drivers/md/dm-writecache.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2018 Red Hat. All rights reserved.
   4  *
   5  * This file is released under the GPL.
   6  */
   7
   8 #include <linux/device-mapper.h>
   9 #include <linux/module.h>
  10 #include <linux/init.h>
  11 #include <linux/vmalloc.h>
  12 #include <linux/kthread.h>
  13 #include <linux/dm-io.h>
  14 #include <linux/dm-kcopyd.h>
  15 #include <linux/dax.h>
  16 #include <linux/pfn_t.h>
  17 #include <linux/libnvdimm.h>
  18
  19 #define DM_MSG_PREFIX "writecache"
  20
  21 #define HIGH_WATERMARK                  50
  22 #define LOW_WATERMARK                   45
  23 #define MAX_WRITEBACK_JOBS              0
  24 #define ENDIO_LATENCY                   16
  25 #define WRITEBACK_LATENCY               64
  26 #define AUTOCOMMIT_BLOCKS_SSD           65536
  27 #define AUTOCOMMIT_BLOCKS_PMEM          64
  28 #define AUTOCOMMIT_MSEC                 1000
  29
  30 #define BITMAP_GRANULARITY      65536
  31 #if BITMAP_GRANULARITY < PAGE_SIZE
  32 #undef BITMAP_GRANULARITY
  33 #define BITMAP_GRANULARITY      PAGE_SIZE
  34 #endif
  35
  36 #if IS_ENABLED(CONFIG_ARCH_HAS_PMEM_API) && IS_ENABLED(CONFIG_DAX_DRIVER)
  37 #define DM_WRITECACHE_HAS_PMEM
  38 #endif
  39
  40 #ifdef DM_WRITECACHE_HAS_PMEM
  41 #define pmem_assign(dest, src)                                  \
  42 do {                                                            \
  43         typeof(dest) uniq = (src);                              \
  44         memcpy_flushcache(&(dest), &uniq, sizeof(dest));        \
  45 } while (0)
  46 #else
  47 #define pmem_assign(dest, src)  ((dest) = (src))
  48 #endif
  49
  50 #if defined(__HAVE_ARCH_MEMCPY_MCSAFE) && defined(DM_WRITECACHE_HAS_PMEM)
  51 #define DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
  52 #endif
  53
  54 #define MEMORY_SUPERBLOCK_MAGIC         0x23489321
  55 #define MEMORY_SUPERBLOCK_VERSION       1
  56
  57 struct wc_memory_entry {
  58         __le64 original_sector;
  59         __le64 seq_count;
  60 };
  61
  62 struct wc_memory_superblock {
  63         union {
  64                 struct {
  65                         __le32 magic;
  66                         __le32 version;
  67                         __le32 block_size;
  68                         __le32 pad;
  69                         __le64 n_blocks;
  70                         __le64 seq_count;
  71                 };
  72                 __le64 padding[8];
  73         };
  74         struct wc_memory_entry entries[0];
  75 };
  76
  77 struct wc_entry {
  78         struct rb_node rb_node;
  79         struct list_head lru;
  80         unsigned short wc_list_contiguous;
  81         bool write_in_progress
  82 #if BITS_PER_LONG == 64
  83                 :1
  84 #endif
  85         ;
  86         unsigned long index
  87 #if BITS_PER_LONG == 64
  88                 :47
  89 #endif
  90         ;
  91 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
  92         uint64_t original_sector;
  93         uint64_t seq_count;
  94 #endif
  95 };
  96
  97 #ifdef DM_WRITECACHE_HAS_PMEM
  98 #define WC_MODE_PMEM(wc)                        ((wc)->pmem_mode)
  99 #define WC_MODE_FUA(wc)                         ((wc)->writeback_fua)
 100 #else
 101 #define WC_MODE_PMEM(wc)                        false
 102 #define WC_MODE_FUA(wc)                         false
 103 #endif
 104 #define WC_MODE_SORT_FREELIST(wc)               (!WC_MODE_PMEM(wc))
 105
 106 struct dm_writecache {
 107         struct mutex lock;
 108         struct list_head lru;
 109         union {
 110                 struct list_head freelist;
 111                 struct {
 112                         struct rb_root freetree;
 113                         struct wc_entry *current_free;
 114                 };
 115         };
 116         struct rb_root tree;
 117
 118         size_t freelist_size;
 119         size_t writeback_size;
 120         size_t freelist_high_watermark;
 121         size_t freelist_low_watermark;
 122
 123         unsigned uncommitted_blocks;
 124         unsigned autocommit_blocks;
 125         unsigned max_writeback_jobs;
 126
 127         int error;
 128
 129         unsigned long autocommit_jiffies;
 130         struct timer_list autocommit_timer;
 131         struct wait_queue_head freelist_wait;
 132
 133         atomic_t bio_in_progress[2];
 134         struct wait_queue_head bio_in_progress_wait[2];
 135
 136         struct dm_target *ti;
 137         struct dm_dev *dev;
 138         struct dm_dev *ssd_dev;
 139         sector_t start_sector;
 140         void *memory_map;
 141         uint64_t memory_map_size;
 142         size_t metadata_sectors;
 143         size_t n_blocks;
 144         uint64_t seq_count;
 145         void *block_start;
 146         struct wc_entry *entries;
 147         unsigned block_size;
 148         unsigned char block_size_bits;
 149
 150         bool pmem_mode:1;
 151         bool writeback_fua:1;
 152
 153         bool overwrote_committed:1;
 154         bool memory_vmapped:1;
 155
 156         bool high_wm_percent_set:1;
 157         bool low_wm_percent_set:1;
 158         bool max_writeback_jobs_set:1;
 159         bool autocommit_blocks_set:1;
 160         bool autocommit_time_set:1;
 161         bool writeback_fua_set:1;
 162         bool flush_on_suspend:1;
 163
 164         unsigned writeback_all;
 165         struct workqueue_struct *writeback_wq;
 166         struct work_struct writeback_work;
 167         struct work_struct flush_work;
 168
 169         struct dm_io_client *dm_io;
 170
 171         raw_spinlock_t endio_list_lock;
 172         struct list_head endio_list;
 173         struct task_struct *endio_thread;
 174
 175         struct task_struct *flush_thread;
 176         struct bio_list flush_list;
 177
 178         struct dm_kcopyd_client *dm_kcopyd;
 179         unsigned long *dirty_bitmap;
 180         unsigned dirty_bitmap_size;
 181
 182         struct bio_set bio_set;
 183         mempool_t copy_pool;
 184 };
 185
 186 #define WB_LIST_INLINE          16
 187
 188 struct writeback_struct {
 189         struct list_head endio_entry;
 190         struct dm_writecache *wc;
 191         struct wc_entry **wc_list;
 192         unsigned wc_list_n;
 193         struct wc_entry *wc_list_inline[WB_LIST_INLINE];
 194         struct bio bio;
 195 };
 196
 197 struct copy_struct {
 198         struct list_head endio_entry;
 199         struct dm_writecache *wc;
 200         struct wc_entry *e;
 201         unsigned n_entries;
 202         int error;
 203 };
 204
 205 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(dm_writecache_throttle,
 206                                             "A percentage of time allocated for data copying");
 207
 208 static void wc_lock(struct dm_writecache *wc)
 209 {
 210         mutex_lock(&wc->lock);
 211 }
 212
 213 static void wc_unlock(struct dm_writecache *wc)
 214 {
 215         mutex_unlock(&wc->lock);
 216 }
 217
 218 #ifdef DM_WRITECACHE_HAS_PMEM
 219 static int persistent_memory_claim(struct dm_writecache *wc)
 220 {
 221         int r;
 222         loff_t s;
 223         long p, da;
 224         pfn_t pfn;
 225         int id;
 226         struct page **pages;
 227
 228         wc->memory_vmapped = false;
 229
 230         if (!wc->ssd_dev->dax_dev) {
 231                 r = -EOPNOTSUPP;
 232                 goto err1;
 233         }
 234         s = wc->memory_map_size;
 235         p = s >> PAGE_SHIFT;
 236         if (!p) {
 237                 r = -EINVAL;
 238                 goto err1;
 239         }
 240         if (p != s >> PAGE_SHIFT) {
 241                 r = -EOVERFLOW;
 242                 goto err1;
 243         }
 244
 245         id = dax_read_lock();
 246
 247         da = dax_direct_access(wc->ssd_dev->dax_dev, 0, p, &wc->memory_map, &pfn);
 248         if (da < 0) {
 249                 wc->memory_map = NULL;
 250                 r = da;
 251                 goto err2;
 252         }
 253         if (!pfn_t_has_page(pfn)) {
 254                 wc->memory_map = NULL;
 255                 r = -EOPNOTSUPP;
 256                 goto err2;
 257         }
 258         if (da != p) {
 259                 long i;
 260                 wc->memory_map = NULL;
 261                 pages = kvmalloc_array(p, sizeof(struct page *), GFP_KERNEL);
 262                 if (!pages) {
 263                         r = -ENOMEM;
 264                         goto err2;
 265                 }
 266                 i = 0;
 267                 do {
 268                         long daa;
 269                         daa = dax_direct_access(wc->ssd_dev->dax_dev, i, p - i,
 270                                                 NULL, &pfn);
 271                         if (daa <= 0) {
 272                                 r = daa ? daa : -EINVAL;
 273                                 goto err3;
 274                         }
 275                         if (!pfn_t_has_page(pfn)) {
 276                                 r = -EOPNOTSUPP;
 277                                 goto err3;
 278                         }
 279                         while (daa-- && i < p) {
 280                                 pages[i++] = pfn_t_to_page(pfn);
 281                                 pfn.val++;
 282                         }
 283                 } while (i < p);
 284                 wc->memory_map = vmap(pages, p, VM_MAP, PAGE_KERNEL);
 285                 if (!wc->memory_map) {
 286                         r = -ENOMEM;
 287                         goto err3;
 288                 }
 289                 kvfree(pages);
 290                 wc->memory_vmapped = true;
 291         }
 292
 293         dax_read_unlock(id);
 294
 295         wc->memory_map += (size_t)wc->start_sector << SECTOR_SHIFT;
 296         wc->memory_map_size -= (size_t)wc->start_sector << SECTOR_SHIFT;
 297
 298         return 0;
 299 err3:
 300         kvfree(pages);
 301 err2:
 302         dax_read_unlock(id);
 303 err1:
 304         return r;
 305 }
 306 #else
 307 static int persistent_memory_claim(struct dm_writecache *wc)
 308 {
 309         BUG();
 310 }
 311 #endif
 312
 313 static void persistent_memory_release(struct dm_writecache *wc)
 314 {
 315         if (wc->memory_vmapped)
 316                 vunmap(wc->memory_map - ((size_t)wc->start_sector << SECTOR_SHIFT));
 317 }
 318
 319 static struct page *persistent_memory_page(void *addr)
 320 {
 321         if (is_vmalloc_addr(addr))
 322                 return vmalloc_to_page(addr);
 323         else
 324                 return virt_to_page(addr);
 325 }
 326
 327 static unsigned persistent_memory_page_offset(void *addr)
 328 {
 329         return (unsigned long)addr & (PAGE_SIZE - 1);
 330 }
 331
 332 static void persistent_memory_flush_cache(void *ptr, size_t size)
 333 {
 334         if (is_vmalloc_addr(ptr))
 335                 flush_kernel_vmap_range(ptr, size);
 336 }
 337
 338 static void persistent_memory_invalidate_cache(void *ptr, size_t size)
 339 {
 340         if (is_vmalloc_addr(ptr))
 341                 invalidate_kernel_vmap_range(ptr, size);
 342 }
 343
 344 static struct wc_memory_superblock *sb(struct dm_writecache *wc)
 345 {
 346         return wc->memory_map;
 347 }
 348
 349 static struct wc_memory_entry *memory_entry(struct dm_writecache *wc, struct wc_entry *e)
 350 {
 351         return &sb(wc)->entries[e->index];
 352 }
 353
 354 static void *memory_data(struct dm_writecache *wc, struct wc_entry *e)
 355 {
 356         return (char *)wc->block_start + (e->index << wc->block_size_bits);
 357 }
 358
 359 static sector_t cache_sector(struct dm_writecache *wc, struct wc_entry *e)
 360 {
 361         return wc->start_sector + wc->metadata_sectors +
 362                 ((sector_t)e->index << (wc->block_size_bits - SECTOR_SHIFT));
 363 }
 364
 365 static uint64_t read_original_sector(struct dm_writecache *wc, struct wc_entry *e)
 366 {
 367 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 368         return e->original_sector;
 369 #else
 370         return le64_to_cpu(memory_entry(wc, e)->original_sector);
 371 #endif
 372 }
 373
 374 static uint64_t read_seq_count(struct dm_writecache *wc, struct wc_entry *e)
 375 {
 376 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 377         return e->seq_count;
 378 #else
 379         return le64_to_cpu(memory_entry(wc, e)->seq_count);
