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         unsigned page_offset;
 194         struct page *page;
 195         struct wc_entry *wc_list_inline[WB_LIST_INLINE];
 196         struct bio bio;
 197 };
 198
 199 struct copy_struct {
 200         struct list_head endio_entry;
 201         struct dm_writecache *wc;
 202         struct wc_entry *e;
 203         unsigned n_entries;
 204         int error;
 205 };
 206
 207 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(dm_writecache_throttle,
 208                                             "A percentage of time allocated for data copying");
 209
 210 static void wc_lock(struct dm_writecache *wc)
 211 {
 212         mutex_lock(&wc->lock);
 213 }
 214
 215 static void wc_unlock(struct dm_writecache *wc)
 216 {
 217         mutex_unlock(&wc->lock);
 218 }
 219
 220 #ifdef DM_WRITECACHE_HAS_PMEM
 221 static int persistent_memory_claim(struct dm_writecache *wc)
 222 {
 223         int r;
 224         loff_t s;
 225         long p, da;
 226         pfn_t pfn;
 227         int id;
 228         struct page **pages;
 229
 230         wc->memory_vmapped = false;
 231
 232         if (!wc->ssd_dev->dax_dev) {
 233                 r = -EOPNOTSUPP;
 234                 goto err1;
 235         }
 236         s = wc->memory_map_size;
 237         p = s >> PAGE_SHIFT;
 238         if (!p) {
 239                 r = -EINVAL;
 240                 goto err1;
 241         }
 242         if (p != s >> PAGE_SHIFT) {
 243                 r = -EOVERFLOW;
 244                 goto err1;
 245         }
 246
 247         id = dax_read_lock();
 248
 249         da = dax_direct_access(wc->ssd_dev->dax_dev, 0, p, &wc->memory_map, &pfn);
 250         if (da < 0) {
 251                 wc->memory_map = NULL;
 252                 r = da;
 253                 goto err2;
 254         }
 255         if (!pfn_t_has_page(pfn)) {
 256                 wc->memory_map = NULL;
 257                 r = -EOPNOTSUPP;
 258                 goto err2;
 259         }
 260         if (da != p) {
 261                 long i;
 262                 wc->memory_map = NULL;
 263                 pages = kvmalloc_array(p, sizeof(struct page *), GFP_KERNEL);
 264                 if (!pages) {
 265                         r = -ENOMEM;
 266                         goto err2;
 267                 }
 268                 i = 0;
 269                 do {
 270                         long daa;
 271                         void *dummy_addr;
 272                         daa = dax_direct_access(wc->ssd_dev->dax_dev, i, p - i,
 273                                                 &dummy_addr, &pfn);
 274                         if (daa <= 0) {
 275                                 r = daa ? daa : -EINVAL;
 276                                 goto err3;
 277                         }
 278                         if (!pfn_t_has_page(pfn)) {
 279                                 r = -EOPNOTSUPP;
 280                                 goto err3;
 281                         }
 282                         while (daa-- && i < p) {
 283                                 pages[i++] = pfn_t_to_page(pfn);
 284                                 pfn.val++;
 285                         }
 286                 } while (i < p);
 287                 wc->memory_map = vmap(pages, p, VM_MAP, PAGE_KERNEL);
 288                 if (!wc->memory_map) {
 289                         r = -ENOMEM;
 290                         goto err3;
 291                 }
 292                 kvfree(pages);
 293                 wc->memory_vmapped = true;
 294         }
 295
 296         dax_read_unlock(id);
 297
 298         wc->memory_map += (size_t)wc->start_sector << SECTOR_SHIFT;
 299         wc->memory_map_size -= (size_t)wc->start_sector << SECTOR_SHIFT;
 300
 301         return 0;
 302 err3:
 303         kvfree(pages);
 304 err2:
 305         dax_read_unlock(id);
 306 err1:
 307         return r;
 308 }
 309 #else
 310 static int persistent_memory_claim(struct dm_writecache *wc)
 311 {
 312         BUG();
 313 }
 314 #endif
 315
 316 static void persistent_memory_release(struct dm_writecache *wc)
 317 {
 318         if (wc->memory_vmapped)
 319                 vunmap(wc->memory_map - ((size_t)wc->start_sector << SECTOR_SHIFT));
 320 }
 321
 322 static struct page *persistent_memory_page(void *addr)
 323 {
 324         if (is_vmalloc_addr(addr))
 325                 return vmalloc_to_page(addr);
 326         else
 327                 return virt_to_page(addr);
 328 }
 329
 330 static unsigned persistent_memory_page_offset(void *addr)
 331 {
 332         return (unsigned long)addr & (PAGE_SIZE - 1);
 333 }
 334
 335 static void persistent_memory_flush_cache(void *ptr, size_t size)
 336 {
 337         if (is_vmalloc_addr(ptr))
 338                 flush_kernel_vmap_range(ptr, size);
 339 }
 340
 341 static void persistent_memory_invalidate_cache(void *ptr, size_t size)
 342 {
 343         if (is_vmalloc_addr(ptr))
 344                 invalidate_kernel_vmap_range(ptr, size);
 345 }
 346
 347 static struct wc_memory_superblock *sb(struct dm_writecache *wc)
 348 {
 349         return wc->memory_map;
 350 }
 351
 352 static struct wc_memory_entry *memory_entry(struct dm_writecache *wc, struct wc_entry *e)
 353 {
 354         if (is_power_of_2(sizeof(struct wc_entry)) && 0)
 355                 return &sb(wc)->entries[e - wc->entries];
 356         else
 357                 return &sb(wc)->entries[e->index];
 358 }
 359
 360 static void *memory_data(struct dm_writecache *wc, struct wc_entry *e)
 361 {
 362         return (char *)wc->block_start + (e->index << wc->block_size_bits);
 363 }
 364
 365 static sector_t cache_sector(struct dm_writecache *wc, struct wc_entry *e)
 366 {
 367         return wc->start_sector + wc->metadata_sectors +
 368                 ((sector_t)e->index << (wc->block_size_bits - SECTOR_SHIFT));
 369 }
 370
 371 static uint64_t read_original_sector(struct dm_writecache *wc, struct wc_entry *e)
 372 {
 373 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 374         return e->original_sector;
 375 #else
 376         return le64_to_cpu(memory_entry(wc, e)->original_sector);
 377 #endif
 378 }
 379
 380 static uint64_t read_seq_count(struct dm_writecache *wc, struct wc_entry *e)
 381 {
 382 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 383         return e->seq_count;
 384 #else
 385         return le64_to_cpu(memory_entry(wc, e)->seq_count);
 386 #endif
 387 }
 388
 389 static void clear_seq_count(struct dm_writecache *wc, struct wc_entry *e)
 390 {
 391 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 392         e->seq_count = -1;
 393 #endif
 394         pmem_assign(memory_entry(wc, e)->seq_count, cpu_to_le64(-1));
 395 }
 396
 397 static void write_original_sector_seq_count(struct dm_writecache *wc, struct wc_entry *e,
 398                                             uint64_t original_sector, uint64_t seq_count)
 399 {
 400         struct wc_memory_entry me;
 401 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 402         e->original_sector = original_sector;
 403         e->seq_count = seq_count;
 404 #endif
 405         me.original_sector = cpu_to_le64(original_sector);
 406         me.seq_count = cpu_to_le64(seq_count);
 407         pmem_assign(*memory_entry(wc, e), me);
 408 }
 409
 410 #define writecache_error(wc, err, msg, arg...)                          \
 411 do {                                                                    \
 412         if (!cmpxchg(&(wc)->error, 0, err))                             \
 413                 DMERR(msg, ##arg);                                      \
 414         wake_up(&(wc)->freelist_wait);                                  \
 415 } while (0)
 416
 417 #define writecache_has_error(wc)        (unlikely(READ_ONCE((wc)->error)))
 418
 419 static void writecache_flush_all_metadata(struct dm_writecache *wc)
 420 {
 421         if (!WC_MODE_PMEM(wc))
 422                 memset(wc->dirty_bitmap, -1, wc->dirty_bitmap_size);
 423 }
 424
 425 static void writecache_flush_region(struct dm_writecache *wc, void *ptr, size_t size)
 426 {
 427         if (!WC_MODE_PMEM(wc))
 428                 __set_bit(((char *)ptr - (char *)wc->memory_map) / BITMAP_GRANULARITY,
 429                           wc->dirty_bitmap);
 430 }
 431
 432 static void writecache_disk_flush(struct dm_writecache *wc, struct dm_dev *dev);
 433
 434 struct io_notify {
 435         struct dm_writecache *wc;
 436         struct completion c;
 437         atomic_t count;
 438 };
 439
 440 static void writecache_notify_io(unsigned long error, void *context)
 441 {
 442         struct io_notify *endio = context;
 443
 444         if (unlikely(error != 0))
 445                 writecache_error(endio->wc, -EIO, "error writing metadata");
 446         BUG_ON(atomic_read(&endio->count) <= 0);
 447         if (atomic_dec_and_test(&endio->count))
 448                 complete(&endio->c);
 449 }
 450
 451 static void ssd_commit_flushed(struct dm_writecache *wc)
 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 * BITS_PER_LONG;
 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         writecache_disk_flush(wc, wc->ssd_dev);
 498
 499         memset(wc->dirty_bitmap, 0, wc->dirty_bitmap_size);
 500 }
 501
 502 static void writecache_commit_flushed(struct dm_writecache *wc)
 503 {
 504         if (WC_MODE_PMEM(wc))
 505                 wmb();
 506         else
 507                 ssd_commit_flushed(wc);
 508 }
 509
 510 static void writecache_disk_flush(struct dm_writecache *wc, struct dm_dev *dev)
 511 {
 512         int r;
 513         struct dm_io_region region;
 514         struct dm_io_request req;
 515
 516         region.bdev = dev->bdev;
 517         region.sector = 0;
 518         region.count = 0;
 519         req.bi_op = REQ_OP_WRITE;
 520         req.bi_op_flags = REQ_PREFLUSH;
 521         req.mem.type = DM_IO_KMEM;
 522         req.mem.ptr.addr = NULL;
 523         req.client = wc->dm_io;
 524         req.notify.fn = NULL;
 525
 526         r = dm_io(&req, 1, &region, NULL);
 527         if (unlikely(r))
 528                 writecache_error(wc, r, "error flushing metadata: %d", r);
 529 }
 530
 531 static void writecache_wait_for_ios(struct dm_writecache *wc, int direction)
 532 {
 533         wait_event(wc->bio_in_progress_wait[direction],
 534                    !atomic_read(&wc->bio_in_progress[direction]));
 535 }
 536
 537 #define WFE_RETURN_FOLLOWING    1
 538 #define WFE_LOWEST_SEQ          2
 539
 540 static struct wc_entry *writecache_find_entry(struct dm_writecache *wc,
 541                                               uint64_t block, int flags)
 542 {
 543         struct wc_entry *e;
 544         struct rb_node *node = wc->tree.rb_node;
 545
 546         if (unlikely(!node))
 547                 return NULL;
 548
 549         while (1) {
 550                 e = container_of(node, struct wc_entry, rb_node);
 551                 if (read_original_sector(wc, e) == block)
 552                         break;
 553                 node = (read_original_sector(wc, e) >= block ?
