605                         if (unlikely(waitqueue_active(&wc->bio_in_progress_wait[dir])))
1606                                 wake_up(&wc->bio_in_progress_wait[dir]);
1607         } else if (bio->bi_private == (void *)2) {
1608                 dm_iot_io_end(&wc->iot, 1);
1609         }
1610         return 0;
1611 }
1612
1613 static int writecache_iterate_devices(struct dm_target *ti,
1614                                       iterate_devices_callout_fn fn, void *data)
1615 {
1616         struct dm_writecache *wc = ti->private;
1617
1618         return fn(ti, wc->dev, 0, ti->len, data);
1619 }
1620
1621 static void writecache_io_hints(struct dm_target *ti, struct queue_limits *limits)
1622 {
1623         struct dm_writecache *wc = ti->private;
1624
1625         if (limits->logical_block_size < wc->block_size)
1626                 limits->logical_block_size = wc->block_size;
1627
1628         if (limits->physical_block_size < wc->block_size)
1629                 limits->physical_block_size = wc->block_size;
1630
1631         if (limits->io_min < wc->block_size)
1632                 limits->io_min = wc->block_size;
1633 }
1634
1635
1636 static void writecache_writeback_endio(struct bio *bio)
1637 {
1638         struct writeback_struct *wb = container_of(bio, struct writeback_struct, bio);
1639         struct dm_writecache *wc = wb->wc;
1640         unsigned long flags;
1641
1642         raw_spin_lock_irqsave(&wc->endio_list_lock, flags);
1643         if (unlikely(list_empty(&wc->endio_list)))
1644                 wake_up_process(wc->endio_thread);
1645         list_add_tail(&wb->endio_entry, &wc->endio_list);
1646         raw_spin_unlock_irqrestore(&wc->endio_list_lock, flags);
1647 }
1648
1649 static void writecache_copy_endio(int read_err, unsigned long write_err, void *ptr)
1650 {
1651         struct copy_struct *c = ptr;
1652         struct dm_writecache *wc = c->wc;
1653
1654         c->error = likely(!(read_err | write_err)) ? 0 : -EIO;
1655
1656         raw_spin_lock_irq(&wc->endio_list_lock);
1657         if (unlikely(list_empty(&wc->endio_list)))
1658                 wake_up_process(wc->endio_thread);
1659         list_add_tail(&c->endio_entry, &wc->endio_list);
1660         raw_spin_unlock_irq(&wc->endio_list_lock);
1661 }
1662
1663 static void __writecache_endio_pmem(struct dm_writecache *wc, struct list_head *list)
1664 {
1665         unsigned i;
1666         struct writeback_struct *wb;
1667         struct wc_entry *e;
1668         unsigned long n_walked = 0;
1669
1670         do {
1671                 wb = list_entry(list->next, struct writeback_struct, endio_entry);
1672                 list_del(&wb->endio_entry);
1673
1674                 if (unlikely(wb->bio.bi_status != BLK_STS_OK))
1675                         writecache_error(wc, blk_status_to_errno(wb->bio.bi_status),
1676                                         "write error %d", wb->bio.bi_status);
1677                 i = 0;
1678                 do {
1679                         e = wb->wc_list[i];
1680                         BUG_ON(!e->write_in_progress);
1681                         e->write_in_progress = false;
1682                         INIT_LIST_HEAD(&e->lru);
1683                         if (!writecache_has_error(wc))
1684                                 writecache_free_entry(wc, e);
1685                         BUG_ON(!wc->writeback_size);
1686                         wc->writeback_size--;
1687                         n_walked++;
1688                         if (unlikely(n_walked >= ENDIO_LATENCY)) {
1689                                 writecache_commit_flushed(wc, false);
1690                                 wc_unlock(wc);
1691                                 wc_lock(wc);
1692                                 n_walked = 0;
1693                         }
1694                 } while (++i < wb->wc_list_n);
1695
1696                 if (wb->wc_list != wb->wc_list_inline)
1697                         kfree(wb->wc_list);
1698                 bio_put(&wb->bio);
1699         } while (!list_empty(list));
1700 }
1701
1702 static void __writecache_endio_ssd(struct dm_writecache *wc, struct list_head *list)
1703 {
1704         struct copy_struct *c;
1705         struct wc_entry *e;
1706
1707         do {
1708                 c = list_entry(list->next, struct copy_struct, endio_entry);
1709                 list_del(&c->endio_entry);
1710
1711                 if (unlikely(c->error))
1712                         writecache_error(wc, c->error, "copy error");
1713
1714                 e = c->e;
1715                 do {
1716                         BUG_ON(!e->write_in_progress);
1717                         e->write_in_progress = false;
1718                         INIT_LIST_HEAD(&e->lru);
1719                         if (!writecache_has_error(wc))
1720                                 writecache_free_entry(wc, e);
1721
1722                         BUG_ON(!wc->writeback_size);
1723                         wc->writeback_size--;
1724                         e++;
1725                 } while (--c->n_entries);
1726                 mempool_free(c, &wc->copy_pool);
1727         } while (!list_empty(list));
1728 }
1729
1730 static int writecache_endio_thread(void *data)
1731 {
1732         struct dm_writecache *wc = data;
1733
1734         while (1) {
1735                 struct list_head list;
1736
1737                 raw_spin_lock_irq(&wc->endio_list_lock);
1738                 if (!list_empty(&wc->endio_list))
1739                         goto pop_from_list;
1740                 set_current_state(TASK_INTERRUPTIBLE);
1741                 raw_spin_unlock_irq(&wc->endio_list_lock);
1742
1743                 if (unlikely(kthread_should_stop())) {
1744                         set_current_state(TASK_RUNNING);
1745                         break;
1746                 }
1747
1748                 schedule();
1749
1750                 continue;
1751
1752 pop_from_list:
1753                 list = wc->endio_list;
1754                 list.next->prev = list.prev->next = &list;
1755                 INIT_LIST_HEAD(&wc->endio_list);
1756                 raw_spin_unlock_irq(&wc->endio_list_lock);
1757
1758                 if (!WC_MODE_FUA(wc))
1759                         writecache_disk_flush(wc, wc->dev);
1760