 380 #endif
 381 }
 382
 383 static void clear_seq_count(struct dm_writecache *wc, struct wc_entry *e)
 384 {
 385 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 386         e->seq_count = -1;
 387 #endif
 388         pmem_assign(memory_entry(wc, e)->seq_count, cpu_to_le64(-1));
 389 }
 390
 391 static void write_original_sector_seq_count(struct dm_writecache *wc, struct wc_entry *e,
 392                                             uint64_t original_sector, uint64_t seq_count)
 393 {
 394         struct wc_memory_entry me;
 395 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 396         e->original_sector = original_sector;
 397         e->seq_count = seq_count;
 398 #endif
 399         me.original_sector = cpu_to_le64(original_sector);
 400         me.seq_count = cpu_to_le64(seq_count);
 401         pmem_assign(*memory_entry(wc, e), me);
 402 }
 403
 404 #define writecache_error(wc, err, msg, arg...)                          \
 405 do {                                                                    \
 406         if (!cmpxchg(&(wc)->error, 0, err))                             \
 407                 DMERR(msg, ##arg);                                      \
 408         wake_up(&(wc)->freelist_wait);                                  \
 409 } while (0)
 410
 411 #define writecache_has_error(wc)        (unlikely(READ_ONCE((wc)->error)))
 412
 413 static void writecache_flush_all_metadata(struct dm_writecache *wc)
 414 {
 415         if (!WC_MODE_PMEM(wc))
 416                 memset(wc->dirty_bitmap, -1, wc->dirty_bitmap_size);
 417 }
 418
 419 static void writecache_flush_region(struct dm_writecache *wc, void *ptr, size_t size)
 420 {
 421         if (!WC_MODE_PMEM(wc))
 422                 __set_bit(((char *)ptr - (char *)wc->memory_map) / BITMAP_GRANULARITY,
 423                           wc->dirty_bitmap);
 424 }
 425
 426 static void writecache_disk_flush(struct dm_writecache *wc, struct dm_dev *dev);
 427
 428 struct io_notify {
 429         struct dm_writecache *wc;
 430         struct completion c;
 431         atomic_t count;
 432 };
 433
 434 static void writecache_notify_io(unsigned long error, void *context)
 435 {
 436         struct io_notify *endio = context;
 437
 438         if (unlikely(error != 0))
 439                 writecache_error(endio->wc, -EIO, "error writing metadata");
 440         BUG_ON(atomic_read(&endio->count) <= 0);
 441         if (atomic_dec_and_test(&endio->count))
 442                 complete(&endio->c);
 443 }
 444
 445 static void writecache_wait_for_ios(struct dm_writecache *wc, int direction)
 446 {
 447         wait_event(wc->bio_in_progress_wait[direction],
 448                    !atomic_read(&wc->bio_in_progress[direction]));
 449 }
 450
 451 static void ssd_commit_flushed(struct dm_writecache *wc, bool wait_for_ios)
 452 {
 453         struct dm_io_region region;
 454         struct dm_io_request req;
 455         struct io_notify endio = {
 456                 wc,
 457                 COMPLETION_INITIALIZER_ONSTACK(endio.c),
 458                 ATOMIC_INIT(1),
 459         };
 460         unsigned bitmap_bits = wc->dirty_bitmap_size * 8;
 461         unsigned i = 0;
 462
 463         while (1) {
 464                 unsigned j;
 465                 i = find_next_bit(wc->dirty_bitmap, bitmap_bits, i);
 466                 if (unlikely(i == bitmap_bits))
 467                         break;
 468                 j = find_next_zero_bit(wc->dirty_bitmap, bitmap_bits, i);
 469
 470                 region.bdev = wc->ssd_dev->bdev;
 471                 region.sector = (sector_t)i * (BITMAP_GRANULARITY >> SECTOR_SHIFT);
 472                 region.count = (sector_t)(j - i) * (BITMAP_GRANULARITY >> SECTOR_SHIFT);
 473
 474                 if (unlikely(region.sector >= wc->metadata_sectors))
 475                         break;
 476                 if (unlikely(region.sector + region.count > wc->metadata_sectors))
 477                         region.count = wc->metadata_sectors - region.sector;
 478
 479                 region.sector += wc->start_sector;
 480                 atomic_inc(&endio.count);
 481                 req.bi_op = REQ_OP_WRITE;
 482                 req.bi_op_flags = REQ_SYNC;
 483                 req.mem.type = DM_IO_VMA;
 484                 req.mem.ptr.vma = (char *)wc->memory_map + (size_t)i * BITMAP_GRANULARITY;
 485                 req.client = wc->dm_io;
 486                 req.notify.fn = writecache_notify_io;
 487                 req.notify.context = &endio;
 488
 489                 /* writing via async dm-io (implied by notify.fn above) won't return an error */
 490                 (void) dm_io(&req, 1, &region, NULL);
 491                 i = j;
 492         }
 493
 494         writecache_notify_io(0, &endio);
 495         wait_for_completion_io(&endio.c);
 496
 497         if (wait_for_ios)
 498                 writecache_wait_for_ios(wc, WRITE);
 499
 500         writecache_disk_flush(wc, wc->ssd_dev);
 501
 502         memset(wc->dirty_bitmap, 0, wc->dirty_bitmap_size);
 503 }
 504
 505 static void writecache_commit_flushed(struct dm_writecache *wc, bool wait_for_ios)
 506 {
 507         if (WC_MODE_PMEM(wc))
 508                 wmb();
 509         else
 510                 ssd_commit_flushed(wc, wait_for_ios);
 511 }
 512
 513 static void writecache_disk_flush(struct dm_writecache *wc, struct dm_dev *dev)
 514 {
 515         int r;
 516         struct dm_io_region region;
 517         struct dm_io_request req;
 518
 519         region.bdev = dev->bdev;
 520         region.sector = 0;
 521         region.count = 0;
 522         req.bi_op = REQ_OP_WRITE;
 523         req.bi_op_flags = REQ_PREFLUSH;
 524         req.mem.type = DM_IO_KMEM;
 525         req.mem.ptr.addr = NULL;
 526         req.client = wc->dm_io;
 527         req.notify.fn = NULL;
 528
 529         r = dm_io(&req, 1, &region, NULL);
 530         if (unlikely(r))
 531                 writecache_error(wc, r, "error flushing metadata: %d", r);
 532 }
 533
 534 #define WFE_RETURN_FOLLOWING    1
 535 #define WFE_LOWEST_SEQ          2
 536
 537 static struct wc_entry *writecache_find_entry(struct dm_writecache *wc,
 538                                               uint64_t block, int flags)
 539 {
 540         struct wc_entry *e;
 541         struct rb_node *node = wc->tree.rb_node;
 542
 543         if (unlikely(!node))
 544                 return NULL;
 545
 546         while (1) {
 547                 e = container_of(node, struct wc_entry, rb_node);
 548                 if (read_original_sector(wc, e) == block)
 549                         break;
 550
 551                 node = (read_original_sector(wc, e) >= block ?
 552                         e->rb_node.rb_left : e->rb_node.rb_right);
 553                 if (unlikely(!node)) {
 554                         if (!(flags & WFE_RETURN_FOLLOWING))
 555                                 return NULL;
 556                         if (read_original_sector(wc, e) >= block) {
 557                                 return e;
 558                         } else {
 559                                 node = rb_next(&e->rb_node);
 560                                 if (unlikely(!node))
 561                                         return NULL;
 562                                 e = container_of(node, struct wc_entry, rb_node);
 563                                 return e;
 564                         }
 565                 }
 566         }
 567
 568         while (1) {
 569                 struct wc_entry *e2;
 570                 if (flags & WFE_LOWEST_SEQ)
 571                         node = rb_prev(&e->rb_node);
 572                 else
 573                         node = rb_next(&e->rb_node);
 574                 if (unlikely(!node))
 575                         return e;
 576                 e2 = container_of(node, struct wc_entry, rb_node);
 577                 if (read_original_sector(wc, e2) != block)
 578                         return e;
 579                 e = e2;
 580         }
 581 }
 582
 583 static void writecache_insert_entry(struct dm_writecache *wc, struct wc_entry *ins)
 584 {
 585         struct wc_entry *e;
 586         struct rb_node **node = &wc->tree.rb_node, *parent = NULL;
 587
 588         while (*node) {
 589                 e = container_of(*node, struct wc_entry, rb_node);
 590                 parent = &e->rb_node;
 591                 if (read_original_sector(wc, e) > read_original_sector(wc, ins))
 592                         node = &parent->rb_left;
 593                 else
 594                         node = &parent->rb_right;
 595         }
 596         rb_link_node(&ins->rb_node, parent, node);
 597         rb_insert_color(&ins->rb_node, &wc->tree);
 598         list_add(&ins->lru, &wc->lru);
 599 }
 600
 601 static void writecache_unlink(struct dm_writecache *wc, struct wc_entry *e)
 602 {
 603         list_del(&e->lru);
 604         rb_erase(&e->rb_node, &wc->tree);
 605 }
 606
 607 static void writecache_add_to_freelist(struct dm_writecache *wc, struct wc_entry *e)
 608 {
 609         if (WC_MODE_SORT_FREELIST(wc)) {
 610                 struct rb_node **node = &wc->freetree.rb_node, *parent = NULL;
 611                 if (unlikely(!*node))
 612                         wc->current_free = e;
 613                 while (*node) {
 614                         parent = *node;
 615                         if (&e->rb_node < *node)
 616                                 node = &parent->rb_left;
 617                         else
 618                                 node = &parent->rb_right;
 619                 }
 620                 rb_link_node(&e->rb_node, parent, node);
 621                 rb_insert_color(&e->rb_node, &wc->freetree);
 622         } else {
 623                 list_add_tail(&e->lru, &wc->freelist);
 624         }
 625         wc->freelist_size++;
 626 }
 627
 628 static struct wc_entry *writecache_pop_from_freelist(struct dm_writecache *wc, sector_t expected_sector)
 629 {
 630         struct wc_entry *e;
 631
 632         if (WC_MODE_SORT_FREELIST(wc)) {
 633                 struct rb_node *next;
 634                 if (unlikely(!wc->current_free))
 635                         return NULL;
 636                 e = wc->current_free;
 637                 if (expected_sector != (sector_t)-1 && unlikely(cache_sector(wc, e) != expected_sector))
 638                         return NULL;
 639                 next = rb_next(&e->rb_node);
 640                 rb_erase(&e->rb_node, &wc->freetree);
 641                 if (unlikely(!next))
 642                         next = rb_first(&wc->freetree);
 643                 wc->current_free = next ? container_of(next, struct wc_entry, rb_node) : NULL;
 644         } else {
 645                 if (unlikely(list_empty(&wc->freelist)))
 646                         return NULL;
 647                 e = container_of(wc->freelist.next, struct wc_entry, lru);
 648                 if (expected_sector != (sector_t)-1 && unlikely(cache_sector(wc, e) != expected_sector))
 649                         return NULL;
 650                 list_del(&e->lru);
 651         }
 652         wc->freelist_size--;
 653         if (unlikely(wc->freelist_size + wc->writeback_size <= wc->freelist_high_watermark))
 654                 queue_work(wc->writeback_wq, &wc->writeback_work);
 655
 656         return e;
 657 }
 658
 659 static void writecache_free_entry(struct dm_writecache *wc, struct wc_entry *e)
 660 {
 661         writecache_unlink(wc, e);
 662         writecache_add_to_freelist(wc, e);
 663         clear_seq_count(wc, e);
 664         writecache_flush_region(wc, memory_entry(wc, e), sizeof(struct wc_memory_entry));
 665         if (unlikely(waitqueue_active(&wc->freelist_wait)))
 666                 wake_up(&wc->freelist_wait);
 667 }
 668
 669 static void writecache_wait_on_freelist(struct dm_writecache *wc)
 670 {
 671         DEFINE_WAIT(wait);
 672
 673         prepare_to_wait(&wc->freelist_wait, &wait, TASK_UNINTERRUPTIBLE);
 674         wc_unlock(wc);
 675         io_schedule();
 676         finish_wait(&wc->freelist_wait, &wait);
 677         wc_lock(wc);
 678 }
 679
 680 static void writecache_poison_lists(struct dm_writecache *wc)
 681 {
 682         /*
 683          * Catch incorrect access to these values while the device is suspended.