 554                         e->rb_node.rb_left : e->rb_node.rb_right);
 555                 if (unlikely(!node)) {
 556                         if (!(flags & WFE_RETURN_FOLLOWING)) {
 557                                 return NULL;
 558                         }
 559                         if (read_original_sector(wc, e) >= block) {
 560                                 break;
 561                         } else {
 562                                 node = rb_next(&e->rb_node);
 563                                 if (unlikely(!node)) {
 564                                         return NULL;
 565                                 }
 566                                 e = container_of(node, struct wc_entry, rb_node);
 567                                 break;
 568                         }
 569                 }
 570         }
 571
 572         while (1) {
 573                 struct wc_entry *e2;
 574                 if (flags & WFE_LOWEST_SEQ)
 575                         node = rb_prev(&e->rb_node);
 576                 else
 577                         node = rb_next(&e->rb_node);
 578                 if (!node)
 579                         return e;
 580                 e2 = container_of(node, struct wc_entry, rb_node);
 581                 if (read_original_sector(wc, e2) != block)
 582                         return e;
 583                 e = e2;
 584         }
 585 }
 586
 587 static void writecache_insert_entry(struct dm_writecache *wc, struct wc_entry *ins)
 588 {
 589         struct wc_entry *e;
 590         struct rb_node **node = &wc->tree.rb_node, *parent = NULL;
 591
 592         while (*node) {
 593                 e = container_of(*node, struct wc_entry, rb_node);
 594                 parent = &e->rb_node;
 595                 if (read_original_sector(wc, e) > read_original_sector(wc, ins))
 596                         node = &parent->rb_left;
 597                 else
 598                         node = &parent->rb_right;
 599         }
 600         rb_link_node(&ins->rb_node, parent, node);
 601         rb_insert_color(&ins->rb_node, &wc->tree);
 602         list_add(&ins->lru, &wc->lru);
 603 }
 604
 605 static void writecache_unlink(struct dm_writecache *wc, struct wc_entry *e)
 606 {
 607         list_del(&e->lru);
 608         rb_erase(&e->rb_node, &wc->tree);
 609 }
 610
 611 static void writecache_add_to_freelist(struct dm_writecache *wc, struct wc_entry *e)
 612 {
 613         if (WC_MODE_SORT_FREELIST(wc)) {
 614                 struct rb_node **node = &wc->freetree.rb_node, *parent = NULL;
 615                 if (unlikely(!*node))
 616                         wc->current_free = e;
 617                 while (*node) {
 618                         parent = *node;
 619                         if (&e->rb_node < *node)
 620                                 node = &parent->rb_left;
 621                         else
 622                                 node = &parent->rb_right;
 623                 }
 624                 rb_link_node(&e->rb_node, parent, node);
 625                 rb_insert_color(&e->rb_node, &wc->freetree);
 626         } else {
 627                 list_add_tail(&e->lru, &wc->freelist);
 628         }
 629         wc->freelist_size++;
 630 }
 631
 632 static struct wc_entry *writecache_pop_from_freelist(struct dm_writecache *wc)
 633 {
 634         struct wc_entry *e;
 635
 636         if (WC_MODE_SORT_FREELIST(wc)) {
 637                 struct rb_node *next;
 638                 if (unlikely(!wc->current_free))
 639                         return NULL;
 640                 e = wc->current_free;
 641                 next = rb_next(&e->rb_node);
 642                 rb_erase(&e->rb_node, &wc->freetree);
 643                 if (unlikely(!next))
 644                         next = rb_first(&wc->freetree);
 645                 wc->current_free = next ? container_of(next, struct wc_entry, rb_node) : NULL;
 646         } else {
 647                 if (unlikely(list_empty(&wc->freelist)))
 648                         return NULL;
 649                 e = container_of(wc->freelist.next, struct wc_entry, lru);
 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);
 735
 736         writecache_wait_for_ios(wc, WRITE);
 737
 738         wc->seq_count++;
 739         pmem_assign(sb(wc)->seq_count, cpu_to_le64(wc->seq_count));
 740         writecache_flush_region(wc, &sb(wc)->seq_count, sizeof sb(wc)->seq_count);
 741         writecache_commit_flushed(wc);
 742
 743         wc->overwrote_committed = false;
 744
 745         need_flush_after_free = false;
 746         while (1) {
 747                 /* Free another committed entry with lower seq-count */
 748                 struct rb_node *rb_node = rb_prev(&e->rb_node);
 749
 750                 if (rb_node) {
 751                         e2 = container_of(rb_node, struct wc_entry, rb_node);
 752                         if (read_original_sector(wc, e2) == read_original_sector(wc, e) &&
 753                             likely(!e2->write_in_progress)) {
 754                                 writecache_free_entry(wc, e2);
 755                                 need_flush_after_free = true;
 756                         }
 757                 }
 758                 if (unlikely(e->lru.prev == &wc->lru))
 759                         break;
 760                 e = container_of(e->lru.prev, struct wc_entry, lru);
 761                 cond_resched();
 762         }
 763
 764         if (need_flush_after_free)
 765                 writecache_commit_flushed(wc);
 766 }
 767
 768 static void writecache_flush_work(struct work_struct *work)
 769 {
 770         struct dm_writecache *wc = container_of(work, struct dm_writecache, flush_work);
 771
 772         wc_lock(wc);
 773         writecache_flush(wc);
 774         wc_unlock(wc);
 775 }
 776
 777 static void writecache_autocommit_timer(struct timer_list *t)
 778 {
 779         struct dm_writecache *wc = from_timer(wc, t, autocommit_timer);
 780         if (!writecache_has_error(wc))
 781                 queue_work(wc->writeback_wq, &wc->flush_work);
 782 }
 783
 784 static void writecache_schedule_autocommit(struct dm_writecache *wc)
 785 {
 786         if (!timer_pending(&wc->autocommit_timer))
 787                 mod_timer(&wc->autocommit_timer, jiffies + wc->autocommit_jiffies);
 788 }
 789
 790 static void writecache_discard(struct dm_writecache *wc, sector_t start, sector_t end)
 791 {
 792         struct wc_entry *e;
 793         bool discarded_something = false;
 794
 795         e = writecache_find_entry(wc, start, WFE_RETURN_FOLLOWING | WFE_LOWEST_SEQ);
 796         if (unlikely(!e))
 797                 return;
 798
 799         while (read_original_sector(wc, e) < end) {
 800                 struct rb_node *node = rb_next(&e->rb_node);
 801
 802                 if (likely(!e->write_in_progress)) {
 803                         if (!discarded_something) {
 804                                 writecache_wait_for_ios(wc, READ);
 805                                 writecache_wait_for_ios(wc, WRITE);
 806                                 discarded_something = true;
 807                         }
 808                         writecache_free_entry(wc, e);
 809                 }
 810
 811                 if (!node)
 812                         break;
 813
 814                 e = container_of(node, struct wc_entry, rb_node);
 815         }
 816
 817         if (discarded_something)
 818                 writecache_commit_flushed(wc);
 819 }
 820
 821 static bool writecache_wait_for_writeback(struct dm_writecache *wc)
 822 {
 823         if (wc->writeback_size) {
 824                 writecache_wait_on_freelist(wc);
 825                 return true;
 826         }
 827         return false;
 828 }
 829
 830 static void writecache_suspend(struct dm_target *ti)
 831 {
 832         struct dm_writecache *wc = ti->private;
 833         bool flush_on_suspend;
 834
 835         del_timer_sync(&wc->autocommit_timer);
 836
 837         wc_lock(wc);
 838         writecache_flush(wc);
 839         flush_on_suspend = wc->flush_on_suspend;
 840         if (flush_on_suspend) {
 841                 wc->flush_on_suspend = false;
 842                 wc->writeback_all++;
 843                 queue_work(wc->writeback_wq, &wc->writeback_work);
 844         }
 845         wc_unlock(wc);
 846
 847         flush_workqueue(wc->writeback_wq);
 848
 849         wc_lock(wc);
 850         if (flush_on_suspend)
 851                 wc->writeback_all--;
 852         while (writecache_wait_for_writeback(wc));
 853
 854         if (WC_MODE_PMEM(wc))
 855                 persistent_memory_flush_cache(wc->memory_map, wc->memory_map_size);
 856
 857         writecache_poison_lists(wc);
 858
 859         wc_unlock(wc);
 860 }
 861
 862 static int writecache_alloc_entries(struct dm_writecache *wc)
 863 {
 864         size_t b;
 865
 866         if (wc->entries)
 867                 return 0;
 868         wc->entries = vmalloc(array_size(sizeof(struct wc_entry), wc->n_blocks));
 869         if (!wc->entries)
 870                 return -ENOMEM;
 871         for (b = 0; b < wc->n_blocks; b++) {
 872                 struct wc_entry *e = &wc->entries[b];
 873                 e->index = b;
 874                 e->write_in_progress = false;
 875         }
 876
 877         return 0;
 878 }
 879
 880 static void writecache_resume(struct dm_target *ti)
 881 {
 882         struct dm_writecache *wc = ti->private;
 883         size_t b;
 884         bool need_flush = false;
 885         __le64 sb_seq_count;
 886         int r;
 887
 888         wc_lock(wc);
 889
 890         if (WC_MODE_PMEM(wc))
 891                 persistent_memory_invalidate_cache(wc->memory_map, wc->memory_map_size);
 892
 893         wc->tree = RB_ROOT;
 894         INIT_LIST_HEAD(&wc->lru);
 895         if (WC_MODE_SORT_FREELIST(wc)) {
 896                 wc->freetree = RB_ROOT;
 897                 wc->current_free = NULL;
 898         } else {
 899                 INIT_LIST_HEAD(&wc->freelist);
 900         }
 901         wc->freelist_size = 0;
 902
 903         r = memcpy_mcsafe(&sb_seq_count, &sb(wc)->seq_count, sizeof(uint64_t));
 904         if (r) {