1761                 wc_lock(wc);
1762
1763                 if (WC_MODE_PMEM(wc)) {
1764                         __writecache_endio_pmem(wc, &list);
1765                 } else {
1766                         __writecache_endio_ssd(wc, &list);
1767                         writecache_wait_for_ios(wc, READ);
1768                 }
1769
1770                 writecache_commit_flushed(wc, false);
1771
1772                 wc_unlock(wc);
1773         }
1774
1775         return 0;
1776 }
1777
1778 static bool wc_add_block(struct writeback_struct *wb, struct wc_entry *e)
1779 {
1780         struct dm_writecache *wc = wb->wc;
1781         unsigned block_size = wc->block_size;
1782         void *address = memory_data(wc, e);
1783
1784         persistent_memory_flush_cache(address, block_size);
1785
1786         if (unlikely(bio_end_sector(&wb->bio) >= wc->data_device_sectors))
1787                 return true;
1788
1789         return bio_add_page(&wb->bio, persistent_memory_page(address),
1790                             block_size, persistent_memory_page_offset(address)) != 0;
1791 }
1792
1793 struct writeback_list {
1794         struct list_head list;
1795         size_t size;
1796 };
1797
1798 static void __writeback_throttle(struct dm_writecache *wc, struct writeback_list *wbl)
1799 {
1800         if (unlikely(wc->max_writeback_jobs)) {
1801                 if (READ_ONCE(wc->writeback_size) - wbl->size >= wc->max_writeback_jobs) {
1802                         wc_lock(wc);
1803                         while (wc->writeback_size - wbl->size >= wc->max_writeback_jobs)
1804                                 writecache_wait_on_freelist(wc);
1805                         wc_unlock(wc);
1806                 }
1807         }
1808         cond_resched();
1809 }
1810
1811 static void __writecache_writeback_pmem(struct dm_writecache *wc, struct writeback_list *wbl)
1812 {
1813         struct wc_entry *e, *f;
1814         struct bio *bio;
1815         struct writeback_struct *wb;
1816         unsigned max_pages;
1817
1818         while (wbl->size) {
1819                 wbl->size--;
1820                 e = container_of(wbl->list.prev, struct wc_entry, lru);
1821                 list_del(&e->lru);
1822
1823                 max_pages = e->wc_list_contiguous;
1824
1825                 bio = bio_alloc_bioset(wc->dev->bdev, max_pages, REQ_OP_WRITE,
1826                                        GFP_NOIO, &wc->bio_set);
1827                 wb = container_of(bio, struct writeback_struct, bio);
1828                 wb->wc = wc;
1829                 bio->bi_end_io = writecache_writeback_endio;
1830                 bio->bi_iter.bi_sector = read_original_sector(wc, e);
1831                 if (max_pages <= WB_LIST_INLINE ||
1832                     unlikely(!(wb->wc_list = kmalloc_array(max_pages, sizeof(struct wc_entry *),
1833                                                            GFP_NOIO | __GFP_NORETRY |
1834                                                            __GFP_NOMEMALLOC | __GFP_NOWARN)))) {
1835                         wb->wc_list = wb->wc_list_inline;
1836                         max_pages = WB_LIST_INLINE;
1837                 }
1838
1839                 BUG_ON(!wc_add_block(wb, e));
1840
1841                 wb->wc_list[0] = e;
1842                 wb->wc_list_n = 1;
1843
1844                 while (wbl->size && wb->wc_list_n < max_pages) {
1845                         f = container_of(wbl->list.prev, struct wc_entry, lru);
1846                         if (read_original_sector(wc, f) !=
1847                             read_original_sector(wc, e) + (wc->block_size >> SECTOR_SHIFT))
1848                                 break;
1849                         if (!wc_add_block(wb, f))
1850                                 break;
1851                         wbl->size--;
1852                         list_del(&f->lru);
1853                         wb->wc_list[wb->wc_list_n++] = f;
1854                         e = f;
1855                 }
1856                 if (WC_MODE_FUA(wc))
1857                         bio->bi_opf |= REQ_FUA;
1858                 if (writecache_has_error(wc)) {
1859                         bio->bi_status = BLK_STS_IOERR;
1860                         bio_endio(bio);
1861                 } else if (unlikely(!bio_sectors(bio))) {
1862                         bio->bi_status = BLK_STS_OK;
1863                         bio_endio(bio);
1864                 } else {
1865                         submit_bio(bio);
1866                 }
1867
1868                 __writeback_throttle(wc, wbl);
1869         }
1870 }
1871
1872 static void __writecache_writeback_ssd(struct dm_writecache *wc, struct writeback_list *wbl)
1873 {
1874         struct wc_entry *e, *f;
1875         struct dm_io_region from, to;
1876         struct copy_struct *c;
1877
1878         while (wbl->size) {
1879                 unsigned n_sectors;
1880
1881                 wbl->size--;
1882                 e = container_of(wbl->list.prev, struct wc_entry, lru);
1883                 list_del(&e->lru);
1884
1885                 n_sectors = e->wc_list_contiguous << (wc->block_size_bits - SECTOR_SHIFT);
1886
1887                 from.bdev = wc->ssd_dev->bdev;
1888                 from.sector = cache_sector(wc, e);
1889                 from.count = n_sectors;
1890                 to.bdev = wc->dev->bdev;
1891                 to.sector = read_original_sector(wc, e);
1892                 to.count = n_sectors;
1893
1894                 c = mempool_alloc(&wc->copy_pool, GFP_NOIO);
1895                 c->wc = wc;
1896                 c->e = e;
1897                 c->n_entries = e->wc_list_contiguous;
1898
1899                 while ((n_sectors -= wc->block_size >> SECTOR_SHIFT)) {
1900                         wbl->size--;
1901                         f = container_of(wbl->list.prev, struct wc_entry, lru);
1902                         BUG_ON(f != e + 1);
1903                         list_del(&f->lru);
1904                         e = f;
1905                 }
1906
1907                 if (unlikely(to.sector + to.count > wc->data_device_sectors)) {
1908                         if (to.sector >= wc->data_device_sectors) {
1909                                 writecache_copy_endio(0, 0, c);
1910                                 continue;