 684          */
 685         memset(&wc->tree, -1, sizeof wc->tree);
 686         wc->lru.next = LIST_POISON1;
 687         wc->lru.prev = LIST_POISON2;
 688         wc->freelist.next = LIST_POISON1;
 689         wc->freelist.prev = LIST_POISON2;
 690 }
 691
 692 static void writecache_flush_entry(struct dm_writecache *wc, struct wc_entry *e)
 693 {
 694         writecache_flush_region(wc, memory_entry(wc, e), sizeof(struct wc_memory_entry));
 695         if (WC_MODE_PMEM(wc))
 696                 writecache_flush_region(wc, memory_data(wc, e), wc->block_size);
 697 }
 698
 699 static bool writecache_entry_is_committed(struct dm_writecache *wc, struct wc_entry *e)
 700 {
 701         return read_seq_count(wc, e) < wc->seq_count;
 702 }
 703
 704 static void writecache_flush(struct dm_writecache *wc)
 705 {
 706         struct wc_entry *e, *e2;
 707         bool need_flush_after_free;
 708
 709         wc->uncommitted_blocks = 0;
 710         del_timer(&wc->autocommit_timer);
 711
 712         if (list_empty(&wc->lru))
 713                 return;
 714
 715         e = container_of(wc->lru.next, struct wc_entry, lru);
 716         if (writecache_entry_is_committed(wc, e)) {
 717                 if (wc->overwrote_committed) {
 718                         writecache_wait_for_ios(wc, WRITE);
 719                         writecache_disk_flush(wc, wc->ssd_dev);
 720                         wc->overwrote_committed = false;
 721                 }
 722                 return;
 723         }
 724         while (1) {
 725                 writecache_flush_entry(wc, e);
 726                 if (unlikely(e->lru.next == &wc->lru))
 727                         break;
 728                 e2 = container_of(e->lru.next, struct wc_entry, lru);
 729                 if (writecache_entry_is_committed(wc, e2))
 730                         break;
 731                 e = e2;
 732                 cond_resched();
 733         }
 734         writecache_commit_flushed(wc, true);
 735
 736         wc->seq_count++;
 737         pmem_assign(sb(wc)->seq_count, cpu_to_le64(wc->seq_count));
 738         writecache_flush_region(wc, &sb(wc)->seq_count, sizeof sb(wc)->seq_count);
 739         writecache_commit_flushed(wc, false);
 740
 741         wc->overwrote_committed = false;
 742
 743         need_flush_after_free = false;
 744         while (1) {
 745                 /* Free another committed entry with lower seq-count */
 746                 struct rb_node *rb_node = rb_prev(&e->rb_node);
 747
 748                 if (rb_node) {
 749                         e2 = container_of(rb_node, struct wc_entry, rb_node);
 750                         if (read_original_sector(wc, e2) == read_original_sector(wc, e) &&
 751                             likely(!e2->write_in_progress)) {
 752                                 writecache_free_entry(wc, e2);
 753                                 need_flush_after_free = true;
 754                         }
 755                 }
 756                 if (unlikely(e->lru.prev == &wc->lru))
 757                         break;
 758                 e = container_of(e->lru.prev, struct wc_entry, lru);
 759                 cond_resched();
 760         }
 761
 762         if (need_flush_after_free)
 763                 writecache_commit_flushed(wc, false);
 764 }
 765
 766 static void writecache_flush_work(struct work_struct *work)
 767 {
 768         struct dm_writecache *wc = container_of(work, struct dm_writecache, flush_work);
 769
 770         wc_lock(wc);
 771         writecache_flush(wc);
 772         wc_unlock(wc);
 773 }
 774
 775 static void writecache_autocommit_timer(struct timer_list *t)
 776 {
 777         struct dm_writecache *wc = from_timer(wc, t, autocommit_timer);
 778         if (!writecache_has_error(wc))
 779                 queue_work(wc->writeback_wq, &wc->flush_work);
 780 }
 781
 782 static void writecache_schedule_autocommit(struct dm_writecache *wc)
 783 {
 784         if (!timer_pending(&wc->autocommit_timer))
 785                 mod_timer(&wc->autocommit_timer, jiffies + wc->autocommit_jiffies);
 786 }
 787
 788 static void writecache_discard(struct dm_writecache *wc, sector_t start, sector_t end)
 789 {
 790         struct wc_entry *e;
 791         bool discarded_something = false;
 792
 793         e = writecache_find_entry(wc, start, WFE_RETURN_FOLLOWING | WFE_LOWEST_SEQ);
 794         if (unlikely(!e))
 795                 return;
 796
 797         while (read_original_sector(wc, e) < end) {
 798                 struct rb_node *node = rb_next(&e->rb_node);
 799
 800                 if (likely(!e->write_in_progress)) {
 801                         if (!discarded_something) {
 802                                 writecache_wait_for_ios(wc, READ);
 803                                 writecache_wait_for_ios(wc, WRITE);
 804                                 discarded_something = true;
 805                         }
 806                         writecache_free_entry(wc, e);
 807                 }
 808
 809                 if (unlikely(!node))
 810                         break;
 811
 812                 e = container_of(node, struct wc_entry, rb_node);
 813         }
 814
 815         if (discarded_something)
 816                 writecache_commit_flushed(wc, false);
 817 }
 818
 819 static bool writecache_wait_for_writeback(struct dm_writecache *wc)
 820 {
 821         if (wc->writeback_size) {
 822                 writecache_wait_on_freelist(wc);
 823                 return true;
 824         }
 825         return false;
 826 }
 827
 828 static void writecache_suspend(struct dm_target *ti)
 829 {
 830         struct dm_writecache *wc = ti->private;
 831         bool flush_on_suspend;
 832
 833         del_timer_sync(&wc->autocommit_timer);
 834
 835         wc_lock(wc);
 836         writecache_flush(wc);
 837         flush_on_suspend = wc->flush_on_suspend;
 838         if (flush_on_suspend) {
 839                 wc->flush_on_suspend = false;
 840                 wc->writeback_all++;
 841                 queue_work(wc->writeback_wq, &wc->writeback_work);
 842         }
 843         wc_unlock(wc);
 844
 845         flush_workqueue(wc->writeback_wq);
 846
 847         wc_lock(wc);
 848         if (flush_on_suspend)
 849                 wc->writeback_all--;
 850         while (writecache_wait_for_writeback(wc));
 851
 852         if (WC_MODE_PMEM(wc))
 853                 persistent_memory_flush_cache(wc->memory_map, wc->memory_map_size);
 854
 855         writecache_poison_lists(wc);
 856
 857         wc_unlock(wc);
 858 }
 859
 860 static int writecache_alloc_entries(struct dm_writecache *wc)
 861 {
 862         size_t b;
 863
 864         if (wc->entries)
 865                 return 0;
 866         wc->entries = vmalloc(array_size(sizeof(struct wc_entry), wc->n_blocks));
 867         if (!wc->entries)
 868                 return -ENOMEM;
 869         for (b = 0; b < wc->n_blocks; b++) {
 870                 struct wc_entry *e = &wc->entries[b];
 871                 e->index = b;
 872                 e->write_in_progress = false;
 873         }
 874
 875         return 0;
 876 }
 877
 878 static void writecache_resume(struct dm_target *ti)
 879 {
 880         struct dm_writecache *wc = ti->private;
 881         size_t b;
 882         bool need_flush = false;
 883         __le64 sb_seq_count;
 884         int r;
 885
 886         wc_lock(wc);
 887
 888         if (WC_MODE_PMEM(wc))
 889                 persistent_memory_invalidate_cache(wc->memory_map, wc->memory_map_size);
 890
 891         wc->tree = RB_ROOT;
 892         INIT_LIST_HEAD(&wc->lru);
 893         if (WC_MODE_SORT_FREELIST(wc)) {
 894                 wc->freetree = RB_ROOT;
 895                 wc->current_free = NULL;
 896         } else {
 897                 INIT_LIST_HEAD(&wc->freelist);
 898         }
 899         wc->freelist_size = 0;
 900
 901         r = memcpy_mcsafe(&sb_seq_count, &sb(wc)->seq_count, sizeof(uint64_t));
 902         if (r) {
 903                 writecache_error(wc, r, "hardware memory error when reading superblock: %d", r);
 904                 sb_seq_count = cpu_to_le64(0);
 905         }
 906         wc->seq_count = le64_to_cpu(sb_seq_count);
 907
 908 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 909         for (b = 0; b < wc->n_blocks; b++) {
 910                 struct wc_entry *e = &wc->entries[b];
 911                 struct wc_memory_entry wme;
 912                 if (writecache_has_error(wc)) {
 913                         e->original_sector = -1;
 914                         e->seq_count = -1;
 915                         continue;
 916                 }
 917                 r = memcpy_mcsafe(&wme, memory_entry(wc, e), sizeof(struct wc_memory_entry));
 918                 if (r) {
 919                         writecache_error(wc, r, "hardware memory error when reading metadata entry %lu: %d",
 920                                          (unsigned long)b, r);
 921                         e->original_sector = -1;
 922                         e->seq_count = -1;
 923                 } else {
 924                         e->original_sector = le64_to_cpu(wme.original_sector);
 925                         e->seq_count = le64_to_cpu(wme.seq_count);
 926                 }
 927         }
 928 #endif
 929         for (b = 0; b < wc->n_blocks; b++) {
 930                 struct wc_entry *e = &wc->entries[b];
 931                 if (!writecache_entry_is_committed(wc, e)) {
 932                         if (read_seq_count(wc, e) != -1) {
 933 erase_this:
 934                                 clear_seq_count(wc, e);
 935                                 need_flush = true;
 936                         }
 937                         writecache_add_to_freelist(wc, e);
 938                 } else {
 939                         struct wc_entry *old;
 940
 941                         old = writecache_find_entry(wc, read_original_sector(wc, e), 0);
 942                         if (!old) {
 943                                 writecache_insert_entry(wc, e);
 944                         } else {
 945                                 if (read_seq_count(wc, old) == read_seq_count(wc, e)) {
 946                                         writecache_error(wc, -EINVAL,
 947                                                  "two identical entries, position %llu, sector %llu, sequence %llu",
 948                                                  (unsigned long long)b, (unsigned long long)read_original_sector(wc, e),
 949                                                  (unsigned long long)read_seq_count(wc, e));
 950                                 }
 951                                 if (read_seq_count(wc, old) > read_seq_count(wc, e)) {
 952                                         goto erase_this;
 953                                 } else {
 954                                         writecache_free_entry(wc, old);
 955                                         writecache_insert_entry(wc, e);
 956                                         need_flush = true;
 957                                 }
 958                         }
 959                 }
 960                 cond_resched();
 961         }
 962
 963         if (need_flush) {
 964                 writecache_flush_all_metadata(wc);
 965                 writecache_commit_flushed(wc, false);
 966         }
 967
 968         wc_unlock(wc);
 969 }
 970
 971 static int process_flush_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
 972 {
 973         if (argc != 1)
 974                 return -EINVAL;
 975
 976         wc_lock(wc);
 977         if (dm_suspended(wc->ti)) {
 978                 wc_unlock(wc);