 905                 writecache_error(wc, r, "hardware memory error when reading superblock: %d", r);
 906                 sb_seq_count = cpu_to_le64(0);
 907         }
 908         wc->seq_count = le64_to_cpu(sb_seq_count);
 909
 910 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 911         for (b = 0; b < wc->n_blocks; b++) {
 912                 struct wc_entry *e = &wc->entries[b];
 913                 struct wc_memory_entry wme;
 914                 if (writecache_has_error(wc)) {
 915                         e->original_sector = -1;
 916                         e->seq_count = -1;
 917                         continue;
 918                 }
 919                 r = memcpy_mcsafe(&wme, memory_entry(wc, e), sizeof(struct wc_memory_entry));
 920                 if (r) {
 921                         writecache_error(wc, r, "hardware memory error when reading metadata entry %lu: %d",
 922                                          (unsigned long)b, r);
 923                         e->original_sector = -1;
 924                         e->seq_count = -1;
 925                 } else {
 926                         e->original_sector = le64_to_cpu(wme.original_sector);
 927                         e->seq_count = le64_to_cpu(wme.seq_count);
 928                 }
 929         }
 930 #endif
 931         for (b = 0; b < wc->n_blocks; b++) {
 932                 struct wc_entry *e = &wc->entries[b];
 933                 if (!writecache_entry_is_committed(wc, e)) {
 934                         if (read_seq_count(wc, e) != -1) {
 935 erase_this:
 936                                 clear_seq_count(wc, e);
 937                                 need_flush = true;
 938                         }
 939                         writecache_add_to_freelist(wc, e);
 940                 } else {
 941                         struct wc_entry *old;
 942
 943                         old = writecache_find_entry(wc, read_original_sector(wc, e), 0);
 944                         if (!old) {
 945                                 writecache_insert_entry(wc, e);
 946                         } else {
 947                                 if (read_seq_count(wc, old) == read_seq_count(wc, e)) {
 948                                         writecache_error(wc, -EINVAL,
 949                                                  "two identical entries, position %llu, sector %llu, sequence %llu",
 950                                                  (unsigned long long)b, (unsigned long long)read_original_sector(wc, e),
 951                                                  (unsigned long long)read_seq_count(wc, e));
 952                                 }
 953                                 if (read_seq_count(wc, old) > read_seq_count(wc, e)) {
 954                                         goto erase_this;
 955                                 } else {
 956                                         writecache_free_entry(wc, old);
 957                                         writecache_insert_entry(wc, e);
 958                                         need_flush = true;
 959                                 }
 960                         }
 961                 }
 962                 cond_resched();
 963         }
 964
 965         if (need_flush) {
 966                 writecache_flush_all_metadata(wc);
 967                 writecache_commit_flushed(wc);
 968         }
 969
 970         wc_unlock(wc);
 971 }
 972
 973 static int process_flush_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
 974 {
 975         if (argc != 1)
 976                 return -EINVAL;
 977
 978         wc_lock(wc);
 979         if (dm_suspended(wc->ti)) {
 980                 wc_unlock(wc);
 981                 return -EBUSY;
 982         }
 983         if (writecache_has_error(wc)) {
 984                 wc_unlock(wc);
 985                 return -EIO;
 986         }
 987
 988         writecache_flush(wc);
 989         wc->writeback_all++;
 990         queue_work(wc->writeback_wq, &wc->writeback_work);
 991         wc_unlock(wc);
 992
 993         flush_workqueue(wc->writeback_wq);
 994
 995         wc_lock(wc);
 996         wc->writeback_all--;
 997         if (writecache_has_error(wc)) {
 998                 wc_unlock(wc);
 999                 return -EIO;
1000         }
1001         wc_unlock(wc);
1002
1003         return 0;
1004 }
1005
1006 static int process_flush_on_suspend_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1007 {
1008         if (argc != 1)
1009                 return -EINVAL;
1010
1011         wc_lock(wc);
1012         wc->flush_on_suspend = true;
1013         wc_unlock(wc);
1014
1015         return 0;
1016 }
1017
1018 static int writecache_message(struct dm_target *ti, unsigned argc, char **argv,
1019                               char *result, unsigned maxlen)
1020 {
1021         int r = -EINVAL;
1022         struct dm_writecache *wc = ti->private;
1023
1024         if (!strcasecmp(argv[0], "flush"))
1025                 r = process_flush_mesg(argc, argv, wc);
1026         else if (!strcasecmp(argv[0], "flush_on_suspend"))
1027                 r = process_flush_on_suspend_mesg(argc, argv, wc);
1028         else
1029                 DMERR("unrecognised message received: %s", argv[0]);
1030
1031         return r;
1032 }
1033
1034 static void bio_copy_block(struct dm_writecache *wc, struct bio *bio, void *data)
1035 {
1036         void *buf;
1037         unsigned long flags;
1038         unsigned size;
1039         int rw = bio_data_dir(bio);
1040         unsigned remaining_size = wc->block_size;
1041
1042         do {
1043                 struct bio_vec bv = bio_iter_iovec(bio, bio->bi_iter);
1044                 buf = bvec_kmap_irq(&bv, &flags);
1045                 size = bv.bv_len;
1046                 if (unlikely(size > remaining_size))
1047                         size = remaining_size;
1048
1049                 if (rw == READ) {
1050                         int r;
1051                         r = memcpy_mcsafe(buf, data, size);
1052                         flush_dcache_page(bio_page(bio));
1053                         if (unlikely(r)) {
1054                                 writecache_error(wc, r, "hardware memory error when reading data: %d", r);
1055                                 bio->bi_status = BLK_STS_IOERR;
1056                         }
1057                 } else {
1058                         flush_dcache_page(bio_page(bio));
1059                         memcpy_flushcache(data, buf, size);
1060                 }
1061
1062                 bvec_kunmap_irq(buf, &flags);
1063
1064                 data = (char *)data + size;
1065                 remaining_size -= size;
1066                 bio_advance(bio, size);
1067         } while (unlikely(remaining_size));
1068 }
1069
1070 static int writecache_flush_thread(void *data)
1071 {
1072         struct dm_writecache *wc = data;
1073
1074         while (1) {
1075                 struct bio *bio;
1076
1077                 wc_lock(wc);
1078                 bio = bio_list_pop(&wc->flush_list);
1079                 if (!bio) {
1080                         set_current_state(TASK_INTERRUPTIBLE);
1081                         wc_unlock(wc);
1082
1083                         if (unlikely(kthread_should_stop())) {
1084                                 set_current_state(TASK_RUNNING);
1085                                 break;
1086                         }
1087
1088                         schedule();
1089                         continue;
1090                 }
1091
1092                 if (bio_op(bio) == REQ_OP_DISCARD) {
1093                         writecache_discard(wc, bio->bi_iter.bi_sector,
1094                                            bio_end_sector(bio));
1095                         wc_unlock(wc);
1096                         bio_set_dev(bio, wc->dev->bdev);
1097                         generic_make_request(bio);
1098                 } else {
1099                         writecache_flush(wc);
1100                         wc_unlock(wc);
1101                         if (writecache_has_error(wc))
1102                                 bio->bi_status = BLK_STS_IOERR;
1103                         bio_endio(bio);
1104                 }
1105         }
1106
1107         return 0;
1108 }
1109
1110 static void writecache_offload_bio(struct dm_writecache *wc, struct bio *bio)
1111 {
1112         if (bio_list_empty(&wc->flush_list))
1113                 wake_up_process(wc->flush_thread);
1114         bio_list_add(&wc->flush_list, bio);
1115 }
1116
1117 static int writecache_map(struct dm_target *ti, struct bio *bio)
1118 {
1119         struct wc_entry *e;
1120         struct dm_writecache *wc = ti->private;
1121
1122         bio->bi_private = NULL;
1123
1124         wc_lock(wc);
1125
1126         if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1127                 if (writecache_has_error(wc))
1128                         goto unlock_error;
1129                 if (WC_MODE_PMEM(wc)) {
1130                         writecache_flush(wc);
1131                         if (writecache_has_error(wc))
1132                                 goto unlock_error;
1133                         goto unlock_submit;
1134                 } else {
1135                         writecache_offload_bio(wc, bio);
1136                         goto unlock_return;
1137                 }
1138         }
1139
1140         bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
1141
1142         if (unlikely((((unsigned)bio->bi_iter.bi_sector | bio_sectors(bio)) &
1143                                 (wc->block_size / 512 - 1)) != 0)) {
1144                 DMERR("I/O is not aligned, sector %llu, size %u, block size %u",
1145                       (unsigned long long)bio->bi_iter.bi_sector,
1146                       bio->bi_iter.bi_size, wc->block_size);
1147                 goto unlock_error;
1148         }
1149
1150         if (unlikely(bio_op(bio) == REQ_OP_DISCARD)) {
1151                 if (writecache_has_error(wc))
1152                         goto unlock_error;
1153                 if (WC_MODE_PMEM(wc)) {
1154                         writecache_discard(wc, bio->bi_iter.bi_sector, bio_end_sector(bio));
1155                         goto unlock_remap_origin;
1156                 } else {