1911                         }
1912                         from.count = to.count = wc->data_device_sectors - to.sector;
1913                 }
1914
1915                 dm_kcopyd_copy(wc->dm_kcopyd, &from, 1, &to, 0, writecache_copy_endio, c);
1916
1917                 __writeback_throttle(wc, wbl);
1918         }
1919 }
1920
1921 static void writecache_writeback(struct work_struct *work)
1922 {
1923         struct dm_writecache *wc = container_of(work, struct dm_writecache, writeback_work);
1924         struct blk_plug plug;
1925         struct wc_entry *f, *g, *e = NULL;
1926         struct rb_node *node, *next_node;
1927         struct list_head skipped;
1928         struct writeback_list wbl;
1929         unsigned long n_walked;
1930
1931         if (!WC_MODE_PMEM(wc)) {
1932                 /* Wait for any active kcopyd work on behalf of ssd writeback */
1933                 dm_kcopyd_client_flush(wc->dm_kcopyd);
1934         }
1935
1936         if (likely(wc->pause != 0)) {
1937                 while (1) {
1938                         unsigned long idle;
1939                         if (unlikely(wc->cleaner) || unlikely(wc->writeback_all) ||
1940                             unlikely(dm_suspended(wc->ti)))
1941                                 break;
1942                         idle = dm_iot_idle_time(&wc->iot);
1943                         if (idle >= wc->pause)
1944                                 break;
1945                         idle = wc->pause - idle;
1946                         if (idle > HZ)
1947                                 idle = HZ;
1948                         schedule_timeout_idle(idle);
1949                 }
1950         }
1951
1952         wc_lock(wc);
1953 restart:
1954         if (writecache_has_error(wc)) {
1955                 wc_unlock(wc);
1956                 return;
1957         }
1958
1959         if (unlikely(wc->writeback_all)) {
1960                 if (writecache_wait_for_writeback(wc))
1961                         goto restart;
1962         }
1963
1964         if (wc->overwrote_committed) {
1965                 writecache_wait_for_ios(wc, WRITE);
1966         }
1967
1968         n_walked = 0;
1969         INIT_LIST_HEAD(&skipped);
1970         INIT_LIST_HEAD(&wbl.list);
1971         wbl.size = 0;
1972         while (!list_empty(&wc->lru) &&
1973                (wc->writeback_all ||
1974                 wc->freelist_size + wc->writeback_size <= wc->freelist_low_watermark ||
1975                 (jiffies - container_of(wc->lru.prev, struct wc_entry, lru)->age >=
1976                  wc->max_age - wc->max_age / MAX_AGE_DIV))) {
1977
1978                 n_walked++;
1979                 if (unlikely(n_walked > WRITEBACK_LATENCY) &&
1980                     likely(!wc->writeback_all)) {
1981                         if (likely(!dm_suspended(wc->ti)))
1982                                 queue_work(wc->writeback_wq, &wc->writeback_work);
1983                         break;
1984                 }
1985
1986                 if (unlikely(wc->writeback_all)) {
1987                         if (unlikely(!e)) {
1988                                 writecache_flush(wc);
1989                                 e = container_of(rb_first(&wc->tree), struct wc_entry, rb_node);
1990                         } else
1991                                 e = g;
1992                 } else
1993                         e = container_of(wc->lru.prev, struct wc_entry, lru);
1994                 BUG_ON(e->write_in_progress);
1995                 if (unlikely(!writecache_entry_is_committed(wc, e))) {
1996                         writecache_flush(wc);
1997                 }
1998                 node = rb_prev(&e->rb_node);
1999                 if (node) {
2000                         f = container_of(node, struct wc_entry, rb_node);
2001                         if (unlikely(read_original_sector(wc, f) ==
2002                                      read_original_sector(wc, e))) {
2003                                 BUG_ON(!f->write_in_progress);
2004                                 list_move(&e->lru, &skipped);
2005                                 cond_resched();
2006                                 continue;
2007                         }
2008                 }
2009                 wc->writeback_size++;
2010                 list_move(&e->lru, &wbl.list);
2011                 wbl.size++;
2012                 e->write_in_progress = true;
2013                 e->wc_list_contiguous = 1;
2014
2015                 f = e;
2016
2017                 while (1) {
2018                         next_node = rb_next(&f->rb_node);
2019                         if (unlikely(!next_node))
2020                                 break;
2021                         g = container_of(next_node, struct wc_entry, rb_node);
2022                         if (unlikely(read_original_sector(wc, g) ==
2023                             read_original_sector(wc, f))) {
2024                                 f = g;
2025                                 continue;
2026                         }
2027                         if (read_original_sector(wc, g) !=
2028                             read_original_sector(wc, f) + (wc->block_size >> SECTOR_SHIFT))
2029                                 break;
2030                         if (unlikely(g->write_in_progress))
2031                                 break;
2032                         if (unlikely(!writecache_entry_is_committed(wc, g)))
2033                                 break;
2034
2035                         if (!WC_MODE_PMEM(wc)) {
2036                                 if (g != f + 1)
2037                                         break;
2038                         }
2039
2040                         n_walked++;
2041                         //if (unlikely(n_walked > WRITEBACK_LATENCY) && likely(!wc->writeback_all))
2042                         //      break;
2043
2044                         wc->writeback_size++;
2045                         list_move(&g->lru, &wbl.list);
2046                         wbl.size++;
2047                         g->write_in_progress = true;
2048                         g->wc_list_contiguous = BIO_MAX_VECS;
2049                         f = g;
2050                         e->wc_list_contiguous++;