 979                 return -EBUSY;
 980         }
 981         if (writecache_has_error(wc)) {
 982                 wc_unlock(wc);
 983                 return -EIO;
 984         }
 985
 986         writecache_flush(wc);
 987         wc->writeback_all++;
 988         queue_work(wc->writeback_wq, &wc->writeback_work);
 989         wc_unlock(wc);
 990
 991         flush_workqueue(wc->writeback_wq);
 992
 993         wc_lock(wc);
 994         wc->writeback_all--;
 995         if (writecache_has_error(wc)) {
 996                 wc_unlock(wc);
 997                 return -EIO;
 998         }
 999         wc_unlock(wc);
1000
1001         return 0;
1002 }
1003
1004 static int process_flush_on_suspend_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1005 {
1006         if (argc != 1)
1007                 return -EINVAL;
1008
1009         wc_lock(wc);
1010         wc->flush_on_suspend = true;
1011         wc_unlock(wc);
1012
1013         return 0;
1014 }
1015
1016 static int writecache_message(struct dm_target *ti, unsigned argc, char **argv,
1017                               char *result, unsigned maxlen)
1018 {
1019         int r = -EINVAL;
1020         struct dm_writecache *wc = ti->private;
1021
1022         if (!strcasecmp(argv[0], "flush"))
1023                 r = process_flush_mesg(argc, argv, wc);
1024         else if (!strcasecmp(argv[0], "flush_on_suspend"))
1025                 r = process_flush_on_suspend_mesg(argc, argv, wc);
1026         else
1027                 DMERR("unrecognised message received: %s", argv[0]);
1028
1029         return r;
1030 }
1031
1032 static void bio_copy_block(struct dm_writecache *wc, struct bio *bio, void *data)
1033 {
1034         void *buf;
1035         unsigned long flags;
1036         unsigned size;
1037         int rw = bio_data_dir(bio);
1038         unsigned remaining_size = wc->block_size;
1039
1040         do {
1041                 struct bio_vec bv = bio_iter_iovec(bio, bio->bi_iter);
1042                 buf = bvec_kmap_irq(&bv, &flags);
1043                 size = bv.bv_len;
1044                 if (unlikely(size > remaining_size))
1045                         size = remaining_size;
1046
1047                 if (rw == READ) {
1048                         int r;
1049                         r = memcpy_mcsafe(buf, data, size);
1050                         flush_dcache_page(bio_page(bio));
1051                         if (unlikely(r)) {
1052                                 writecache_error(wc, r, "hardware memory error when reading data: %d", r);
1053                                 bio->bi_status = BLK_STS_IOERR;
1054                         }
1055                 } else {
1056                         flush_dcache_page(bio_page(bio));
1057                         memcpy_flushcache(data, buf, size);
1058                 }
1059
1060                 bvec_kunmap_irq(buf, &flags);
1061
1062                 data = (char *)data + size;
1063                 remaining_size -= size;
1064                 bio_advance(bio, size);
1065         } while (unlikely(remaining_size));
1066 }
1067
1068 static int writecache_flush_thread(void *data)
1069 {
1070         struct dm_writecache *wc = data;
1071
1072         while (1) {
1073                 struct bio *bio;
1074
1075                 wc_lock(wc);
1076                 bio = bio_list_pop(&wc->flush_list);
1077                 if (!bio) {
1078                         set_current_state(TASK_INTERRUPTIBLE);
1079                         wc_unlock(wc);
1080
1081                         if (unlikely(kthread_should_stop())) {
1082                                 set_current_state(TASK_RUNNING);
1083                                 break;
1084                         }
1085
1086                         schedule();
1087                         continue;
1088                 }
1089
1090                 if (bio_op(bio) == REQ_OP_DISCARD) {
1091                         writecache_discard(wc, bio->bi_iter.bi_sector,
1092                                            bio_end_sector(bio));
1093                         wc_unlock(wc);
1094                         bio_set_dev(bio, wc->dev->bdev);
1095                         generic_make_request(bio);
1096                 } else {
1097                         writecache_flush(wc);
1098                         wc_unlock(wc);
1099                         if (writecache_has_error(wc))
1100                                 bio->bi_status = BLK_STS_IOERR;
1101                         bio_endio(bio);
1102                 }
1103         }
1104
1105         return 0;
1106 }
1107
1108 static void writecache_offload_bio(struct dm_writecache *wc, struct bio *bio)
1109 {
1110         if (bio_list_empty(&wc->flush_list))
1111                 wake_up_process(wc->flush_thread);
1112         bio_list_add(&wc->flush_list, bio);
1113 }
1114
1115 static int writecache_map(struct dm_target *ti, struct bio *bio)
1116 {
1117         struct wc_entry *e;
1118         struct dm_writecache *wc = ti->private;
1119
1120         bio->bi_private = NULL;
1121
1122         wc_lock(wc);
1123
1124         if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1125                 if (writecache_has_error(wc))
1126                         goto unlock_error;
1127                 if (WC_MODE_PMEM(wc)) {
1128                         writecache_flush(wc);
1129                         if (writecache_has_error(wc))
1130                                 goto unlock_error;
1131                         goto unlock_submit;
1132                 } else {
1133                         writecache_offload_bio(wc, bio);
1134                         goto unlock_return;
1135                 }
1136         }
1137
1138         bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
1139
1140         if (unlikely((((unsigned)bio->bi_iter.bi_sector | bio_sectors(bio)) &
1141                                 (wc->block_size / 512 - 1)) != 0)) {
1142                 DMERR("I/O is not aligned, sector %llu, size %u, block size %u",
1143                       (unsigned long long)bio->bi_iter.bi_sector,
1144                       bio->bi_iter.bi_size, wc->block_size);
1145                 goto unlock_error;
1146         }
1147
1148         if (unlikely(bio_op(bio) == REQ_OP_DISCARD)) {
1149                 if (writecache_has_error(wc))
1150                         goto unlock_error;
1151                 if (WC_MODE_PMEM(wc)) {
1152                         writecache_discard(wc, bio->bi_iter.bi_sector, bio_end_sector(bio));
1153                         goto unlock_remap_origin;
1154                 } else {
1155                         writecache_offload_bio(wc, bio);
1156                         goto unlock_return;
1157                 }
1158         }
1159
1160         if (bio_data_dir(bio) == READ) {
1161 read_next_block:
1162                 e = writecache_find_entry(wc, bio->bi_iter.bi_sector, WFE_RETURN_FOLLOWING);
1163                 if (e && read_original_sector(wc, e) == bio->bi_iter.bi_sector) {
1164                         if (WC_MODE_PMEM(wc)) {
1165                                 bio_copy_block(wc, bio, memory_data(wc, e));
1166                                 if (bio->bi_iter.bi_size)
1167                                         goto read_next_block;
1168                                 goto unlock_submit;
1169                         } else {
1170                                 dm_accept_partial_bio(bio, wc->block_size >> SECTOR_SHIFT);
1171                                 bio_set_dev(bio, wc->ssd_dev->bdev);
1172                                 bio->bi_iter.bi_sector = cache_sector(wc, e);
1173                                 if (!writecache_entry_is_committed(wc, e))
1174                                         writecache_wait_for_ios(wc, WRITE);
1175                                 goto unlock_remap;
1176                         }
1177                 } else {
1178                         if (e) {
1179                                 sector_t next_boundary =
1180                                         read_original_sector(wc, e) - bio->bi_iter.bi_sector;
1181                                 if (next_boundary < bio->bi_iter.bi_size >> SECTOR_SHIFT) {
1182                                         dm_accept_partial_bio(bio, next_boundary);
1183                                 }
1184                         }
1185                         goto unlock_remap_origin;
1186                 }
1187         } else {
1188                 do {
1189                         if (writecache_has_error(wc))
1190                                 goto unlock_error;
1191                         e = writecache_find_entry(wc, bio->bi_iter.bi_sector, 0);
1192                         if (e) {
1193                                 if (!writecache_entry_is_committed(wc, e))
1194                                         goto bio_copy;
1195                                 if (!WC_MODE_PMEM(wc) && !e->write_in_progress) {
1196                                         wc->overwrote_committed = true;
1197                                         goto bio_copy;
1198                                 }
1199                         }
1200                         e = writecache_pop_from_freelist(wc, (sector_t)-1);
1201                         if (unlikely(!e)) {
1202                                 writecache_wait_on_freelist(wc);
1203                                 continue;
1204                         }
1205                         write_original_sector_seq_count(wc, e, bio->bi_iter.bi_sector, wc->seq_count);
1206                         writecache_insert_entry(wc, e);
1207                         wc->uncommitted_blocks++;
1208 bio_copy:
1209                         if (WC_MODE_PMEM(wc)) {
1210                                 bio_copy_block(wc, bio, memory_data(wc, e));
1211                         } else {
1212                                 unsigned bio_size = wc->block_size;
1213                                 sector_t start_cache_sec = cache_sector(wc, e);
1214                                 sector_t current_cache_sec = start_cache_sec + (bio_size >> SECTOR_SHIFT);
1215
1216                                 while (bio_size < bio->bi_iter.bi_size) {
1217                                         struct wc_entry *f = writecache_pop_from_freelist(wc, current_cache_sec);
1218                                         if (!f)
1219                                                 break;
1220                                         write_original_sector_seq_count(wc, f, bio->bi_iter.bi_sector +
1221                                                                         (bio_size >> SECTOR_SHIFT), wc->seq_count);
1222                                         writecache_insert_entry(wc, f);
1223                                         wc->uncommitted_blocks++;
1224                                         bio_size += wc->block_size;
1225                                         current_cache_sec += wc->block_size >> SECTOR_SHIFT;
1226                                 }
1227
1228                                 bio_set_dev(bio, wc->ssd_dev->bdev);
1229                                 bio->bi_iter.bi_sector = start_cache_sec;
1230                                 dm_accept_partial_bio(bio, bio_size >> SECTOR_SHIFT);
1231
1232                                 if (unlikely(wc->uncommitted_blocks >= wc->autocommit_blocks)) {
1233                                         wc->uncommitted_blocks = 0;
1234                                         queue_work(wc->writeback_wq, &wc->flush_work);
1235                                 } else {
1236                                         writecache_schedule_autocommit(wc);
1237                                 }