1157                         writecache_offload_bio(wc, bio);
1158                         goto unlock_return;
1159                 }
1160         }
1161
1162         if (bio_data_dir(bio) == READ) {
1163 read_next_block:
1164                 e = writecache_find_entry(wc, bio->bi_iter.bi_sector, WFE_RETURN_FOLLOWING);
1165                 if (e && read_original_sector(wc, e) == bio->bi_iter.bi_sector) {
1166                         if (WC_MODE_PMEM(wc)) {
1167                                 bio_copy_block(wc, bio, memory_data(wc, e));
1168                                 if (bio->bi_iter.bi_size)
1169                                         goto read_next_block;
1170                                 goto unlock_submit;
1171                         } else {
1172                                 dm_accept_partial_bio(bio, wc->block_size >> SECTOR_SHIFT);
1173                                 bio_set_dev(bio, wc->ssd_dev->bdev);
1174                                 bio->bi_iter.bi_sector = cache_sector(wc, e);
1175                                 if (!writecache_entry_is_committed(wc, e))
1176                                         writecache_wait_for_ios(wc, WRITE);
1177                                 goto unlock_remap;
1178                         }
1179                 } else {
1180                         if (e) {
1181                                 sector_t next_boundary =
1182                                         read_original_sector(wc, e) - bio->bi_iter.bi_sector;
1183                                 if (next_boundary < bio->bi_iter.bi_size >> SECTOR_SHIFT) {
1184                                         dm_accept_partial_bio(bio, next_boundary);
1185                                 }
1186                         }
1187                         goto unlock_remap_origin;
1188                 }
1189         } else {
1190                 do {
1191                         if (writecache_has_error(wc))
1192                                 goto unlock_error;
1193                         e = writecache_find_entry(wc, bio->bi_iter.bi_sector, 0);
1194                         if (e) {
1195                                 if (!writecache_entry_is_committed(wc, e))
1196                                         goto bio_copy;
1197                                 if (!WC_MODE_PMEM(wc) && !e->write_in_progress) {
1198                                         wc->overwrote_committed = true;
1199                                         goto bio_copy;
1200                                 }
1201                         }
1202                         e = writecache_pop_from_freelist(wc);
1203                         if (unlikely(!e)) {
1204                                 writecache_wait_on_freelist(wc);
1205                                 continue;
1206                         }
1207                         write_original_sector_seq_count(wc, e, bio->bi_iter.bi_sector, wc->seq_count);
1208                         writecache_insert_entry(wc, e);
1209                         wc->uncommitted_blocks++;
1210 bio_copy:
1211                         if (WC_MODE_PMEM(wc)) {
1212                                 bio_copy_block(wc, bio, memory_data(wc, e));
1213                         } else {
1214                                 dm_accept_partial_bio(bio, wc->block_size >> SECTOR_SHIFT);
1215                                 bio_set_dev(bio, wc->ssd_dev->bdev);
1216                                 bio->bi_iter.bi_sector = cache_sector(wc, e);
1217                                 if (unlikely(wc->uncommitted_blocks >= wc->autocommit_blocks)) {
1218                                         wc->uncommitted_blocks = 0;
1219                                         queue_work(wc->writeback_wq, &wc->flush_work);
1220                                 } else {
1221                                         writecache_schedule_autocommit(wc);
1222                                 }
1223                                 goto unlock_remap;
1224                         }
1225                 } while (bio->bi_iter.bi_size);
1226
1227                 if (unlikely(wc->uncommitted_blocks >= wc->autocommit_blocks))
1228                         writecache_flush(wc);
1229                 else
1230                         writecache_schedule_autocommit(wc);
1231                 goto unlock_submit;
1232         }
1233
1234 unlock_remap_origin:
1235         bio_set_dev(bio, wc->dev->bdev);
1236         wc_unlock(wc);
1237         return DM_MAPIO_REMAPPED;
1238
1239 unlock_remap:
1240         /* make sure that writecache_end_io decrements bio_in_progress: */
1241         bio->bi_private = (void *)1;
1242         atomic_inc(&wc->bio_in_progress[bio_data_dir(bio)]);
1243         wc_unlock(wc);
1244         return DM_MAPIO_REMAPPED;
1245
1246 unlock_submit:
1247         wc_unlock(wc);
1248         bio_endio(bio);
1249         return DM_MAPIO_SUBMITTED;
1250
1251 unlock_return:
1252         wc_unlock(wc);
1253         return DM_MAPIO_SUBMITTED;
1254
1255 unlock_error:
1256         wc_unlock(wc);
1257         bio_io_error(bio);
1258         return DM_MAPIO_SUBMITTED;
1259 }
1260
1261 static int writecache_end_io(struct dm_target *ti, struct bio *bio, blk_status_t *status)
1262 {
1263         struct dm_writecache *wc = ti->private;
1264
1265         if (bio->bi_private != NULL) {
1266                 int dir = bio_data_dir(bio);
1267                 if (atomic_dec_and_test(&wc->bio_in_progress[dir]))
1268                         if (unlikely(waitqueue_active(&wc->bio_in_progress_wait[dir])))
1269                                 wake_up(&wc->bio_in_progress_wait[dir]);
1270         }
1271         return 0;
1272 }
1273
1274 static int writecache_iterate_devices(struct dm_target *ti,
1275                                       iterate_devices_callout_fn fn, void *data)
1276 {
1277         struct dm_writecache *wc = ti->private;
1278
1279         return fn(ti, wc->dev, 0, ti->len, data);
1280 }
1281
1282 static void writecache_io_hints(struct dm_target *ti, struct queue_limits *limits)
1283 {
1284         struct dm_writecache *wc = ti->private;
1285
1286         if (limits->logical_block_size < wc->block_size)
1287                 limits->logical_block_size = wc->block_size;
1288
1289         if (limits->physical_block_size < wc->block_size)
1290                 limits->physical_block_size = wc->block_size;
1291
1292         if (limits->io_min < wc->block_size)
1293                 limits->io_min = wc->block_size;
1294 }
1295
1296
1297 static void writecache_writeback_endio(struct bio *bio)
1298 {
1299         struct writeback_struct *wb = container_of(bio, struct writeback_struct, bio);
1300         struct dm_writecache *wc = wb->wc;
1301         unsigned long flags;
1302
1303         raw_spin_lock_irqsave(&wc->endio_list_lock, flags);
1304         if (unlikely(list_empty(&wc->endio_list)))
1305                 wake_up_process(wc->endio_thread);
1306         list_add_tail(&wb->endio_entry, &wc->endio_list);
1307         raw_spin_unlock_irqrestore(&wc->endio_list_lock, flags);
1308 }
1309
1310 static void writecache_copy_endio(int read_err, unsigned long write_err, void *ptr)
1311 {
1312         struct copy_struct *c = ptr;
1313         struct dm_writecache *wc = c->wc;
1314
1315         c->error = likely(!(read_err | write_err)) ? 0 : -EIO;
1316
1317         raw_spin_lock_irq(&wc->endio_list_lock);
1318         if (unlikely(list_empty(&wc->endio_list)))
1319                 wake_up_process(wc->endio_thread);
1320         list_add_tail(&c->endio_entry, &wc->endio_list);
1321         raw_spin_unlock_irq(&wc->endio_list_lock);
1322 }
1323
1324 static void __writecache_endio_pmem(struct dm_writecache *wc, struct list_head *list)
1325 {
1326         unsigned i;
1327         struct writeback_struct *wb;
1328         struct wc_entry *e;
1329         unsigned long n_walked = 0;
1330
1331         do {
1332                 wb = list_entry(list->next, struct writeback_struct, endio_entry);
1333                 list_del(&wb->endio_entry);
1334
1335                 if (unlikely(wb->bio.bi_status != BLK_STS_OK))
1336                         writecache_error(wc, blk_status_to_errno(wb->bio.bi_status),
1337                                         "write error %d", wb->bio.bi_status);
1338                 i = 0;
1339                 do {
1340                         e = wb->wc_list[i];
1341                         BUG_ON(!e->write_in_progress);
1342                         e->write_in_progress = false;
1343                         INIT_LIST_HEAD(&e->lru);
1344                         if (!writecache_has_error(wc))
1345                                 writecache_free_entry(wc, e);
1346                         BUG_ON(!wc->writeback_size);
1347                         wc->writeback_size--;
1348                         n_walked++;
1349                         if (unlikely(n_walked >= ENDIO_LATENCY)) {
1350                                 writecache_commit_flushed(wc);
1351                                 wc_unlock(wc);
1352                                 wc_lock(wc);
1353                                 n_walked = 0;
1354                         }
1355                 } while (++i < wb->wc_list_n);
1356
1357                 if (wb->wc_list != wb->wc_list_inline)
1358                         kfree(wb->wc_list);
1359                 bio_put(&wb->bio);
1360         } while (!list_empty(list));
1361 }
1362
1363 static void __writecache_endio_ssd(struct dm_writecache *wc, struct list_head *list)
1364 {
1365         struct copy_struct *c;
1366         struct wc_entry *e;
1367
1368         do {
1369                 c = list_entry(list->next, struct copy_struct, endio_entry);
1370                 list_del(&c->endio_entry);
1371
1372                 if (unlikely(c->error))
1373                         writecache_error(wc, c->error, "copy error");
1374
1375                 e = c->e;
1376                 do {
1377                         BUG_ON(!e->write_in_progress);
1378                         e->write_in_progress = false;
1379                         INIT_LIST_HEAD(&e->lru);
1380                         if (!writecache_has_error(wc))
1381                                 writecache_free_entry(wc, e);
1382