2051                         if (unlikely(e->wc_list_contiguous == BIO_MAX_VECS)) {
2052                                 if (unlikely(wc->writeback_all)) {
2053                                         next_node = rb_next(&f->rb_node);
2054                                         if (likely(next_node))
2055                                                 g = container_of(next_node, struct wc_entry, rb_node);
2056                                 }
2057                                 break;
2058                         }
2059                 }
2060                 cond_resched();
2061         }
2062
2063         if (!list_empty(&skipped)) {
2064                 list_splice_tail(&skipped, &wc->lru);
2065                 /*
2066                  * If we didn't do any progress, we must wait until some
2067                  * writeback finishes to avoid burning CPU in a loop
2068                  */
2069                 if (unlikely(!wbl.size))
2070                         writecache_wait_for_writeback(wc);
2071         }
2072
2073         wc_unlock(wc);
2074
2075         blk_start_plug(&plug);
2076
2077         if (WC_MODE_PMEM(wc))
2078                 __writecache_writeback_pmem(wc, &wbl);
2079         else
2080                 __writecache_writeback_ssd(wc, &wbl);
2081
2082         blk_finish_plug(&plug);
2083
2084         if (unlikely(wc->writeback_all)) {
2085                 wc_lock(wc);
2086                 while (writecache_wait_for_writeback(wc));
2087                 wc_unlock(wc);
2088         }
2089 }
2090
2091 static int calculate_memory_size(uint64_t device_size, unsigned block_size,
2092                                  size_t *n_blocks_p, size_t *n_metadata_blocks_p)
2093 {
2094         uint64_t n_blocks, offset;
2095         struct wc_entry e;
2096
2097         n_blocks = device_size;
2098         do_div(n_blocks, block_size + sizeof(struct wc_memory_entry));
2099
2100         while (1) {
2101                 if (!n_blocks)
2102                         return -ENOSPC;
2103                 /* Verify the following entries[n_blocks] won't overflow */
2104                 if (n_blocks >= ((size_t)-sizeof(struct wc_memory_superblock) /
2105                                  sizeof(struct wc_memory_entry)))
2106                         return -EFBIG;
2107                 offset = offsetof(struct wc_memory_superblock, entries[n_blocks]);
2108                 offset = (offset + block_size - 1) & ~(uint64_t)(block_size - 1);
2109                 if (offset + n_blocks * block_size <= device_size)
2110                         break;
2111                 n_blocks--;
2112         }
2113
2114         /* check if the bit field overflows */
2115         e.index = n_blocks;
2116         if (e.index != n_blocks)
2117                 return -EFBIG;
2118
2119         if (n_blocks_p)
2120                 *n_blocks_p = n_blocks;
2121         if (n_metadata_blocks_p)
2122                 *n_metadata_blocks_p = offset >> __ffs(block_size);
2123         return 0;
2124 }
2125
2126 static int init_memory(struct dm_writecache *wc)
2127 {
2128         size_t b;
2129         int r;
2130
2131         r = calculate_memory_size(wc->memory_map_size, wc->block_size, &wc->n_blocks, NULL);
2132         if (r)
2133                 return r;
2134
2135         r = writecache_alloc_entries(wc);
2136         if (r)
2137                 return r;
2138
2139         for (b = 0; b < ARRAY_SIZE(sb(wc)->padding); b++)
2140                 pmem_assign(sb(wc)->padding[b], cpu_to_le64(0));
2141         pmem_assign(sb(wc)->version, cpu_to_le32(MEMORY_SUPERBLOCK_VERSION));
2142         pmem_assign(sb(wc)->block_size, cpu_to_le32(wc->block_size));
2143         pmem_assign(sb(wc)->n_blocks, cpu_to_le64(wc->n_blocks));
2144         pmem_assign(sb(wc)->seq_count, cpu_to_le64(0));
2145
2146         for (b = 0; b < wc->n_blocks; b++) {
2147                 write_original_sector_seq_count(wc, &wc->entries[b], -1, -1);
2148                 cond_resched();
2149         }
2150
2151         writecache_flush_all_metadata(wc);
2152         writecache_commit_flushed(wc, false);
2153         pmem_assign(sb(wc)->magic, cpu_to_le32(MEMORY_SUPERBLOCK_MAGIC));
2154         writecache_flush_region(wc, &sb(wc)->magic, sizeof sb(wc)->magic);
2155         writecache_commit_flushed(wc, false);
2156
2157         return 0;
2158 }
2159
2160 static void writecache_dtr(struct dm_target *ti)
2161 {
2162         struct dm_writecache *wc = ti->private;
2163
2164         if (!wc)
2165                 return;
2166
2167         if (wc->endio_thread)
2168                 kthread_stop(wc->endio_thread);
2169
2170         if (wc->flush_thread)
2171                 kthread_stop(wc->flush_thread);
2172
2173         bioset_exit(&wc->bio_set);
2174
2175         mempool_exit(&wc->copy_pool);
2176
2177         if (wc->writeback_wq)
2178                 destroy_workqueue(wc->writeback_wq);
2179
2180         if (wc->dev)
2181                 dm_put_device(ti, wc->dev);
2182
2183         if (wc->ssd_dev)
2184                 dm_put_device(ti, wc->ssd_dev);
2185
2186         vfree(wc->entries);
2187
2188         if (wc->memory_map) {
2189                 if (WC_MODE_PMEM(wc))
2190                         persistent_memory_release(wc);
2191                 else
2192                         vfree(wc->memory_map);
2193         }
2194
2195         if (wc->dm_kcopyd)
2196                 dm_kcopyd_client_destroy(wc->dm_kcopyd);
2197
2198         if (wc->dm_io)
2199                 dm_io_client_destroy(wc->dm_io);
2200
2201         vfree(wc->dirty_bitmap);
2202
2203         kfree(wc);
2204 }
2205
2206 static int writecache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2207 {
2208         struct dm_writecache *wc;
2209         struct dm_arg_set as;
2210         const char *string;
2211         unsigned opt_params;
2212         size_t offset, data_size;
2213         int i, r;
2214         char dummy;
2215         int high_wm_percent = HIGH_WATERMARK;
2216         int low_wm_percent = LOW_WATERMARK;
2217         uint64_t x;
2218         struct wc_memory_superblock s;
2219
2220         static struct dm_arg _args[] = {
2221                 {0, 18, "Invalid number of feature args"},
2222         };
2223
2224         as.argc = argc;
2225         as.argv = argv;
2226