1238                                 goto unlock_remap;
1239                         }
1240                 } while (bio->bi_iter.bi_size);
1241
1242                 if (unlikely(bio->bi_opf & REQ_FUA ||
1243                              wc->uncommitted_blocks >= wc->autocommit_blocks))
1244                         writecache_flush(wc);
1245                 else
1246                         writecache_schedule_autocommit(wc);
1247                 goto unlock_submit;
1248         }
1249
1250 unlock_remap_origin:
1251         bio_set_dev(bio, wc->dev->bdev);
1252         wc_unlock(wc);
1253         return DM_MAPIO_REMAPPED;
1254
1255 unlock_remap:
1256         /* make sure that writecache_end_io decrements bio_in_progress: */
1257         bio->bi_private = (void *)1;
1258         atomic_inc(&wc->bio_in_progress[bio_data_dir(bio)]);
1259         wc_unlock(wc);
1260         return DM_MAPIO_REMAPPED;
1261
1262 unlock_submit:
1263         wc_unlock(wc);
1264         bio_endio(bio);
1265         return DM_MAPIO_SUBMITTED;
1266
1267 unlock_return:
1268         wc_unlock(wc);
1269         return DM_MAPIO_SUBMITTED;
1270
1271 unlock_error:
1272         wc_unlock(wc);
1273         bio_io_error(bio);
1274         return DM_MAPIO_SUBMITTED;
1275 }
1276
1277 static int writecache_end_io(struct dm_target *ti, struct bio *bio, blk_status_t *status)
1278 {
1279         struct dm_writecache *wc = ti->private;
1280
1281         if (bio->bi_private != NULL) {
1282                 int dir = bio_data_dir(bio);
1283                 if (atomic_dec_and_test(&wc->bio_in_progress[dir]))
1284                         if (unlikely(waitqueue_active(&wc->bio_in_progress_wait[dir])))
1285                                 wake_up(&wc->bio_in_progress_wait[dir]);
1286         }
1287         return 0;
1288 }
1289
1290 static int writecache_iterate_devices(struct dm_target *ti,
1291                                       iterate_devices_callout_fn fn, void *data)
1292 {
1293         struct dm_writecache *wc = ti->private;
1294
1295         return fn(ti, wc->dev, 0, ti->len, data);
1296 }
1297
1298 static void writecache_io_hints(struct dm_target *ti, struct queue_limits *limits)
1299 {
1300         struct dm_writecache *wc = ti->private;
1301
1302         if (limits->logical_block_size < wc->block_size)
1303                 limits->logical_block_size = wc->block_size;
1304
1305         if (limits->physical_block_size < wc->block_size)
1306                 limits->physical_block_size = wc->block_size;
1307
1308         if (limits->io_min < wc->block_size)
1309                 limits->io_min = wc->block_size;
1310 }
1311
1312
1313 static void writecache_writeback_endio(struct bio *bio)
1314 {
1315         struct writeback_struct *wb = container_of(bio, struct writeback_struct, bio);
1316         struct dm_writecache *wc = wb->wc;
1317         unsigned long flags;
1318
1319         raw_spin_lock_irqsave(&wc->endio_list_lock, flags);
1320         if (unlikely(list_empty(&wc->endio_list)))
1321                 wake_up_process(wc->endio_thread);
1322         list_add_tail(&wb->endio_entry, &wc->endio_list);
1323         raw_spin_unlock_irqrestore(&wc->endio_list_lock, flags);
1324 }
1325
1326 static void writecache_copy_endio(int read_err, unsigned long write_err, void *ptr)
1327 {
1328         struct copy_struct *c = ptr;
1329         struct dm_writecache *wc = c->wc;
1330
1331         c->error = likely(!(read_err | write_err)) ? 0 : -EIO;
1332
1333         raw_spin_lock_irq(&wc->endio_list_lock);
1334         if (unlikely(list_empty(&wc->endio_list)))
1335                 wake_up_process(wc->endio_thread);
1336         list_add_tail(&c->endio_entry, &wc->endio_list);
1337         raw_spin_unlock_irq(&wc->endio_list_lock);
1338 }
1339
1340 static void __writecache_endio_pmem(struct dm_writecache *wc, struct list_head *list)
1341 {
1342         unsigned i;
1343         struct writeback_struct *wb;
1344         struct wc_entry *e;
1345         unsigned long n_walked = 0;
1346
1347         do {
1348                 wb = list_entry(list->next, struct writeback_struct, endio_entry);
1349                 list_del(&wb->endio_entry);
1350
1351                 if (unlikely(wb->bio.bi_status != BLK_STS_OK))
1352                         writecache_error(wc, blk_status_to_errno(wb->bio.bi_status),
1353                                         "write error %d", wb->bio.bi_status);
1354                 i = 0;
1355                 do {
1356                         e = wb->wc_list[i];
1357                         BUG_ON(!e->write_in_progress);
1358                         e->write_in_progress = false;
1359                         INIT_LIST_HEAD(&e->lru);
1360                         if (!writecache_has_error(wc))
1361                                 writecache_free_entry(wc, e);
1362                         BUG_ON(!wc->writeback_size);
1363                         wc->writeback_size--;
1364                         n_walked++;
1365                         if (unlikely(n_walked >= ENDIO_LATENCY)) {
1366                                 writecache_commit_flushed(wc, false);
1367                                 wc_unlock(wc);
1368                                 wc_lock(wc);
1369                                 n_walked = 0;
1370                         }
1371                 } while (++i < wb->wc_list_n);
1372
1373                 if (wb->wc_list != wb->wc_list_inline)
1374                         kfree(wb->wc_list);
1375                 bio_put(&wb->bio);
1376         } while (!list_empty(list));
1377 }
1378
1379 static void __writecache_endio_ssd(struct dm_writecache *wc, struct list_head *list)
1380 {
1381         struct copy_struct *c;
1382         struct wc_entry *e;
1383
1384         do {
1385                 c = list_entry(list->next, struct copy_struct, endio_entry);
1386                 list_del(&c->endio_entry);
1387
1388                 if (unlikely(c->error))
1389                         writecache_error(wc, c->error, "copy error");
1390
1391                 e = c->e;
1392                 do {
1393                         BUG_ON(!e->write_in_progress);
1394                         e->write_in_progress = false;
1395                         INIT_LIST_HEAD(&e->lru);
1396                         if (!writecache_has_error(wc))
1397                                 writecache_free_entry(wc, e);
1398
1399                         BUG_ON(!wc->writeback_size);
1400                         wc->writeback_size--;
1401                         e++;
1402                 } while (--c->n_entries);
1403                 mempool_free(c, &wc->copy_pool);
1404         } while (!list_empty(list));
1405 }
1406
1407 static int writecache_endio_thread(void *data)
1408 {
1409         struct dm_writecache *wc = data;
1410
1411         while (1) {
1412                 struct list_head list;
1413
1414                 raw_spin_lock_irq(&wc->endio_list_lock);
1415                 if (!list_empty(&wc->endio_list))
1416                         goto pop_from_list;
1417                 set_current_state(TASK_INTERRUPTIBLE);
1418                 raw_spin_unlock_irq(&wc->endio_list_lock);
1419
1420                 if (unlikely(kthread_should_stop())) {
1421                         set_current_state(TASK_RUNNING);
1422                         break;
1423                 }
1424
1425                 schedule();
1426
1427                 continue;
1428
1429 pop_from_list:
1430                 list = wc->endio_list;
1431                 list.next->prev = list.prev->next = &list;
1432                 INIT_LIST_HEAD(&wc->endio_list);
1433                 raw_spin_unlock_irq(&wc->endio_list_lock);
1434
1435                 if (!WC_MODE_FUA(wc))
1436                         writecache_disk_flush(wc, wc->dev);
1437
1438                 wc_lock(wc);
1439
1440                 if (WC_MODE_PMEM(wc)) {
1441                         __writecache_endio_pmem(wc, &list);
1442                 } else {
1443                         __writecache_endio_ssd(wc, &list);
1444                         writecache_wait_for_ios(wc, READ);
1445                 }
1446
1447                 writecache_commit_flushed(wc, false);
1448
1449                 wc_unlock(wc);
1450         }
1451
1452         return 0;
1453 }
1454
1455 static bool wc_add_block(struct writeback_struct *wb, struct wc_entry *e, gfp_t gfp)
1456 {
1457         struct dm_writecache *wc = wb->wc;
1458         unsigned block_size = wc->block_size;
1459         void *address = memory_data(wc, e);
1460
1461         persistent_memory_flush_cache(address, block_size);
1462         return bio_add_page(&wb->bio, persistent_memory_page(address),
1463                             block_size, persistent_memory_page_offset(address)) != 0;
1464 }
1465
1466 struct writeback_list {
1467         struct list_head list;
1468         size_t size;
1469 };
1470
1471 static void __writeback_throttle(struct dm_writecache *wc, struct writeback_list *wbl)
1472 {
1473         if (unlikely(wc->max_writeback_jobs)) {
1474                 if (READ_ONCE(wc->writeback_size) - wbl->size >= wc->max_writeback_jobs) {
1475                         wc_lock(wc);
1476                         while (wc->writeback_size - wbl->size >= wc->max_writeback_jobs)
1477                                 writecache_wait_on_freelist(wc);
1478                         wc_unlock(wc);
1479                 }
1480         }
1481         cond_resched();
1482 }
1483
1484 static void __writecache_writeback_pmem(struct dm_writecache *wc, struct writeback_list *wbl)
1485 {
1486         struct wc_entry *e, *f;
1487         struct bio *bio;
1488         struct writeback_struct *wb;
1489         unsigned max_pages;
1490
1491         while (wbl->size) {
1492                 wbl->size--;
1493                 e = container_of(wbl->list.prev, struct wc_entry, lru);
1494                 list_del(&e->lru);
1495
1496                 max_pages = e->wc_list_contiguous;
1497
1498                 bio = bio_alloc_bioset(GFP_NOIO, max_pages, &wc->bio_set);
1499                 wb = container_of(bio, struct writeback_struct, bio);
1500                 wb->wc = wc;
1501                 bio->bi_end_io = writecache_writeback_endio;
1502                 bio_set_dev(bio, wc->dev->bdev);
1503                 bio->bi_iter.bi_sector = read_original_sector(wc, e);
1504                 if (max_pages <= WB_LIST_INLINE ||
1505                     unlikely(!(wb->wc_list = kmalloc_array(max_pages, sizeof(struct wc_entry *),
1506                                                            GFP_NOIO | __GFP_NORETRY |
1507                                                            __GFP_NOMEMALLOC | __GFP_NOWARN)))) {
1508                         wb->wc_list = wb->wc_list_inline;
1509                         max_pages = WB_LIST_INLINE;
1510                 }
1511
1512                 BUG_ON(!wc_add_block(wb, e, GFP_NOIO));
1513
1514                 wb->wc_list[0] = e;
1515                 wb->wc_list_n = 1;
1516
1517                 while (wbl->size && wb->wc_list_n < max_pages) {
1518                         f = container_of(wbl->list.prev, struct wc_entry, lru);
1519                         if (read_original_sector(wc, f) !=
1520                             read_original_sector(wc, e) + (wc->block_size >> SECTOR_SHIFT))
1521                                 break;
1522                         if (!wc_add_block(wb, f, GFP_NOWAIT | __GFP_NOWARN))
1523                                 break;
1524                         wbl->size--;
1525                         list_del(&f->lru);
1526                         wb->wc_list[wb->wc_list_n++] = f;
1527                         e = f;
1528                 }