1383                         BUG_ON(!wc->writeback_size);
1384                         wc->writeback_size--;
1385                         e++;
1386                 } while (--c->n_entries);
1387                 mempool_free(c, &wc->copy_pool);
1388         } while (!list_empty(list));
1389 }
1390
1391 static int writecache_endio_thread(void *data)
1392 {
1393         struct dm_writecache *wc = data;
1394
1395         while (1) {
1396                 struct list_head list;
1397
1398                 raw_spin_lock_irq(&wc->endio_list_lock);
1399                 if (!list_empty(&wc->endio_list))
1400                         goto pop_from_list;
1401                 set_current_state(TASK_INTERRUPTIBLE);
1402                 raw_spin_unlock_irq(&wc->endio_list_lock);
1403
1404                 if (unlikely(kthread_should_stop())) {
1405                         set_current_state(TASK_RUNNING);
1406                         break;
1407                 }
1408
1409                 schedule();
1410
1411                 continue;
1412
1413 pop_from_list:
1414                 list = wc->endio_list;
1415                 list.next->prev = list.prev->next = &list;
1416                 INIT_LIST_HEAD(&wc->endio_list);
1417                 raw_spin_unlock_irq(&wc->endio_list_lock);
1418
1419                 if (!WC_MODE_FUA(wc))
1420                         writecache_disk_flush(wc, wc->dev);
1421
1422                 wc_lock(wc);
1423
1424                 if (WC_MODE_PMEM(wc)) {
1425                         __writecache_endio_pmem(wc, &list);
1426                 } else {
1427                         __writecache_endio_ssd(wc, &list);
1428                         writecache_wait_for_ios(wc, READ);
1429                 }
1430
1431                 writecache_commit_flushed(wc);
1432
1433                 wc_unlock(wc);
1434         }
1435
1436         return 0;
1437 }
1438
1439 static bool wc_add_block(struct writeback_struct *wb, struct wc_entry *e, gfp_t gfp)
1440 {
1441         struct dm_writecache *wc = wb->wc;
1442         unsigned block_size = wc->block_size;
1443         void *address = memory_data(wc, e);
1444
1445         persistent_memory_flush_cache(address, block_size);
1446         return bio_add_page(&wb->bio, persistent_memory_page(address),
1447                             block_size, persistent_memory_page_offset(address)) != 0;
1448 }
1449
1450 struct writeback_list {
1451         struct list_head list;
1452         size_t size;
1453 };
1454
1455 static void __writeback_throttle(struct dm_writecache *wc, struct writeback_list *wbl)
1456 {
1457         if (unlikely(wc->max_writeback_jobs)) {
1458                 if (READ_ONCE(wc->writeback_size) - wbl->size >= wc->max_writeback_jobs) {
1459                         wc_lock(wc);
1460                         while (wc->writeback_size - wbl->size >= wc->max_writeback_jobs)
1461                                 writecache_wait_on_freelist(wc);
1462                         wc_unlock(wc);
1463                 }
1464         }
1465         cond_resched();
1466 }
1467
1468 static void __writecache_writeback_pmem(struct dm_writecache *wc, struct writeback_list *wbl)
1469 {
1470         struct wc_entry *e, *f;
1471         struct bio *bio;
1472         struct writeback_struct *wb;
1473         unsigned max_pages;
1474
1475         while (wbl->size) {
1476                 wbl->size--;
1477                 e = container_of(wbl->list.prev, struct wc_entry, lru);
1478                 list_del(&e->lru);
1479
1480                 max_pages = e->wc_list_contiguous;
1481
1482                 bio = bio_alloc_bioset(GFP_NOIO, max_pages, &wc->bio_set);
1483                 wb = container_of(bio, struct writeback_struct, bio);
1484                 wb->wc = wc;
1485                 wb->bio.bi_end_io = writecache_writeback_endio;
1486                 bio_set_dev(&wb->bio, wc->dev->bdev);
1487                 wb->bio.bi_iter.bi_sector = read_original_sector(wc, e);
1488                 wb->page_offset = PAGE_SIZE;
1489                 if (max_pages <= WB_LIST_INLINE ||
1490                     unlikely(!(wb->wc_list = kmalloc_array(max_pages, sizeof(struct wc_entry *),
1491                                                            GFP_NOIO | __GFP_NORETRY |
1492                                                            __GFP_NOMEMALLOC | __GFP_NOWARN)))) {
1493                         wb->wc_list = wb->wc_list_inline;
1494                         max_pages = WB_LIST_INLINE;
1495                 }
1496
1497                 BUG_ON(!wc_add_block(wb, e, GFP_NOIO));
1498
1499                 wb->wc_list[0] = e;
1500                 wb->wc_list_n = 1;
1501
1502                 while (wbl->size && wb->wc_list_n < max_pages) {
1503                         f = container_of(wbl->list.prev, struct wc_entry, lru);
1504                         if (read_original_sector(wc, f) !=
1505                             read_original_sector(wc, e) + (wc->block_size >> SECTOR_SHIFT))
1506                                 break;
1507                         if (!wc_add_block(wb, f, GFP_NOWAIT | __GFP_NOWARN))
1508                                 break;
1509                         wbl->size--;
1510                         list_del(&f->lru);
1511                         wb->wc_list[wb->wc_list_n++] = f;
1512                         e = f;
1513                 }
1514                 bio_set_op_attrs(&wb->bio, REQ_OP_WRITE, WC_MODE_FUA(wc) * REQ_FUA);
1515                 if (writecache_has_error(wc)) {
1516                         bio->bi_status = BLK_STS_IOERR;
1517                         bio_endio(&wb->bio);
1518                 } else {
1519                         submit_bio(&wb->bio);
1520                 }
1521
1522                 __writeback_throttle(wc, wbl);
1523         }
1524 }
1525
1526 static void __writecache_writeback_ssd(struct dm_writecache *wc, struct writeback_list *wbl)
1527 {
1528         struct wc_entry *e, *f;
1529         struct dm_io_region from, to;
1530         struct copy_struct *c;
1531
1532         while (wbl->size) {
1533                 unsigned n_sectors;
1534
1535                 wbl->size--;
1536                 e = container_of(wbl->list.prev, struct wc_entry, lru);
1537                 list_del(&e->lru);
1538
1539                 n_sectors = e->wc_list_contiguous << (wc->block_size_bits - SECTOR_SHIFT);
1540
1541                 from.bdev = wc->ssd_dev->bdev;
1542                 from.sector = cache_sector(wc, e);
1543                 from.count = n_sectors;
1544                 to.bdev = wc->dev->bdev;
1545                 to.sector = read_original_sector(wc, e);
1546                 to.count = n_sectors;
1547
1548                 c = mempool_alloc(&wc->copy_pool, GFP_NOIO);
1549                 c->wc = wc;
1550                 c->e = e;
1551                 c->n_entries = e->wc_list_contiguous;
1552
1553                 while ((n_sectors -= wc->block_size >> SECTOR_SHIFT)) {
1554                         wbl->size--;
1555                         f = container_of(wbl->list.prev, struct wc_entry, lru);
1556                         BUG_ON(f != e + 1);
1557                         list_del(&f->lru);
1558                         e = f;
1559                 }
1560
1561                 dm_kcopyd_copy(wc->dm_kcopyd, &from, 1, &to, 0, writecache_copy_endio, c);
1562
1563                 __writeback_throttle(wc, wbl);
1564         }
1565 }
1566
1567 static void writecache_writeback(struct work_struct *work)
1568 {
1569         struct dm_writecache *wc = container_of(work, struct dm_writecache, writeback_work);
1570         struct blk_plug plug;
1571         struct wc_entry *e, *f, *g;
1572         struct rb_node *node, *next_node;
1573         struct list_head skipped;
1574         struct writeback_list wbl;
1575         unsigned long n_walked;
1576
1577         wc_lock(wc);
1578 restart:
1579         if (writecache_has_error(wc)) {
1580                 wc_unlock(wc);
1581                 return;
1582         }
1583
1584         if (unlikely(wc->writeback_all)) {
1585                 if (writecache_wait_for_writeback(wc))
1586                         goto restart;
1587         }
1588
1589         if (wc->overwrote_committed) {
1590                 writecache_wait_for_ios(wc, WRITE);
1591         }
1592
1593         n_walked = 0;
1594         INIT_LIST_HEAD(&skipped);
1595         INIT_LIST_HEAD(&wbl.list);
1596         wbl.size = 0;
1597         while (!list_empty(&wc->lru) &&
1598                (wc->writeback_all ||
1599                 wc->freelist_size + wc->writeback_size <= wc->freelist_low_watermark)) {
1600
1601                 n_walked++;
1602                 if (unlikely(n_walked > WRITEBACK_LATENCY) &&
1603                     likely(!wc->writeback_all) && likely(!dm_suspended(wc->ti))) {
1604                         queue_work(wc->writeback_wq, &wc->writeback_work);
1605                         break;
1606                 }
1607
1608                 e = container_of(wc->lru.prev, struct wc_entry, lru);
1609                 BUG_ON(e->write_in_progress);
1610                 if (unlikely(!writecache_entry_is_committed(wc, e))) {
1611                         writecache_flush(wc);
1612                 }
1613                 node = rb_prev(&e->rb_node);
1614                 if (node) {
1615                         f = container_of(node, struct wc_entry, rb_node);
1616                         if (unlikely(read_original_sector(wc, f) ==
1617                                      read_original_sector(wc, e))) {
1618                                 BUG_ON(!f->write_in_progress);
1619                                 list_del(&e->lru);
1620                                 list_add(&e->lru, &skipped);
1621                                 cond_resched();
1622                                 continue;
1623                         }
1624                 }
1625                 wc->writeback_size++;
1626                 list_del(&e->lru);
1627                 list_add(&e->lru, &wbl.list);
1628                 wbl.size++;
1629                 e->write_in_progress = true;
1630                 e->wc_list_contiguous = 1;
1631
1632                 f = e;
1633
1634                 while (1) {
1635                         next_node = rb_next(&f->rb_node);