2227         wc = kzalloc(sizeof(struct dm_writecache), GFP_KERNEL);
2228         if (!wc) {
2229                 ti->error = "Cannot allocate writecache structure";
2230                 r = -ENOMEM;
2231                 goto bad;
2232         }
2233         ti->private = wc;
2234         wc->ti = ti;
2235
2236         mutex_init(&wc->lock);
2237         wc->max_age = MAX_AGE_UNSPECIFIED;
2238         writecache_poison_lists(wc);
2239         init_waitqueue_head(&wc->freelist_wait);
2240         timer_setup(&wc->autocommit_timer, writecache_autocommit_timer, 0);
2241         timer_setup(&wc->max_age_timer, writecache_max_age_timer, 0);
2242
2243         for (i = 0; i < 2; i++) {
2244                 atomic_set(&wc->bio_in_progress[i], 0);
2245                 init_waitqueue_head(&wc->bio_in_progress_wait[i]);
2246         }
2247
2248         wc->dm_io = dm_io_client_create();
2249         if (IS_ERR(wc->dm_io)) {
2250                 r = PTR_ERR(wc->dm_io);
2251                 ti->error = "Unable to allocate dm-io client";
2252                 wc->dm_io = NULL;
2253                 goto bad;
2254         }
2255
2256         wc->writeback_wq = alloc_workqueue("writecache-writeback", WQ_MEM_RECLAIM, 1);
2257         if (!wc->writeback_wq) {
2258                 r = -ENOMEM;
2259                 ti->error = "Could not allocate writeback workqueue";
2260                 goto bad;
2261         }
2262         INIT_WORK(&wc->writeback_work, writecache_writeback);
2263         INIT_WORK(&wc->flush_work, writecache_flush_work);
2264
2265         dm_iot_init(&wc->iot);
2266
2267         raw_spin_lock_init(&wc->endio_list_lock);
2268         INIT_LIST_HEAD(&wc->endio_list);
2269         wc->endio_thread = kthread_run(writecache_endio_thread, wc, "writecache_endio");
2270         if (IS_ERR(wc->endio_thread)) {
2271                 r = PTR_ERR(wc->endio_thread);
2272                 wc->endio_thread = NULL;
2273                 ti->error = "Couldn't spawn endio thread";
2274                 goto bad;
2275         }
2276
2277         /*
2278          * Parse the mode (pmem or ssd)
2279          */
2280         string = dm_shift_arg(&as);
2281         if (!string)
2282                 goto bad_arguments;
2283
2284         if (!strcasecmp(string, "s")) {
2285                 wc->pmem_mode = false;
2286         } else if (!strcasecmp(string, "p")) {
2287 #ifdef DM_WRITECACHE_HAS_PMEM
2288                 wc->pmem_mode = true;
2289                 wc->writeback_fua = true;
2290 #else
2291                 /*
2292                  * If the architecture doesn't support persistent memory or
2293                  * the kernel doesn't support any DAX drivers, this driver can
2294                  * only be used in SSD-only mode.
2295                  */
2296                 r = -EOPNOTSUPP;
2297                 ti->error = "Persistent memory or DAX not supported on this system";
2298                 goto bad;
2299 #endif
2300         } else {
2301                 goto bad_arguments;
2302         }
2303
2304         if (WC_MODE_PMEM(wc)) {
2305                 r = bioset_init(&wc->bio_set, BIO_POOL_SIZE,
2306                                 offsetof(struct writeback_struct, bio),
2307                                 BIOSET_NEED_BVECS);
2308                 if (r) {
2309                         ti->error = "Could not allocate bio set";
2310                         goto bad;
2311                 }
2312         } else {
2313                 wc->pause = PAUSE_WRITEBACK;
2314                 r = mempool_init_kmalloc_pool(&wc->copy_pool, 1, sizeof(struct copy_struct));
2315                 if (r) {
2316                         ti->error = "Could not allocate mempool";
2317                         goto bad;
2318                 }
2319         }
2320
2321         /*
2322          * Parse the origin data device
2323          */
2324         string = dm_shift_arg(&as);
2325         if (!string)
2326                 goto bad_arguments;
2327         r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->dev);
2328         if (r) {
2329                 ti->error = "Origin data device lookup failed";
2330                 goto bad;
2331         }
2332
2333         /*
2334          * Parse cache data device (be it pmem or ssd)
2335          */
2336         string = dm_shift_arg(&as);
2337         if (!string)
2338                 goto bad_arguments;
2339
2340         r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->ssd_dev);
2341         if (r) {
2342                 ti->error = "Cache data device lookup failed";
2343                 goto bad;
2344         }
2345         wc->memory_map_size = bdev_nr_bytes(wc->ssd_dev->bdev);
2346
2347         /*
2348          * Parse the cache block size
2349          */
2350         string = dm_shift_arg(&as);
2351         if (!string)
2352                 goto bad_arguments;
2353         if (sscanf(string, "%u%c", &wc->block_size, &dummy) != 1 ||
2354             wc->block_size < 512 || wc->block_size > PAGE_SIZE ||
2355             (wc->block_size & (wc->block_size - 1))) {
2356                 r = -EINVAL;
2357                 ti->error = "Invalid block size";
2358                 goto bad;
2359         }
2360         if (wc->block_size < bdev_logical_block_size(wc->dev->bdev) ||
2361             wc->block_size < bdev_logical_block_size(wc->ssd_dev->bdev)) {
2362                 r = -EINVAL;
2363                 ti->error = "Block size is smaller than device logical block size";
2364                 goto bad;
2365         }
2366         wc->block_size_bits = __ffs(wc->block_size);
2367
2368         wc->max_writeback_jobs = MAX_WRITEBACK_JOBS;
2369         wc->autocommit_blocks = !WC_MODE_PMEM(wc) ? AUTOCOMMIT_BLOCKS_SSD : AUTOCOMMIT_BLOCKS_PMEM;
2370         wc->autocommit_jiffies = msecs_to_jiffies(AUTOCOMMIT_MSEC);
2371
2372         /*
2373          * Parse optional arguments
2374          */
2375         r = dm_read_arg_group(_args, &as, &opt_params, &ti->error);
2376         if (r)
2377                 goto bad;
2378
2379         while (opt_params) {
2380                 string = dm_shift_arg(&as), opt_params--;
2381                 if (!strcasecmp(string, "start_sector") && opt_params >= 1) {
2382                         unsigned long long start_sector;