1529                 bio_set_op_attrs(bio, REQ_OP_WRITE, WC_MODE_FUA(wc) * REQ_FUA);
1530                 if (writecache_has_error(wc)) {
1531                         bio->bi_status = BLK_STS_IOERR;
1532                         bio_endio(bio);
1533                 } else {
1534                         submit_bio(bio);
1535                 }
1536
1537                 __writeback_throttle(wc, wbl);
1538         }
1539 }
1540
1541 static void __writecache_writeback_ssd(struct dm_writecache *wc, struct writeback_list *wbl)
1542 {
1543         struct wc_entry *e, *f;
1544         struct dm_io_region from, to;
1545         struct copy_struct *c;
1546
1547         while (wbl->size) {
1548                 unsigned n_sectors;
1549
1550                 wbl->size--;
1551                 e = container_of(wbl->list.prev, struct wc_entry, lru);
1552                 list_del(&e->lru);
1553
1554                 n_sectors = e->wc_list_contiguous << (wc->block_size_bits - SECTOR_SHIFT);
1555
1556                 from.bdev = wc->ssd_dev->bdev;
1557                 from.sector = cache_sector(wc, e);
1558                 from.count = n_sectors;
1559                 to.bdev = wc->dev->bdev;
1560                 to.sector = read_original_sector(wc, e);
1561                 to.count = n_sectors;
1562
1563                 c = mempool_alloc(&wc->copy_pool, GFP_NOIO);
1564                 c->wc = wc;
1565                 c->e = e;
1566                 c->n_entries = e->wc_list_contiguous;
1567
1568                 while ((n_sectors -= wc->block_size >> SECTOR_SHIFT)) {
1569                         wbl->size--;
1570                         f = container_of(wbl->list.prev, struct wc_entry, lru);
1571                         BUG_ON(f != e + 1);
1572                         list_del(&f->lru);
1573                         e = f;
1574                 }
1575
1576                 dm_kcopyd_copy(wc->dm_kcopyd, &from, 1, &to, 0, writecache_copy_endio, c);
1577
1578                 __writeback_throttle(wc, wbl);
1579         }
1580 }
1581
1582 static void writecache_writeback(struct work_struct *work)
1583 {
1584         struct dm_writecache *wc = container_of(work, struct dm_writecache, writeback_work);
1585         struct blk_plug plug;
1586         struct wc_entry *f, *uninitialized_var(g), *e = NULL;
1587         struct rb_node *node, *next_node;
1588         struct list_head skipped;
1589         struct writeback_list wbl;
1590         unsigned long n_walked;
1591
1592         wc_lock(wc);
1593 restart:
1594         if (writecache_has_error(wc)) {
1595                 wc_unlock(wc);
1596                 return;
1597         }
1598
1599         if (unlikely(wc->writeback_all)) {
1600                 if (writecache_wait_for_writeback(wc))
1601                         goto restart;
1602         }
1603
1604         if (wc->overwrote_committed) {
1605                 writecache_wait_for_ios(wc, WRITE);
1606         }
1607
1608         n_walked = 0;
1609         INIT_LIST_HEAD(&skipped);
1610         INIT_LIST_HEAD(&wbl.list);
1611         wbl.size = 0;
1612         while (!list_empty(&wc->lru) &&
1613                (wc->writeback_all ||
1614                 wc->freelist_size + wc->writeback_size <= wc->freelist_low_watermark)) {
1615
1616                 n_walked++;
1617                 if (unlikely(n_walked > WRITEBACK_LATENCY) &&
1618                     likely(!wc->writeback_all) && likely(!dm_suspended(wc->ti))) {
1619                         queue_work(wc->writeback_wq, &wc->writeback_work);
1620                         break;
1621                 }
1622
1623                 if (unlikely(wc->writeback_all)) {
1624                         if (unlikely(!e)) {
1625                                 writecache_flush(wc);
1626                                 e = container_of(rb_first(&wc->tree), struct wc_entry, rb_node);
1627                         } else
1628                                 e = g;
1629                 } else
1630                         e = container_of(wc->lru.prev, struct wc_entry, lru);
1631                 BUG_ON(e->write_in_progress);
1632                 if (unlikely(!writecache_entry_is_committed(wc, e))) {
1633                         writecache_flush(wc);
1634                 }
1635                 node = rb_prev(&e->rb_node);
1636                 if (node) {
1637                         f = container_of(node, struct wc_entry, rb_node);
1638                         if (unlikely(read_original_sector(wc, f) ==
1639                                      read_original_sector(wc, e))) {
1640                                 BUG_ON(!f->write_in_progress);
1641                                 list_del(&e->lru);
1642                                 list_add(&e->lru, &skipped);
1643                                 cond_resched();
1644                                 continue;
1645                         }
1646                 }
1647                 wc->writeback_size++;
1648                 list_del(&e->lru);
1649                 list_add(&e->lru, &wbl.list);
1650                 wbl.size++;
1651                 e->write_in_progress = true;
1652                 e->wc_list_contiguous = 1;
1653
1654                 f = e;
1655
1656                 while (1) {
1657                         next_node = rb_next(&f->rb_node);
1658                         if (unlikely(!next_node))
1659                                 break;
1660                         g = container_of(next_node, struct wc_entry, rb_node);
1661                         if (unlikely(read_original_sector(wc, g) ==
1662                             read_original_sector(wc, f))) {
1663                                 f = g;
1664                                 continue;
1665                         }
1666                         if (read_original_sector(wc, g) !=
1667                             read_original_sector(wc, f) + (wc->block_size >> SECTOR_SHIFT))
1668                                 break;
1669                         if (unlikely(g->write_in_progress))
1670                                 break;
1671                         if (unlikely(!writecache_entry_is_committed(wc, g)))
1672                                 break;
1673
1674                         if (!WC_MODE_PMEM(wc)) {
1675                                 if (g != f + 1)
1676                                         break;
1677                         }
1678
1679                         n_walked++;
1680                         //if (unlikely(n_walked > WRITEBACK_LATENCY) && likely(!wc->writeback_all))
1681                         //      break;
1682
1683                         wc->writeback_size++;
1684                         list_del(&g->lru);
1685                         list_add(&g->lru, &wbl.list);
1686                         wbl.size++;
1687                         g->write_in_progress = true;
1688                         g->wc_list_contiguous = BIO_MAX_PAGES;
1689                         f = g;
1690                         e->wc_list_contiguous++;
1691                         if (unlikely(e->wc_list_contiguous == BIO_MAX_PAGES)) {
1692                                 if (unlikely(wc->writeback_all)) {
1693                                         next_node = rb_next(&f->rb_node);
1694                                         if (likely(next_node))
1695                                                 g = container_of(next_node, struct wc_entry, rb_node);
1696                                 }
1697                                 break;
1698                         }
1699                 }
1700                 cond_resched();
1701         }
1702
1703         if (!list_empty(&skipped)) {
1704                 list_splice_tail(&skipped, &wc->lru);
1705                 /*
1706                  * If we didn't do any progress, we must wait until some
1707                  * writeback finishes to avoid burning CPU in a loop
1708                  */
1709                 if (unlikely(!wbl.size))
1710                         writecache_wait_for_writeback(wc);
1711         }
1712
1713         wc_unlock(wc);
1714
1715         blk_start_plug(&plug);
1716
1717         if (WC_MODE_PMEM(wc))
1718                 __writecache_writeback_pmem(wc, &wbl);
1719         else
1720                 __writecache_writeback_ssd(wc, &wbl);
1721
1722         blk_finish_plug(&plug);
1723
1724         if (unlikely(wc->writeback_all)) {
1725                 wc_lock(wc);
1726                 while (writecache_wait_for_writeback(wc));
1727                 wc_unlock(wc);
1728         }
1729 }
1730
1731 static int calculate_memory_size(uint64_t device_size, unsigned block_size,
1732                                  size_t *n_blocks_p, size_t *n_metadata_blocks_p)
1733 {
1734         uint64_t n_blocks, offset;
1735         struct wc_entry e;
1736
1737         n_blocks = device_size;
1738         do_div(n_blocks, block_size + sizeof(struct wc_memory_entry));
1739
1740         while (1) {
1741                 if (!n_blocks)
1742                         return -ENOSPC;
1743                 /* Verify the following entries[n_blocks] won't overflow */
1744                 if (n_blocks >= ((size_t)-sizeof(struct wc_memory_superblock) /
1745                                  sizeof(struct wc_memory_entry)))
1746                         return -EFBIG;
1747                 offset = offsetof(struct wc_memory_superblock, entries[n_blocks]);
1748                 offset = (offset + block_size - 1) & ~(uint64_t)(block_size - 1);
1749                 if (offset + n_blocks * block_size <= device_size)
1750                         break;
1751                 n_blocks--;
1752         }
1753
1754         /* check if the bit field overflows */
1755         e.index = n_blocks;
1756         if (e.index != n_blocks)
1757                 return -EFBIG;
1758
1759         if (n_blocks_p)
1760                 *n_blocks_p = n_blocks;
1761         if (n_metadata_blocks_p)
1762                 *n_metadata_blocks_p = offset >> __ffs(block_size);
1763         return 0;
1764 }
1765
1766 static int init_memory(struct dm_writecache *wc)
1767 {
1768         size_t b;
1769         int r;
1770
1771         r = calculate_memory_size(wc->memory_map_size, wc->block_size, &wc->n_blocks, NULL);
1772         if (r)
1773                 return r;
1774
1775         r = writecache_alloc_entries(wc);
1776         if (r)
1777                 return r;
1778
1779         for (b = 0; b < ARRAY_SIZE(sb(wc)->padding); b++)
1780                 pmem_assign(sb(wc)->padding[b], cpu_to_le64(0));
1781         pmem_assign(sb(wc)->version, cpu_to_le32(MEMORY_SUPERBLOCK_VERSION));
1782         pmem_assign(sb(wc)->block_size, cpu_to_le32(wc->block_size));
1783         pmem_assign(sb(wc)->n_blocks, cpu_to_le64(wc->n_blocks));
1784         pmem_assign(sb(wc)->seq_count, cpu_to_le64(0));
1785
1786         for (b = 0; b < wc->n_blocks; b++)
1787                 write_original_sector_seq_count(wc, &wc->entries[b], -1, -1);
1788
1789         writecache_flush_all_metadata(wc);
1790         writecache_commit_flushed(wc, false);
1791         pmem_assign(sb(wc)->magic, cpu_to_le32(MEMORY_SUPERBLOCK_MAGIC));
1792         writecache_flush_region(wc, &sb(wc)->magic, sizeof sb(wc)->magic);
1793         writecache_commit_flushed(wc, false);
1794
1795         return 0;
1796 }
1797
1798 static void writecache_dtr(struct dm_target *ti)
1799 {
1800         struct dm_writecache *wc = ti->private;
1801
1802         if (!wc)
1803                 return;
1804
1805         if (wc->endio_thread)
1806                 kthread_stop(wc->endio_thread);
1807
1808         if (wc->flush_thread)
1809                 kthread_stop(wc->flush_thread);
1810
1811         bioset_exit(&wc->bio_set);
1812
1813         mempool_exit(&wc->copy_pool);