1636                         if (unlikely(!next_node))
1637                                 break;
1638                         g = container_of(next_node, struct wc_entry, rb_node);
1639                         if (read_original_sector(wc, g) ==
1640                             read_original_sector(wc, f)) {
1641                                 f = g;
1642                                 continue;
1643                         }
1644                         if (read_original_sector(wc, g) !=
1645                             read_original_sector(wc, f) + (wc->block_size >> SECTOR_SHIFT))
1646                                 break;
1647                         if (unlikely(g->write_in_progress))
1648                                 break;
1649                         if (unlikely(!writecache_entry_is_committed(wc, g)))
1650                                 break;
1651
1652                         if (!WC_MODE_PMEM(wc)) {
1653                                 if (g != f + 1)
1654                                         break;
1655                         }
1656
1657                         n_walked++;
1658                         //if (unlikely(n_walked > WRITEBACK_LATENCY) && likely(!wc->writeback_all))
1659                         //      break;
1660
1661                         wc->writeback_size++;
1662                         list_del(&g->lru);
1663                         list_add(&g->lru, &wbl.list);
1664                         wbl.size++;
1665                         g->write_in_progress = true;
1666                         g->wc_list_contiguous = BIO_MAX_PAGES;
1667                         f = g;
1668                         e->wc_list_contiguous++;
1669                         if (unlikely(e->wc_list_contiguous == BIO_MAX_PAGES))
1670                                 break;
1671                 }
1672                 cond_resched();
1673         }
1674
1675         if (!list_empty(&skipped)) {
1676                 list_splice_tail(&skipped, &wc->lru);
1677                 /*
1678                  * If we didn't do any progress, we must wait until some
1679                  * writeback finishes to avoid burning CPU in a loop
1680                  */
1681                 if (unlikely(!wbl.size))
1682                         writecache_wait_for_writeback(wc);
1683         }
1684
1685         wc_unlock(wc);
1686
1687         blk_start_plug(&plug);
1688
1689         if (WC_MODE_PMEM(wc))
1690                 __writecache_writeback_pmem(wc, &wbl);
1691         else
1692                 __writecache_writeback_ssd(wc, &wbl);
1693
1694         blk_finish_plug(&plug);
1695
1696         if (unlikely(wc->writeback_all)) {
1697                 wc_lock(wc);
1698                 while (writecache_wait_for_writeback(wc));
1699                 wc_unlock(wc);
1700         }
1701 }
1702
1703 static int calculate_memory_size(uint64_t device_size, unsigned block_size,
1704                                  size_t *n_blocks_p, size_t *n_metadata_blocks_p)
1705 {
1706         uint64_t n_blocks, offset;
1707         struct wc_entry e;
1708
1709         n_blocks = device_size;
1710         do_div(n_blocks, block_size + sizeof(struct wc_memory_entry));
1711
1712         while (1) {
1713                 if (!n_blocks)
1714                         return -ENOSPC;
1715                 /* Verify the following entries[n_blocks] won't overflow */
1716                 if (n_blocks >= ((size_t)-sizeof(struct wc_memory_superblock) /
1717                                  sizeof(struct wc_memory_entry)))
1718                         return -EFBIG;
1719                 offset = offsetof(struct wc_memory_superblock, entries[n_blocks]);
1720                 offset = (offset + block_size - 1) & ~(uint64_t)(block_size - 1);
1721                 if (offset + n_blocks * block_size <= device_size)
1722                         break;
1723                 n_blocks--;
1724         }
1725
1726         /* check if the bit field overflows */
1727         e.index = n_blocks;
1728         if (e.index != n_blocks)
1729                 return -EFBIG;
1730
1731         if (n_blocks_p)
1732                 *n_blocks_p = n_blocks;
1733         if (n_metadata_blocks_p)
1734                 *n_metadata_blocks_p = offset >> __ffs(block_size);
1735         return 0;
1736 }
1737
1738 static int init_memory(struct dm_writecache *wc)
1739 {
1740         size_t b;
1741         int r;
1742
1743         r = calculate_memory_size(wc->memory_map_size, wc->block_size, &wc->n_blocks, NULL);
1744         if (r)
1745                 return r;
1746
1747         r = writecache_alloc_entries(wc);
1748         if (r)
1749                 return r;
1750
1751         for (b = 0; b < ARRAY_SIZE(sb(wc)->padding); b++)
1752                 pmem_assign(sb(wc)->padding[b], cpu_to_le64(0));
1753         pmem_assign(sb(wc)->version, cpu_to_le32(MEMORY_SUPERBLOCK_VERSION));
1754         pmem_assign(sb(wc)->block_size, cpu_to_le32(wc->block_size));
1755         pmem_assign(sb(wc)->n_blocks, cpu_to_le64(wc->n_blocks));
1756         pmem_assign(sb(wc)->seq_count, cpu_to_le64(0));
1757
1758         for (b = 0; b < wc->n_blocks; b++)
1759                 write_original_sector_seq_count(wc, &wc->entries[b], -1, -1);
1760
1761         writecache_flush_all_metadata(wc);
1762         writecache_commit_flushed(wc);
1763         pmem_assign(sb(wc)->magic, cpu_to_le32(MEMORY_SUPERBLOCK_MAGIC));
1764         writecache_flush_region(wc, &sb(wc)->magic, sizeof sb(wc)->magic);
1765         writecache_commit_flushed(wc);
1766
1767         return 0;
1768 }
1769
1770 static void writecache_dtr(struct dm_target *ti)
1771 {
1772         struct dm_writecache *wc = ti->private;
1773
1774         if (!wc)
1775                 return;
1776
1777         if (wc->endio_thread)
1778                 kthread_stop(wc->endio_thread);
1779
1780         if (wc->flush_thread)
1781                 kthread_stop(wc->flush_thread);
1782
1783         bioset_exit(&wc->bio_set);
1784
1785         mempool_exit(&wc->copy_pool);
1786
1787         if (wc->writeback_wq)
1788                 destroy_workqueue(wc->writeback_wq);
1789
1790         if (wc->dev)
1791                 dm_put_device(ti, wc->dev);
1792
1793         if (wc->ssd_dev)
1794                 dm_put_device(ti, wc->ssd_dev);
1795
1796         if (wc->entries)
1797                 vfree(wc->entries);
1798
1799         if (wc->memory_map) {
1800                 if (WC_MODE_PMEM(wc))
1801                         persistent_memory_release(wc);
1802                 else
1803                         vfree(wc->memory_map);
1804         }
1805
1806         if (wc->dm_kcopyd)
1807                 dm_kcopyd_client_destroy(wc->dm_kcopyd);
1808
1809         if (wc->dm_io)
1810                 dm_io_client_destroy(wc->dm_io);
1811
1812         if (wc->dirty_bitmap)
1813                 vfree(wc->dirty_bitmap);
1814
1815         kfree(wc);
1816 }
1817
1818 static int writecache_ctr(struct dm_target *ti, unsigned argc, char **argv)
1819 {
1820         struct dm_writecache *wc;
1821         struct dm_arg_set as;
1822         const char *string;
1823         unsigned opt_params;
1824         size_t offset, data_size;
1825         int i, r;
1826         char dummy;
1827         int high_wm_percent = HIGH_WATERMARK;
1828         int low_wm_percent = LOW_WATERMARK;
1829         uint64_t x;
1830         struct wc_memory_superblock s;
1831
1832         static struct dm_arg _args[] = {
1833                 {0, 10, "Invalid number of feature args"},
1834         };
1835
1836         as.argc = argc;
1837         as.argv = argv;
1838
1839         wc = kzalloc(sizeof(struct dm_writecache), GFP_KERNEL);
1840         if (!wc) {
1841                 ti->error = "Cannot allocate writecache structure";
1842                 r = -ENOMEM;
1843                 goto bad;
1844         }
1845         ti->private = wc;
1846         wc->ti = ti;
1847
1848         mutex_init(&wc->lock);
1849         writecache_poison_lists(wc);
1850         init_waitqueue_head(&wc->freelist_wait);
1851         timer_setup(&wc->autocommit_timer, writecache_autocommit_timer, 0);
1852
1853         for (i = 0; i < 2; i++) {
1854                 atomic_set(&wc->bio_in_progress[i], 0);
1855                 init_waitqueue_head(&wc->bio_in_progress_wait[i]);
1856         }
1857
1858         wc->dm_io = dm_io_client_create();
1859         if (IS_ERR(wc->dm_io)) {
1860                 r = PTR_ERR(wc->dm_io);
1861                 ti->error = "Unable to allocate dm-io client";
1862                 wc->dm_io = NULL;
1863                 goto bad;
1864         }
1865
1866         wc->writeback_wq = alloc_workqueue("writecache-writeabck", WQ_MEM_RECLAIM, 1);
1867         if (!wc->writeback_wq) {
1868                 r = -ENOMEM;
1869                 ti->error = "Could not allocate writeback workqueue";
1870                 goto bad;
1871         }
1872         INIT_WORK(&wc->writeback_work, writecache_writeback);
1873         INIT_WORK(&wc->flush_work, writecache_flush_work);
1874
1875         raw_spin_lock_init(&wc->endio_list_lock);
1876         INIT_LIST_HEAD(&wc->endio_list);
1877         wc->endio_thread = kthread_create(writecache_endio_thread, wc, "writecache_endio");
1878         if (IS_ERR(wc->endio_thread)) {
1879                 r = PTR_ERR(wc->endio_thread);
1880                 wc->endio_thread = NULL;
1881                 ti->error = "Couldn't spawn endio thread";
1882                 goto bad;
1883         }
1884         wake_up_process(wc->endio_thread);
1885
1886         /*
1887          * Parse the mode (pmem or ssd)
1888          */
1889         string = dm_shift_arg(&as);
1890         if (!string)
1891                 goto bad_arguments;
1892
1893         if (!strcasecmp(string, "s")) {
1894                 wc->pmem_mode = false;
1895         } else if (!strcasecmp(string, "p")) {
1896 #ifdef DM_WRITECACHE_HAS_PMEM
1897                 wc->pmem_mode = true;
1898                 wc->writeback_fua = true;
1899 #else
1900                 /*
1901                  * If the architecture doesn't support persistent memory or
1902                  * the kernel doesn't support any DAX drivers, this driver can
1903                  * only be used in SSD-only mode.