2383                         string = dm_shift_arg(&as), opt_params--;
2384                         if (sscanf(string, "%llu%c", &start_sector, &dummy) != 1)
2385                                 goto invalid_optional;
2386                         wc->start_sector = start_sector;
2387                         wc->start_sector_set = true;
2388                         if (wc->start_sector != start_sector ||
2389                             wc->start_sector >= wc->memory_map_size >> SECTOR_SHIFT)
2390                                 goto invalid_optional;
2391                 } else if (!strcasecmp(string, "high_watermark") && opt_params >= 1) {
2392                         string = dm_shift_arg(&as), opt_params--;
2393                         if (sscanf(string, "%d%c", &high_wm_percent, &dummy) != 1)
2394                                 goto invalid_optional;
2395                         if (high_wm_percent < 0 || high_wm_percent > 100)
2396                                 goto invalid_optional;
2397                         wc->high_wm_percent_value = high_wm_percent;
2398                         wc->high_wm_percent_set = true;
2399                 } else if (!strcasecmp(string, "low_watermark") && opt_params >= 1) {
2400                         string = dm_shift_arg(&as), opt_params--;
2401                         if (sscanf(string, "%d%c", &low_wm_percent, &dummy) != 1)
2402                                 goto invalid_optional;
2403                         if (low_wm_percent < 0 || low_wm_percent > 100)
2404                                 goto invalid_optional;
2405                         wc->low_wm_percent_value = low_wm_percent;
2406                         wc->low_wm_percent_set = true;
2407                 } else if (!strcasecmp(string, "writeback_jobs") && opt_params >= 1) {
2408                         string = dm_shift_arg(&as), opt_params--;
2409                         if (sscanf(string, "%u%c", &wc->max_writeback_jobs, &dummy) != 1)
2410                                 goto invalid_optional;
2411                         wc->max_writeback_jobs_set = true;
2412                 } else if (!strcasecmp(string, "autocommit_blocks") && opt_params >= 1) {
2413                         string = dm_shift_arg(&as), opt_params--;
2414                         if (sscanf(string, "%u%c", &wc->autocommit_blocks, &dummy) != 1)
2415                                 goto invalid_optional;
2416                         wc->autocommit_blocks_set = true;
2417                 } else if (!strcasecmp(string, "autocommit_time") && opt_params >= 1) {
2418                         unsigned autocommit_msecs;
2419                         string = dm_shift_arg(&as), opt_params--;
2420                         if (sscanf(string, "%u%c", &autocommit_msecs, &dummy) != 1)
2421                                 goto invalid_optional;
2422                         if (autocommit_msecs > 3600000)
2423                                 goto invalid_optional;
2424                         wc->autocommit_jiffies = msecs_to_jiffies(autocommit_msecs);
2425                         wc->autocommit_time_value = autocommit_msecs;
2426                         wc->autocommit_time_set = true;
2427                 } else if (!strcasecmp(string, "max_age") && opt_params >= 1) {
2428                         unsigned max_age_msecs;
2429                         string = dm_shift_arg(&as), opt_params--;
2430                         if (sscanf(string, "%u%c", &max_age_msecs, &dummy) != 1)
2431                                 goto invalid_optional;
2432                         if (max_age_msecs > 86400000)
2433                                 goto invalid_optional;
2434                         wc->max_age = msecs_to_jiffies(max_age_msecs);
2435                         wc->max_age_set = true;
2436                         wc->max_age_value = max_age_msecs;
2437                 } else if (!strcasecmp(string, "cleaner")) {
2438                         wc->cleaner_set = true;
2439                         wc->cleaner = true;
2440                 } else if (!strcasecmp(string, "fua")) {
2441                         if (WC_MODE_PMEM(wc)) {
2442                                 wc->writeback_fua = true;
2443                                 wc->writeback_fua_set = true;
2444                         } else goto invalid_optional;
2445                 } else if (!strcasecmp(string, "nofua")) {
2446                         if (WC_MODE_PMEM(wc)) {
2447                                 wc->writeback_fua = false;
2448                                 wc->writeback_fua_set = true;
2449                         } else goto invalid_optional;
2450                 } else if (!strcasecmp(string, "metadata_only")) {
2451                         wc->metadata_only = true;
2452                 } else if (!strcasecmp(string, "pause_writeback") && opt_params >= 1) {
2453                         unsigned pause_msecs;
2454                         if (WC_MODE_PMEM(wc))
2455                                 goto invalid_optional;
2456                         string = dm_shift_arg(&as), opt_params--;
2457                         if (sscanf(string, "%u%c", &pause_msecs, &dummy) != 1)
2458                                 goto invalid_optional;
2459                         if (pause_msecs > 60000)
2460                                 goto invalid_optional;
2461                         wc->pause = msecs_to_jiffies(pause_msecs);
2462                         wc->pause_set = true;
2463                         wc->pause_value = pause_msecs;
2464                 } else {
2465 invalid_optional:
2466                         r = -EINVAL;
2467                         ti->error = "Invalid optional argument";
2468                         goto bad;
2469                 }
2470         }
2471
2472         if (high_wm_percent < low_wm_percent) {
2473                 r = -EINVAL;
2474                 ti->error = "High watermark must be greater than or equal to low watermark";
2475                 goto bad;
2476         }
2477
2478         if (WC_MODE_PMEM(wc)) {
2479                 if (!dax_synchronous(wc->ssd_dev->dax_dev)) {
2480                         r = -EOPNOTSUPP;
2481                         ti->error = "Asynchronous persistent memory not supported as pmem cache";
2482                         goto bad;