1814
1815         if (wc->writeback_wq)
1816                 destroy_workqueue(wc->writeback_wq);
1817
1818         if (wc->dev)
1819                 dm_put_device(ti, wc->dev);
1820
1821         if (wc->ssd_dev)
1822                 dm_put_device(ti, wc->ssd_dev);
1823
1824         if (wc->entries)
1825                 vfree(wc->entries);
1826
1827         if (wc->memory_map) {
1828                 if (WC_MODE_PMEM(wc))
1829                         persistent_memory_release(wc);
1830                 else
1831                         vfree(wc->memory_map);
1832         }
1833
1834         if (wc->dm_kcopyd)
1835                 dm_kcopyd_client_destroy(wc->dm_kcopyd);
1836
1837         if (wc->dm_io)
1838                 dm_io_client_destroy(wc->dm_io);
1839
1840         if (wc->dirty_bitmap)
1841                 vfree(wc->dirty_bitmap);
1842
1843         kfree(wc);
1844 }
1845
1846 static int writecache_ctr(struct dm_target *ti, unsigned argc, char **argv)
1847 {
1848         struct dm_writecache *wc;
1849         struct dm_arg_set as;
1850         const char *string;
1851         unsigned opt_params;
1852         size_t offset, data_size;
1853         int i, r;
1854         char dummy;
1855         int high_wm_percent = HIGH_WATERMARK;
1856         int low_wm_percent = LOW_WATERMARK;
1857         uint64_t x;
1858         struct wc_memory_superblock s;
1859
1860         static struct dm_arg _args[] = {
1861                 {0, 10, "Invalid number of feature args"},
1862         };
1863
1864         as.argc = argc;
1865         as.argv = argv;
1866
1867         wc = kzalloc(sizeof(struct dm_writecache), GFP_KERNEL);
1868         if (!wc) {
1869                 ti->error = "Cannot allocate writecache structure";
1870                 r = -ENOMEM;
1871                 goto bad;
1872         }
1873         ti->private = wc;
1874         wc->ti = ti;
1875
1876         mutex_init(&wc->lock);
1877         writecache_poison_lists(wc);
1878         init_waitqueue_head(&wc->freelist_wait);
1879         timer_setup(&wc->autocommit_timer, writecache_autocommit_timer, 0);
1880
1881         for (i = 0; i < 2; i++) {
1882                 atomic_set(&wc->bio_in_progress[i], 0);
1883                 init_waitqueue_head(&wc->bio_in_progress_wait[i]);
1884         }
1885
1886         wc->dm_io = dm_io_client_create();
1887         if (IS_ERR(wc->dm_io)) {
1888                 r = PTR_ERR(wc->dm_io);
1889                 ti->error = "Unable to allocate dm-io client";
1890                 wc->dm_io = NULL;
1891                 goto bad;
1892         }
1893
1894         wc->writeback_wq = alloc_workqueue("writecache-writeback", WQ_MEM_RECLAIM, 1);
1895         if (!wc->writeback_wq) {
1896                 r = -ENOMEM;
1897                 ti->error = "Could not allocate writeback workqueue";
1898                 goto bad;
1899         }
1900         INIT_WORK(&wc->writeback_work, writecache_writeback);
1901         INIT_WORK(&wc->flush_work, writecache_flush_work);
1902
1903         raw_spin_lock_init(&wc->endio_list_lock);
1904         INIT_LIST_HEAD(&wc->endio_list);
1905         wc->endio_thread = kthread_create(writecache_endio_thread, wc, "writecache_endio");
1906         if (IS_ERR(wc->endio_thread)) {
1907                 r = PTR_ERR(wc->endio_thread);
1908                 wc->endio_thread = NULL;
1909                 ti->error = "Couldn't spawn endio thread";
1910                 goto bad;
1911         }
1912         wake_up_process(wc->endio_thread);
1913
1914         /*
1915          * Parse the mode (pmem or ssd)
1916          */
1917         string = dm_shift_arg(&as);
1918         if (!string)
1919                 goto bad_arguments;
1920
1921         if (!strcasecmp(string, "s")) {
1922                 wc->pmem_mode = false;
1923         } else if (!strcasecmp(string, "p")) {
1924 #ifdef DM_WRITECACHE_HAS_PMEM
1925                 wc->pmem_mode = true;
1926                 wc->writeback_fua = true;
1927 #else
1928                 /*
1929                  * If the architecture doesn't support persistent memory or
1930                  * the kernel doesn't support any DAX drivers, this driver can
1931                  * only be used in SSD-only mode.
1932                  */
1933                 r = -EOPNOTSUPP;
1934                 ti->error = "Persistent memory or DAX not supported on this system";
1935                 goto bad;
1936 #endif
1937         } else {
1938                 goto bad_arguments;
1939         }
1940
1941         if (WC_MODE_PMEM(wc)) {
1942                 r = bioset_init(&wc->bio_set, BIO_POOL_SIZE,
1943                                 offsetof(struct writeback_struct, bio),
1944                                 BIOSET_NEED_BVECS);
1945                 if (r) {
1946                         ti->error = "Could not allocate bio set";
1947                         goto bad;
1948                 }
1949         } else {
1950                 r = mempool_init_kmalloc_pool(&wc->copy_pool, 1, sizeof(struct copy_struct));
1951                 if (r) {
1952                         ti->error = "Could not allocate mempool";
1953                         goto bad;
1954                 }
1955         }
1956
1957         /*
1958          * Parse the origin data device
1959          */
1960         string = dm_shift_arg(&as);
1961         if (!string)
1962                 goto bad_arguments;
1963         r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->dev);
1964         if (r) {
1965                 ti->error = "Origin data device lookup failed";
1966                 goto bad;
1967         }
1968
1969         /*
1970          * Parse cache data device (be it pmem or ssd)
1971          */
1972         string = dm_shift_arg(&as);
1973         if (!string)
1974                 goto bad_arguments;
1975
1976         r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->ssd_dev);
1977         if (r) {
1978                 ti->error = "Cache data device lookup failed";
1979                 goto bad;
1980         }
1981         wc->memory_map_size = i_size_read(wc->ssd_dev->bdev->bd_inode);
1982
1983         /*
1984          * Parse the cache block size
1985          */
1986         string = dm_shift_arg(&as);
1987         if (!string)
1988                 goto bad_arguments;
1989         if (sscanf(string, "%u%c", &wc->block_size, &dummy) != 1 ||
1990             wc->block_size < 512 || wc->block_size > PAGE_SIZE ||
1991             (wc->block_size & (wc->block_size - 1))) {
1992                 r = -EINVAL;
1993                 ti->error = "Invalid block size";
1994                 goto bad;
1995         }
1996         wc->block_size_bits = __ffs(wc->block_size);
1997
1998         wc->max_writeback_jobs = MAX_WRITEBACK_JOBS;
1999         wc->autocommit_blocks = !WC_MODE_PMEM(wc) ? AUTOCOMMIT_BLOCKS_SSD : AUTOCOMMIT_BLOCKS_PMEM;
2000         wc->autocommit_jiffies = msecs_to_jiffies(AUTOCOMMIT_MSEC);
2001
2002         /*
2003          * Parse optional arguments
2004          */
2005         r = dm_read_arg_group(_args, &as, &opt_params, &ti->error);
2006         if (r)
2007                 goto bad;
2008
2009         while (opt_params) {
2010                 string = dm_shift_arg(&as), opt_params--;
2011                 if (!strcasecmp(string, "start_sector") && opt_params >= 1) {
2012                         unsigned long long start_sector;
2013                         string = dm_shift_arg(&as), opt_params--;
2014                         if (sscanf(string, "%llu%c", &start_sector, &dummy) != 1)
2015                                 goto invalid_optional;
2016                         wc->start_sector = start_sector;
2017                         if (wc->start_sector != start_sector ||
2018                             wc->start_sector >= wc->memory_map_size >> SECTOR_SHIFT)
2019                                 goto invalid_optional;
2020                 } else if (!strcasecmp(string, "high_watermark") && opt_params >= 1) {
2021                         string = dm_shift_arg(&as), opt_params--;
2022                         if (sscanf(string, "%d%c", &high_wm_percent, &dummy) != 1)
2023                                 goto invalid_optional;
2024                         if (high_wm_percent < 0 || high_wm_percent > 100)
2025                                 goto invalid_optional;
2026                         wc->high_wm_percent_set = true;
2027                 } else if (!strcasecmp(string, "low_watermark") && opt_params >= 1) {
2028                         string = dm_shift_arg(&as), opt_params--;
2029                         if (sscanf(string, "%d%c", &low_wm_percent, &dummy) != 1)
2030                                 goto invalid_optional;
2031                         if (low_wm_percent < 0 || low_wm_percent > 100)
2032                                 goto invalid_optional;
2033                         wc->low_wm_percent_set = true;
2034                 } else if (!strcasecmp(string, "writeback_jobs") && opt_params >= 1) {
2035                         string = dm_shift_arg(&as), opt_params--;
2036                         if (sscanf(string, "%u%c", &wc->max_writeback_jobs, &dummy) != 1)
2037                                 goto invalid_optional;
2038                         wc->max_writeback_jobs_set = true;
2039                 } else if (!strcasecmp(string, "autocommit_blocks") && opt_params >= 1) {
2040                         string = dm_shift_arg(&as), opt_params--;
2041                         if (sscanf(string, "%u%c", &wc->autocommit_blocks, &dummy) != 1)
2042                                 goto invalid_optional;
2043                         wc->autocommit_blocks_set = true;
2044                 } else if (!strcasecmp(string, "autocommit_time") && opt_params >= 1) {
2045                         unsigned autocommit_msecs;
2046                         string = dm_shift_arg(&as), opt_params--;
2047                         if (sscanf(string, "%u%c", &autocommit_msecs, &dummy) != 1)
2048                                 goto invalid_optional;
2049                         if (autocommit_msecs > 3600000)
2050                                 goto invalid_optional;
2051                         wc->autocommit_jiffies = msecs_to_jiffies(autocommit_msecs);
2052                         wc->autocommit_time_set = true;
2053                 } else if (!strcasecmp(string, "fua")) {
2054                         if (WC_MODE_PMEM(wc)) {
2055                                 wc->writeback_fua = true;
2056                                 wc->writeback_fua_set = true;
2057                         } else goto invalid_optional;
2058                 } else if (!strcasecmp(string, "nofua")) {
2059                         if (WC_MODE_PMEM(wc)) {
2060                                 wc->writeback_fua = false;
2061                                 wc->writeback_fua_set = true;
2062                         } else goto invalid_optional;
2063                 } else {
2064 invalid_optional:
2065                         r = -EINVAL;
2066                         ti->error = "Invalid optional argument";
2067                         goto bad;
2068                 }
2069         }
2070
2071         if (high_wm_percent < low_wm_percent) {
2072                 r = -EINVAL;
2073                 ti->error = "High watermark must be greater than or equal to low watermark";
2074                 goto bad;
2075         }
2076
2077         if (WC_MODE_PMEM(wc)) {
2078                 r = persistent_memory_claim(wc);
2079                 if (r) {
2080                         ti->error = "Unable to map persistent memory for cache";
2081                         goto bad;
2082                 }
2083         } else {