1904                  */
1905                 r = -EOPNOTSUPP;
1906                 ti->error = "Persistent memory or DAX not supported on this system";
1907                 goto bad;
1908 #endif
1909         } else {
1910                 goto bad_arguments;
1911         }
1912
1913         if (WC_MODE_PMEM(wc)) {
1914                 r = bioset_init(&wc->bio_set, BIO_POOL_SIZE,
1915                                 offsetof(struct writeback_struct, bio),
1916                                 BIOSET_NEED_BVECS);
1917                 if (r) {
1918                         ti->error = "Could not allocate bio set";
1919                         goto bad;
1920                 }
1921         } else {
1922                 r = mempool_init_kmalloc_pool(&wc->copy_pool, 1, sizeof(struct copy_struct));
1923                 if (r) {
1924                         ti->error = "Could not allocate mempool";
1925                         goto bad;
1926                 }
1927         }
1928
1929         /*
1930          * Parse the origin data device
1931          */
1932         string = dm_shift_arg(&as);
1933         if (!string)
1934                 goto bad_arguments;
1935         r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->dev);
1936         if (r) {
1937                 ti->error = "Origin data device lookup failed";
1938                 goto bad;
1939         }
1940
1941         /*
1942          * Parse cache data device (be it pmem or ssd)
1943          */
1944         string = dm_shift_arg(&as);
1945         if (!string)
1946                 goto bad_arguments;
1947
1948         r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->ssd_dev);
1949         if (r) {
1950                 ti->error = "Cache data device lookup failed";
1951                 goto bad;
1952         }
1953         wc->memory_map_size = i_size_read(wc->ssd_dev->bdev->bd_inode);
1954
1955         /*
1956          * Parse the cache block size
1957          */
1958         string = dm_shift_arg(&as);
1959         if (!string)
1960                 goto bad_arguments;
1961         if (sscanf(string, "%u%c", &wc->block_size, &dummy) != 1 ||
1962             wc->block_size < 512 || wc->block_size > PAGE_SIZE ||
1963             (wc->block_size & (wc->block_size - 1))) {
1964                 r = -EINVAL;
1965                 ti->error = "Invalid block size";
1966                 goto bad;
1967         }
1968         wc->block_size_bits = __ffs(wc->block_size);
1969
1970         wc->max_writeback_jobs = MAX_WRITEBACK_JOBS;
1971         wc->autocommit_blocks = !WC_MODE_PMEM(wc) ? AUTOCOMMIT_BLOCKS_SSD : AUTOCOMMIT_BLOCKS_PMEM;
1972         wc->autocommit_jiffies = msecs_to_jiffies(AUTOCOMMIT_MSEC);
1973
1974         /*
1975          * Parse optional arguments
1976          */
1977         r = dm_read_arg_group(_args, &as, &opt_params, &ti->error);
1978         if (r)
1979                 goto bad;
1980
1981         while (opt_params) {
1982                 string = dm_shift_arg(&as), opt_params--;
1983                 if (!strcasecmp(string, "start_sector") && opt_params >= 1) {
1984                         unsigned long long start_sector;
1985                         string = dm_shift_arg(&as), opt_params--;
1986                         if (sscanf(string, "%llu%c", &start_sector, &dummy) != 1)
1987                                 goto invalid_optional;
1988                         wc->start_sector = start_sector;
1989                         if (wc->start_sector != start_sector ||
1990                             wc->start_sector >= wc->memory_map_size >> SECTOR_SHIFT)
1991                                 goto invalid_optional;
1992                 } else if (!strcasecmp(string, "high_watermark") && opt_params >= 1) {
1993                         string = dm_shift_arg(&as), opt_params--;
1994                         if (sscanf(string, "%d%c", &high_wm_percent, &dummy) != 1)
1995                                 goto invalid_optional;
1996                         if (high_wm_percent < 0 || high_wm_percent > 100)
1997                                 goto invalid_optional;
1998                         wc->high_wm_percent_set = true;
1999                 } else if (!strcasecmp(string, "low_watermark") && opt_params >= 1) {
2000                         string = dm_shift_arg(&as), opt_params--;
2001                         if (sscanf(string, "%d%c", &low_wm_percent, &dummy) != 1)
2002                                 goto invalid_optional;
2003                         if (low_wm_percent < 0 || low_wm_percent > 100)
2004                                 goto invalid_optional;
2005                         wc->low_wm_percent_set = true;
2006                 } else if (!strcasecmp(string, "writeback_jobs") && opt_params >= 1) {
2007                         string = dm_shift_arg(&as), opt_params--;
2008                         if (sscanf(string, "%u%c", &wc->max_writeback_jobs, &dummy) != 1)
2009                                 goto invalid_optional;
2010                         wc->max_writeback_jobs_set = true;
2011                 } else if (!strcasecmp(string, "autocommit_blocks") && opt_params >= 1) {
2012                         string = dm_shift_arg(&as), opt_params--;
2013                         if (sscanf(string, "%u%c", &wc->autocommit_blocks, &dummy) != 1)
2014                                 goto invalid_optional;
2015                         wc->autocommit_blocks_set = true;
2016                 } else if (!strcasecmp(string, "autocommit_time") && opt_params >= 1) {
2017                         unsigned autocommit_msecs;
2018                         string = dm_shift_arg(&as), opt_params--;
2019                         if (sscanf(string, "%u%c", &autocommit_msecs, &dummy) != 1)
2020                                 goto invalid_optional;
2021                         if (autocommit_msecs > 3600000)
2022                                 goto invalid_optional;
2023                         wc->autocommit_jiffies = msecs_to_jiffies(autocommit_msecs);
2024                         wc->autocommit_time_set = true;
2025                 } else if (!strcasecmp(string, "fua")) {
2026                         if (WC_MODE_PMEM(wc)) {
2027                                 wc->writeback_fua = true;
2028                                 wc->writeback_fua_set = true;
2029                         } else goto invalid_optional;
2030                 } else if (!strcasecmp(string, "nofua")) {
2031                         if (WC_MODE_PMEM(wc)) {
2032                                 wc->writeback_fua = false;
2033                                 wc->writeback_fua_set = true;
2034                         } else goto invalid_optional;
2035                 } else {
2036 invalid_optional:
2037                         r = -EINVAL;
2038                         ti->error = "Invalid optional argument";
2039                         goto bad;
2040                 }
2041         }
2042
2043         if (high_wm_percent < low_wm_percent) {
2044                 r = -EINVAL;
2045                 ti->error = "High watermark must be greater than or equal to low watermark";
2046                 goto bad;
2047         }
2048
2049         if (WC_MODE_PMEM(wc)) {
2050                 r = persistent_memory_claim(wc);
2051                 if (r) {
2052                         ti->error = "Unable to map persistent memory for cache";
2053                         goto bad;
2054                 }
2055         } else {
2056                 struct dm_io_region region;
2057                 struct dm_io_request req;
2058                 size_t n_blocks, n_metadata_blocks;
2059                 uint64_t n_bitmap_bits;
2060
2061                 wc->memory_map_size -= (uint64_t)wc->start_sector << SECTOR_SHIFT;
2062
2063                 bio_list_init(&wc->flush_list);
2064                 wc->flush_thread = kthread_create(writecache_flush_thread, wc, "dm_writecache_flush");
2065                 if (IS_ERR(wc->flush_thread)) {
2066                         r = PTR_ERR(wc->flush_thread);
2067                         wc->flush_thread = NULL;
2068                         ti->error = "Couldn't spawn endio thread";
2069                         goto bad;
2070                 }
2071                 wake_up_process(wc->flush_thread);
2072
2073                 r = calculate_memory_size(wc->memory_map_size, wc->block_size,
2074                                           &n_blocks, &n_metadata_blocks);
2075                 if (r) {
2076                         ti->error = "Invalid device size";
2077                         goto bad;
2078                 }
2079
2080                 n_bitmap_bits = (((uint64_t)n_metadata_blocks << wc->block_size_bits) +
2081                                  BITMAP_GRANULARITY - 1) / BITMAP_GRANULARITY;
2082                 /* this is limitation of test_bit functions */
2083                 if (n_bitmap_bits > 1U << 31) {
2084                         r = -EFBIG;
2085                         ti->error = "Invalid device size";
2086                         goto bad;
2087                 }
2088
2089                 wc->memory_map = vmalloc(n_metadata_blocks << wc->block_size_bits);
2090                 if (!wc->memory_map) {
2091                         r = -ENOMEM;
2092                         ti->error = "Unable to allocate memory for metadata";
2093                         goto bad;
2094                 }
2095