2483                 }
2484
2485                 r = persistent_memory_claim(wc);
2486                 if (r) {
2487                         ti->error = "Unable to map persistent memory for cache";
2488                         goto bad;
2489                 }
2490         } else {
2491                 size_t n_blocks, n_metadata_blocks;
2492                 uint64_t n_bitmap_bits;
2493
2494                 wc->memory_map_size -= (uint64_t)wc->start_sector << SECTOR_SHIFT;
2495
2496                 bio_list_init(&wc->flush_list);
2497                 wc->flush_thread = kthread_run(writecache_flush_thread, wc, "dm_writecache_flush");
2498                 if (IS_ERR(wc->flush_thread)) {
2499                         r = PTR_ERR(wc->flush_thread);
2500                         wc->flush_thread = NULL;
2501                         ti->error = "Couldn't spawn flush thread";
2502                         goto bad;
2503                 }
2504
2505                 r = calculate_memory_size(wc->memory_map_size, wc->block_size,
2506                                           &n_blocks, &n_metadata_blocks);
2507                 if (r) {
2508                         ti->error = "Invalid device size";
2509                         goto bad;
2510                 }
2511
2512                 n_bitmap_bits = (((uint64_t)n_metadata_blocks << wc->block_size_bits) +
2513                                  BITMAP_GRANULARITY - 1) / BITMAP_GRANULARITY;
2514                 /* this is limitation of test_bit functions */
2515                 if (n_bitmap_bits > 1U << 31) {
2516                         r = -EFBIG;
2517                         ti->error = "Invalid device size";
2518                         goto bad;
2519                 }
2520
2521                 wc->memory_map = vmalloc(n_metadata_blocks << wc->block_size_bits);
2522                 if (!wc->memory_map) {
2523                         r = -ENOMEM;
2524                         ti->error = "Unable to allocate memory for metadata";
2525                         goto bad;
2526                 }
2527
2528                 wc->dm_kcopyd = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2529                 if (IS_ERR(wc->dm_kcopyd)) {
2530                         r = PTR_ERR(wc->dm_kcopyd);
2531                         ti->error = "Unable to allocate dm-kcopyd client";
2532                         wc->dm_kcopyd = NULL;
2533                         goto bad;
2534                 }
2535
2536                 wc->metadata_sectors = n_metadata_blocks << (wc->block_size_bits - SECTOR_SHIFT);
2537                 wc->dirty_bitmap_size = (n_bitmap_bits + BITS_PER_LONG - 1) /
2538                         BITS_PER_LONG * sizeof(unsigned long);
2539                 wc->dirty_bitmap = vzalloc(wc->dirty_bitmap_size);
2540                 if (!wc->dirty_bitmap) {
2541                         r = -ENOMEM;
2542                         ti->error = "Unable to allocate dirty bitmap";
2543                         goto bad;
2544                 }
2545
2546                 r = writecache_read_metadata(wc, wc->block_size >> SECTOR_SHIFT);
2547                 if (r) {
2548                         ti->error = "Unable to read first block of metadata";
2549                         goto bad;
2550                 }
2551         }
2552
2553         r = copy_mc_to_kernel(&s, sb(wc), sizeof(struct wc_memory_superblock));
2554         if (r) {
2555                 ti->error = "Hardware memory error when reading superblock";
2556                 goto bad;
2557         }
2558         if (!le32_to_cpu(s.magic) && !le32_to_cpu(s.version)) {
2559                 r = init_memory(wc);
2560                 if (r) {
2561                         ti->error = "Unable to initialize device";
2562                         goto bad;
2563                 }
2564                 r = copy_mc_to_kernel(&s, sb(wc),
2565                                       sizeof(struct wc_memory_superblock));
2566                 if (r) {
2567                         ti->error = "Hardware memory error when reading superblock";
2568                         goto bad;
2569                 }
2570         }
2571
2572         if (le32_to_cpu(s.magic) != MEMORY_SUPERBLOCK_MAGIC) {
2573                 ti->error = "Invalid magic in the superblock";
2574                 r = -EINVAL;
2575                 goto bad;
2576         }
2577
2578         if (le32_to_cpu(s.version) != MEMORY_SUPERBLOCK_VERSION) {
2579                 ti->error = "Invalid version in the superblock";
2580                 r = -EINVAL;
2581                 goto bad;
2582         }
2583
2584         if (le32_to_cpu(s.block_size) != wc->block_size) {
2585                 ti->error = "Block size does not match superblock";
2586                 r = -EINVAL;
2587                 goto bad;
2588         }
2589
2590         wc->n_blocks = le64_to_cpu(s.n_blocks);
2591
2592         offset = wc->n_blocks * sizeof(struct wc_memory_entry);
2593         if (offset / sizeof(struct wc_memory_entry) != le64_to_cpu(sb(wc)->n_blocks)) {
2594 overflow:
2595                 ti->error = "Overflow in size calculation";
2596                 r = -EINVAL;
2597                 goto bad;
2598         }
2599         offset += sizeof(struct wc_memory_superblock);
2600         if (offset < sizeof(struct wc_memory_superblock))
2601                 goto overflow;
2602         offset = (offset + wc->block_size - 1) & ~(size_t)(wc->block_size - 1);
2603         data_size = wc->n_blocks * (size_t)wc->block_size;
2604         if (!offset || (data_size / wc->block_size != wc->n_blocks) ||
2605             (offset + data_size < offset))
2606                 goto overflow;
2607         if (offset + data_size > wc->memory_map_size) {
2608                 ti->error = "Memory area is too small";
2609                 r = -EINVAL;
2610                 goto bad;
2611         }
2612
2613         wc->metadata_sectors = offset >> SECTOR_SHIFT;
2614         wc->block_start = (char *)sb(wc) + offset;
2615
2616         x = (uint64_t)wc->n_blocks * (100 - high_wm_percent);
2617         x += 50;
2618         do_div(x, 100);
2619         wc->freelist_high_watermark = x;
2620         x = (uint64_t)wc->n_blocks * (100 - low_wm_percent);
2621         x += 50;
2622         do_div(x, 100);