2084                 struct dm_io_region region;
2085                 struct dm_io_request req;
2086                 size_t n_blocks, n_metadata_blocks;
2087                 uint64_t n_bitmap_bits;
2088
2089                 wc->memory_map_size -= (uint64_t)wc->start_sector << SECTOR_SHIFT;
2090
2091                 bio_list_init(&wc->flush_list);
2092                 wc->flush_thread = kthread_create(writecache_flush_thread, wc, "dm_writecache_flush");
2093                 if (IS_ERR(wc->flush_thread)) {
2094                         r = PTR_ERR(wc->flush_thread);
2095                         wc->flush_thread = NULL;
2096                         ti->error = "Couldn't spawn flush thread";
2097                         goto bad;
2098                 }
2099                 wake_up_process(wc->flush_thread);
2100
2101                 r = calculate_memory_size(wc->memory_map_size, wc->block_size,
2102                                           &n_blocks, &n_metadata_blocks);
2103                 if (r) {
2104                         ti->error = "Invalid device size";
2105                         goto bad;
2106                 }
2107
2108                 n_bitmap_bits = (((uint64_t)n_metadata_blocks << wc->block_size_bits) +
2109                                  BITMAP_GRANULARITY - 1) / BITMAP_GRANULARITY;
2110                 /* this is limitation of test_bit functions */
2111                 if (n_bitmap_bits > 1U << 31) {
2112                         r = -EFBIG;
2113                         ti->error = "Invalid device size";
2114                         goto bad;
2115                 }
2116
2117                 wc->memory_map = vmalloc(n_metadata_blocks << wc->block_size_bits);
2118                 if (!wc->memory_map) {
2119                         r = -ENOMEM;
2120                         ti->error = "Unable to allocate memory for metadata";
2121                         goto bad;
2122                 }
2123
2124                 wc->dm_kcopyd = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2125                 if (IS_ERR(wc->dm_kcopyd)) {
2126                         r = PTR_ERR(wc->dm_kcopyd);
2127                         ti->error = "Unable to allocate dm-kcopyd client";
2128                         wc->dm_kcopyd = NULL;
2129                         goto bad;
2130                 }
2131
2132                 wc->metadata_sectors = n_metadata_blocks << (wc->block_size_bits - SECTOR_SHIFT);
2133                 wc->dirty_bitmap_size = (n_bitmap_bits + BITS_PER_LONG - 1) /
2134                         BITS_PER_LONG * sizeof(unsigned long);
2135                 wc->dirty_bitmap = vzalloc(wc->dirty_bitmap_size);
2136                 if (!wc->dirty_bitmap) {
2137                         r = -ENOMEM;
2138                         ti->error = "Unable to allocate dirty bitmap";
2139                         goto bad;
2140                 }
2141
2142                 region.bdev = wc->ssd_dev->bdev;
2143                 region.sector = wc->start_sector;
2144                 region.count = wc->metadata_sectors;
2145                 req.bi_op = REQ_OP_READ;
2146                 req.bi_op_flags = REQ_SYNC;
2147                 req.mem.type = DM_IO_VMA;
2148                 req.mem.ptr.vma = (char *)wc->memory_map;
2149                 req.client = wc->dm_io;
2150                 req.notify.fn = NULL;
2151
2152                 r = dm_io(&req, 1, &region, NULL);
2153                 if (r) {
2154                         ti->error = "Unable to read metadata";
2155                         goto bad;
2156                 }
2157         }
2158
2159         r = memcpy_mcsafe(&s, sb(wc), sizeof(struct wc_memory_superblock));
2160         if (r) {
2161                 ti->error = "Hardware memory error when reading superblock";
2162                 goto bad;
2163         }
2164         if (!le32_to_cpu(s.magic) && !le32_to_cpu(s.version)) {
2165                 r = init_memory(wc);
2166                 if (r) {
2167                         ti->error = "Unable to initialize device";
2168                         goto bad;
2169                 }
2170                 r = memcpy_mcsafe(&s, sb(wc), sizeof(struct wc_memory_superblock));
2171                 if (r) {
2172                         ti->error = "Hardware memory error when reading superblock";
2173                         goto bad;
2174                 }
2175         }
2176
2177         if (le32_to_cpu(s.magic) != MEMORY_SUPERBLOCK_MAGIC) {
2178                 ti->error = "Invalid magic in the superblock";
2179                 r = -EINVAL;
2180                 goto bad;
2181         }
2182
2183         if (le32_to_cpu(s.version) != MEMORY_SUPERBLOCK_VERSION) {
2184                 ti->error = "Invalid version in the superblock";
2185                 r = -EINVAL;
2186                 goto bad;
2187         }
2188
2189         if (le32_to_cpu(s.block_size) != wc->block_size) {
2190                 ti->error = "Block size does not match superblock";
2191                 r = -EINVAL;
2192                 goto bad;
2193         }
2194
2195         wc->n_blocks = le64_to_cpu(s.n_blocks);
2196
2197         offset = wc->n_blocks * sizeof(struct wc_memory_entry);
2198         if (offset / sizeof(struct wc_memory_entry) != le64_to_cpu(sb(wc)->n_blocks)) {
2199 overflow:
2200                 ti->error = "Overflow in size calculation";
2201                 r = -EINVAL;
2202                 goto bad;
2203         }
2204         offset += sizeof(struct wc_memory_superblock);
2205         if (offset < sizeof(struct wc_memory_superblock))
2206                 goto overflow;
2207         offset = (offset + wc->block_size - 1) & ~(size_t)(wc->block_size - 1);
2208         data_size = wc->n_blocks * (size_t)wc->block_size;
2209         if (!offset || (data_size / wc->block_size != wc->n_blocks) ||
2210             (offset + data_size < offset))
2211                 goto overflow;
2212         if (offset + data_size > wc->memory_map_size) {
2213                 ti->error = "Memory area is too small";
2214                 r = -EINVAL;
2215                 goto bad;
2216         }
2217
2218         wc->metadata_sectors = offset >> SECTOR_SHIFT;
2219         wc->block_start = (char *)sb(wc) + offset;
2220
2221         x = (uint64_t)wc->n_blocks * (100 - high_wm_percent);
2222         x += 50;
2223         do_div(x, 100);
2224         wc->freelist_high_watermark = x;
2225         x = (uint64_t)wc->n_blocks * (100 - low_wm_percent);
2226         x += 50;
2227         do_div(x, 100);
2228         wc->freelist_low_watermark = x;
2229
2230         r = writecache_alloc_entries(wc);
2231         if (r) {
2232                 ti->error = "Cannot allocate memory";
2233                 goto bad;
2234         }
2235
2236         ti->num_flush_bios = 1;
2237         ti->flush_supported = true;
2238         ti->num_discard_bios = 1;
2239
2240         if (WC_MODE_PMEM(wc))
2241                 persistent_memory_flush_cache(wc->memory_map, wc->memory_map_size);
2242
2243         return 0;
2244
2245 bad_arguments:
2246         r = -EINVAL;
2247         ti->error = "Bad arguments";
2248 bad:
2249         writecache_dtr(ti);
2250         return r;
2251 }
2252
2253 static void writecache_status(struct dm_target *ti, status_type_t type,
2254                               unsigned status_flags, char *result, unsigned maxlen)
2255 {
2256         struct dm_writecache *wc = ti->private;
2257         unsigned extra_args;
2258         unsigned sz = 0;
2259         uint64_t x;
2260
2261         switch (type) {
2262         case STATUSTYPE_INFO:
2263                 DMEMIT("%ld %llu %llu %llu", writecache_has_error(wc),
2264                        (unsigned long long)wc->n_blocks, (unsigned long long)wc->freelist_size,
2265                        (unsigned long long)wc->writeback_size);
2266                 break;
2267         case STATUSTYPE_TABLE:
2268                 DMEMIT("%c %s %s %u ", WC_MODE_PMEM(wc) ? 'p' : 's',
2269                                 wc->dev->name, wc->ssd_dev->name, wc->block_size);
2270                 extra_args = 0;
2271                 if (wc->start_sector)
2272                         extra_args += 2;
2273                 if (wc->high_wm_percent_set)
2274                         extra_args += 2;
2275                 if (wc->low_wm_percent_set)
2276                         extra_args += 2;
2277                 if (wc->max_writeback_jobs_set)
2278                         extra_args += 2;
2279                 if (wc->autocommit_blocks_set)
2280                         extra_args += 2;
2281                 if (wc->autocommit_time_set)
2282                         extra_args += 2;
2283                 if (wc->writeback_fua_set)
2284                         extra_args++;
2285
2286                 DMEMIT("%u", extra_args);
2287                 if (wc->start_sector)
2288                         DMEMIT(" start_sector %llu", (unsigned long long)wc->start_sector);
2289                 if (wc->high_wm_percent_set) {
2290                         x = (uint64_t)wc->freelist_high_watermark * 100;
2291                         x += wc->n_blocks / 2;
2292                         do_div(x, (size_t)wc->n_blocks);
2293                         DMEMIT(" high_watermark %u", 100 - (unsigned)x);
2294                 }
2295                 if (wc->low_wm_percent_set) {
2296                         x = (uint64_t)wc->freelist_low_watermark * 100;
2297                         x += wc->n_blocks / 2;
2298                         do_div(x, (size_t)wc->n_blocks);
2299                         DMEMIT(" low_watermark %u", 100 - (unsigned)x);
2300                 }
2301                 if (wc->max_writeback_jobs_set)
2302                         DMEMIT(" writeback_jobs %u", wc->max_writeback_jobs);
2303                 if (wc->autocommit_blocks_set)
2304                         DMEMIT(" autocommit_blocks %u", wc->autocommit_blocks);
2305                 if (wc->autocommit_time_set)
2306                         DMEMIT(" autocommit_time %u", jiffies_to_msecs(wc->autocommit_jiffies));
2307                 if (wc->writeback_fua_set)
2308                         DMEMIT(" %sfua", wc->writeback_fua ? "" : "no");
2309                 break;
2310         }
2311 }
2312
2313 static struct target_type writecache_target = {
2314         .name                   = "writecache",
2315         .version                = {1, 1, 1},
2316         .module                 = THIS_MODULE,
2317         .ctr                    = writecache_ctr,
2318         .dtr                    = writecache_dtr,
2319         .status                 = writecache_status,
2320         .postsuspend            = writecache_suspend,
2321         .resume                 = writecache_resume,
2322         .message                = writecache_message,
2323         .map                    = writecache_map,
2324         .end_io                 = writecache_end_io,
2325         .iterate_devices        = writecache_iterate_devices,
2326         .io_hints               = writecache_io_hints,
2327 };
2328
2329 static int __init dm_writecache_init(void)
2330 {
2331         int r;
2332
2333         r = dm_register_target(&writecache_target);
2334         if (r < 0) {
2335                 DMERR("register failed %d", r);
2336                 return r;
2337         }
2338
2339         return 0;
2340 }
2341
2342 static void __exit dm_writecache_exit(void)
2343 {
2344         dm_unregister_target(&writecache_target);
2345 }
2346
2347 module_init(dm_writecache_init);
2348 module_exit(dm_writecache_exit);
2349
2350 MODULE_DESCRIPTION(DM_NAME " writecache target");
2351 MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
2352 MODULE_LICENSE("GPL");