2096                 wc->dm_kcopyd = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2097                 if (IS_ERR(wc->dm_kcopyd)) {
2098                         r = PTR_ERR(wc->dm_kcopyd);
2099                         ti->error = "Unable to allocate dm-kcopyd client";
2100                         wc->dm_kcopyd = NULL;
2101                         goto bad;
2102                 }
2103
2104                 wc->metadata_sectors = n_metadata_blocks << (wc->block_size_bits - SECTOR_SHIFT);
2105                 wc->dirty_bitmap_size = (n_bitmap_bits + BITS_PER_LONG - 1) /
2106                         BITS_PER_LONG * sizeof(unsigned long);
2107                 wc->dirty_bitmap = vzalloc(wc->dirty_bitmap_size);
2108                 if (!wc->dirty_bitmap) {
2109                         r = -ENOMEM;
2110                         ti->error = "Unable to allocate dirty bitmap";
2111                         goto bad;
2112                 }
2113
2114                 region.bdev = wc->ssd_dev->bdev;
2115                 region.sector = wc->start_sector;
2116                 region.count = wc->metadata_sectors;
2117                 req.bi_op = REQ_OP_READ;
2118                 req.bi_op_flags = REQ_SYNC;
2119                 req.mem.type = DM_IO_VMA;
2120                 req.mem.ptr.vma = (char *)wc->memory_map;
2121                 req.client = wc->dm_io;
2122                 req.notify.fn = NULL;
2123
2124                 r = dm_io(&req, 1, &region, NULL);
2125                 if (r) {
2126                         ti->error = "Unable to read metadata";
2127                         goto bad;
2128                 }
2129         }
2130
2131         r = memcpy_mcsafe(&s, sb(wc), sizeof(struct wc_memory_superblock));
2132         if (r) {
2133                 ti->error = "Hardware memory error when reading superblock";
2134                 goto bad;
2135         }
2136         if (!le32_to_cpu(s.magic) && !le32_to_cpu(s.version)) {
2137                 r = init_memory(wc);
2138                 if (r) {
2139                         ti->error = "Unable to initialize device";
2140                         goto bad;
2141                 }
2142                 r = memcpy_mcsafe(&s, sb(wc), sizeof(struct wc_memory_superblock));
2143                 if (r) {
2144                         ti->error = "Hardware memory error when reading superblock";
2145                         goto bad;
2146                 }
2147         }
2148
2149         if (le32_to_cpu(s.magic) != MEMORY_SUPERBLOCK_MAGIC) {
2150                 ti->error = "Invalid magic in the superblock";
2151                 r = -EINVAL;
2152                 goto bad;
2153         }
2154
2155         if (le32_to_cpu(s.version) != MEMORY_SUPERBLOCK_VERSION) {
2156                 ti->error = "Invalid version in the superblock";
2157                 r = -EINVAL;
2158                 goto bad;
2159         }
2160
2161         if (le32_to_cpu(s.block_size) != wc->block_size) {
2162                 ti->error = "Block size does not match superblock";
2163                 r = -EINVAL;
2164                 goto bad;
2165         }
2166
2167         wc->n_blocks = le64_to_cpu(s.n_blocks);
2168
2169         offset = wc->n_blocks * sizeof(struct wc_memory_entry);
2170         if (offset / sizeof(struct wc_memory_entry) != le64_to_cpu(sb(wc)->n_blocks)) {
2171 overflow:
2172                 ti->error = "Overflow in size calculation";
2173                 r = -EINVAL;
2174                 goto bad;
2175         }
2176         offset += sizeof(struct wc_memory_superblock);
2177         if (offset < sizeof(struct wc_memory_superblock))
2178                 goto overflow;
2179         offset = (offset + wc->block_size - 1) & ~(size_t)(wc->block_size - 1);
2180         data_size = wc->n_blocks * (size_t)wc->block_size;
2181         if (!offset || (data_size / wc->block_size != wc->n_blocks) ||
2182             (offset + data_size < offset))
2183                 goto overflow;
2184         if (offset + data_size > wc->memory_map_size) {
2185                 ti->error = "Memory area is too small";
2186                 r = -EINVAL;
2187                 goto bad;
2188         }
2189
2190         wc->metadata_sectors = offset >> SECTOR_SHIFT;
2191         wc->block_start = (char *)sb(wc) + offset;
2192
2193         x = (uint64_t)wc->n_blocks * (100 - high_wm_percent);
2194         x += 50;
2195         do_div(x, 100);
2196         wc->freelist_high_watermark = x;
2197         x = (uint64_t)wc->n_blocks * (100 - low_wm_percent);
2198         x += 50;
2199         do_div(x, 100);
2200         wc->freelist_low_watermark = x;
2201
2202         r = writecache_alloc_entries(wc);
2203         if (r) {
2204                 ti->error = "Cannot allocate memory";
2205                 goto bad;
2206         }
2207
2208         ti->num_flush_bios = 1;
2209         ti->flush_supported = true;
2210         ti->num_discard_bios = 1;
2211
2212         if (WC_MODE_PMEM(wc))
2213                 persistent_memory_flush_cache(wc->memory_map, wc->memory_map_size);
2214
2215         return 0;
2216
2217 bad_arguments:
2218         r = -EINVAL;
2219         ti->error = "Bad arguments";
2220 bad:
2221         writecache_dtr(ti);
2222         return r;
2223 }
2224
2225 static void writecache_status(struct dm_target *ti, status_type_t type,
2226                               unsigned status_flags, char *result, unsigned maxlen)
2227 {
2228         struct dm_writecache *wc = ti->private;
2229         unsigned extra_args;
2230         unsigned sz = 0;
2231         uint64_t x;
2232
2233         switch (type) {
2234         case STATUSTYPE_INFO:
2235                 DMEMIT("%ld %llu %llu %llu", writecache_has_error(wc),
2236                        (unsigned long long)wc->n_blocks, (unsigned long long)wc->freelist_size,
2237                        (unsigned long long)wc->writeback_size);
2238                 break;
2239         case STATUSTYPE_TABLE:
2240                 DMEMIT("%c %s %s %u ", WC_MODE_PMEM(wc) ? 'p' : 's',
2241                                 wc->dev->name, wc->ssd_dev->name, wc->block_size);
2242                 extra_args = 0;
2243                 if (wc->high_wm_percent_set)
2244                         extra_args += 2;
2245                 if (wc->low_wm_percent_set)
2246                         extra_args += 2;
2247                 if (wc->max_writeback_jobs_set)
2248                         extra_args += 2;
2249                 if (wc->autocommit_blocks_set)
2250                         extra_args += 2;
2251                 if (wc->autocommit_time_set)
2252                         extra_args += 2;
2253                 if (wc->writeback_fua_set)
2254                         extra_args++;
2255
2256                 DMEMIT("%u", extra_args);
2257                 if (wc->high_wm_percent_set) {
2258                         x = (uint64_t)wc->freelist_high_watermark * 100;
2259                         x += wc->n_blocks / 2;
2260                         do_div(x, (size_t)wc->n_blocks);
2261                         DMEMIT(" high_watermark %u", 100 - (unsigned)x);
2262                 }
2263                 if (wc->low_wm_percent_set) {
2264                         x = (uint64_t)wc->freelist_low_watermark * 100;
2265                         x += wc->n_blocks / 2;
2266                         do_div(x, (size_t)wc->n_blocks);
2267                         DMEMIT(" low_watermark %u", 100 - (unsigned)x);
2268                 }
2269                 if (wc->max_writeback_jobs_set)
2270                         DMEMIT(" writeback_jobs %u", wc->max_writeback_jobs);
2271                 if (wc->autocommit_blocks_set)
2272                         DMEMIT(" autocommit_blocks %u", wc->autocommit_blocks);
2273                 if (wc->autocommit_time_set)
2274                         DMEMIT(" autocommit_time %u", jiffies_to_msecs(wc->autocommit_jiffies));
2275                 if (wc->writeback_fua_set)
2276                         DMEMIT(" %sfua", wc->writeback_fua ? "" : "no");
2277                 break;
2278         }
2279 }
2280
2281 static struct target_type writecache_target = {
2282         .name                   = "writecache",
2283         .version                = {1, 1, 0},
2284         .module                 = THIS_MODULE,
2285         .ctr                    = writecache_ctr,
2286         .dtr                    = writecache_dtr,
2287         .status                 = writecache_status,
2288         .postsuspend            = writecache_suspend,
2289         .resume                 = writecache_resume,
2290         .message                = writecache_message,
2291         .map                    = writecache_map,
2292         .end_io                 = writecache_end_io,
2293         .iterate_devices        = writecache_iterate_devices,
2294         .io_hints               = writecache_io_hints,
2295 };
2296
2297 static int __init dm_writecache_init(void)
2298 {
2299         int r;
2300
2301         r = dm_register_target(&writecache_target);
2302         if (r < 0) {
2303                 DMERR("register failed %d", r);
2304                 return r;
2305         }
2306
2307         return 0;
2308 }
2309
2310 static void __exit dm_writecache_exit(void)
2311 {
2312         dm_unregister_target(&writecache_target);
2313 }
2314
2315 module_init(dm_writecache_init);
2316 module_exit(dm_writecache_exit);
2317
2318 MODULE_DESCRIPTION(DM_NAME " writecache target");
2319 MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
2320 MODULE_LICENSE("GPL");