2623         wc->freelist_low_watermark = x;
2624
2625         if (wc->cleaner)
2626                 activate_cleaner(wc);
2627
2628         r = writecache_alloc_entries(wc);
2629         if (r) {
2630                 ti->error = "Cannot allocate memory";
2631                 goto bad;
2632         }
2633
2634         ti->num_flush_bios = WC_MODE_PMEM(wc) ? 1 : 2;
2635         ti->flush_supported = true;
2636         ti->num_discard_bios = 1;
2637
2638         if (WC_MODE_PMEM(wc))
2639                 persistent_memory_flush_cache(wc->memory_map, wc->memory_map_size);
2640
2641         return 0;
2642
2643 bad_arguments:
2644         r = -EINVAL;
2645         ti->error = "Bad arguments";
2646 bad:
2647         writecache_dtr(ti);
2648         return r;
2649 }
2650
2651 static void writecache_status(struct dm_target *ti, status_type_t type,
2652                               unsigned status_flags, char *result, unsigned maxlen)
2653 {
2654         struct dm_writecache *wc = ti->private;
2655         unsigned extra_args;
2656         unsigned sz = 0;
2657
2658         switch (type) {
2659         case STATUSTYPE_INFO:
2660                 DMEMIT("%ld %llu %llu %llu %llu %llu %llu %llu %llu %llu %llu %llu %llu %llu",
2661                        writecache_has_error(wc),
2662                        (unsigned long long)wc->n_blocks, (unsigned long long)wc->freelist_size,
2663                        (unsigned long long)wc->writeback_size,
2664                        wc->stats.reads,
2665                        wc->stats.read_hits,
2666                        wc->stats.writes,
2667                        wc->stats.write_hits_uncommitted,
2668                        wc->stats.write_hits_committed,
2669                        wc->stats.writes_around,
2670                        wc->stats.writes_allocate,
2671                        wc->stats.writes_blocked_on_freelist,
2672                        wc->stats.flushes,
2673                        wc->stats.discards);
2674                 break;
2675         case STATUSTYPE_TABLE:
2676                 DMEMIT("%c %s %s %u ", WC_MODE_PMEM(wc) ? 'p' : 's',
2677                                 wc->dev->name, wc->ssd_dev->name, wc->block_size);
2678                 extra_args = 0;
2679                 if (wc->start_sector_set)
2680                         extra_args += 2;
2681                 if (wc->high_wm_percent_set)
2682                         extra_args += 2;
2683                 if (wc->low_wm_percent_set)
2684                         extra_args += 2;
2685                 if (wc->max_writeback_jobs_set)
2686                         extra_args += 2;
2687                 if (wc->autocommit_blocks_set)
2688                         extra_args += 2;
2689                 if (wc->autocommit_time_set)
2690                         extra_args += 2;
2691                 if (wc->max_age_set)
2692                         extra_args += 2;
2693                 if (wc->cleaner_set)
2694                         extra_args++;
2695                 if (wc->writeback_fua_set)
2696                         extra_args++;
2697                 if (wc->metadata_only)
2698                         extra_args++;
2699                 if (wc->pause_set)
2700                         extra_args += 2;
2701
2702                 DMEMIT("%u", extra_args);
2703                 if (wc->start_sector_set)
2704                         DMEMIT(" start_sector %llu", (unsigned long long)wc->start_sector);
2705                 if (wc->high_wm_percent_set)
2706                         DMEMIT(" high_watermark %u", wc->high_wm_percent_value);
2707                 if (wc->low_wm_percent_set)
2708                         DMEMIT(" low_watermark %u", wc->low_wm_percent_value);
2709                 if (wc->max_writeback_jobs_set)
2710                         DMEMIT(" writeback_jobs %u", wc->max_writeback_jobs);
2711                 if (wc->autocommit_blocks_set)
2712                         DMEMIT(" autocommit_blocks %u", wc->autocommit_blocks);
2713                 if (wc->autocommit_time_set)
2714                         DMEMIT(" autocommit_time %u", wc->autocommit_time_value);
2715                 if (wc->max_age_set)
2716                         DMEMIT(" max_age %u", wc->max_age_value);
2717                 if (wc->cleaner_set)
2718                         DMEMIT(" cleaner");
2719                 if (wc->writeback_fua_set)
2720                         DMEMIT(" %sfua", wc->writeback_fua ? "" : "no");
2721                 if (wc->metadata_only)
2722                         DMEMIT(" metadata_only");
2723                 if (wc->pause_set)
2724                         DMEMIT(" pause_writeback %u", wc->pause_value);
2725                 break;
2726         case STATUSTYPE_IMA:
2727                 *result = '\0';
2728                 break;
2729         }
2730 }
2731
2732 static struct target_type writecache_target = {
2733         .name                   = "writecache",
2734         .version                = {1, 6, 0},
2735         .module                 = THIS_MODULE,
2736         .ctr                    = writecache_ctr,
2737         .dtr                    = writecache_dtr,
2738         .status                 = writecache_status,
2739         .postsuspend            = writecache_suspend,
2740         .resume                 = writecache_resume,
2741         .message                = writecache_message,
2742         .map                    = writecache_map,
2743         .end_io                 = writecache_end_io,
2744         .iterate_devices        = writecache_iterate_devices,
2745         .io_hints               = writecache_io_hints,
2746 };
2747
2748 static int __init dm_writecache_init(void)
2749 {
2750         int r;
2751
2752         r = dm_register_target(&writecache_target);
2753         if (r < 0) {
2754                 DMERR("register failed %d", r);
2755                 return r;
2756         }
2757
2758         return 0;
2759 }
2760
2761 static void __exit dm_writecache_exit(void)
2762 {
2763         dm_unregister_target(&writecache_target);
2764 }
2765
2766 module_init(dm_writecache_init);
2767 module_exit(dm_writecache_exit);
2768
2769 MODULE_DESCRIPTION(DM_NAME " writecache target");
2770 MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
2771 MODULE_LICENSE("GPL");