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 #define MAX_AGE_DIV                     16
  30 #define MAX_AGE_UNSPECIFIED             -1UL
  31
  32 #define BITMAP_GRANULARITY      65536
  33 #if BITMAP_GRANULARITY < PAGE_SIZE
  34 #undef BITMAP_GRANULARITY
  35 #define BITMAP_GRANULARITY      PAGE_SIZE
  36 #endif
  37
  38 #if IS_ENABLED(CONFIG_ARCH_HAS_PMEM_API) && IS_ENABLED(CONFIG_DAX_DRIVER)
  39 #define DM_WRITECACHE_HAS_PMEM
  40 #endif
  41
  42 #ifdef DM_WRITECACHE_HAS_PMEM
  43 #define pmem_assign(dest, src)                                  \
  44 do {                                                            \
  45         typeof(dest) uniq = (src);                              \
  46         memcpy_flushcache(&(dest), &uniq, sizeof(dest));        \
  47 } while (0)
  48 #else
  49 #define pmem_assign(dest, src)  ((dest) = (src))
  50 #endif
  51
  52 #if defined(__HAVE_ARCH_MEMCPY_MCSAFE) && defined(DM_WRITECACHE_HAS_PMEM)
  53 #define DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
  54 #endif
  55
  56 #define MEMORY_SUPERBLOCK_MAGIC         0x23489321
  57 #define MEMORY_SUPERBLOCK_VERSION       1
  58
  59 struct wc_memory_entry {
  60         __le64 original_sector;
  61         __le64 seq_count;
  62 };
  63
  64 struct wc_memory_superblock {
  65         union {
  66                 struct {
  67                         __le32 magic;
  68                         __le32 version;
  69                         __le32 block_size;
  70                         __le32 pad;
  71                         __le64 n_blocks;
  72                         __le64 seq_count;
  73                 };
  74                 __le64 padding[8];
  75         };
  76         struct wc_memory_entry entries[0];
  77 };
  78
  79 struct wc_entry {
  80         struct rb_node rb_node;
  81         struct list_head lru;
  82         unsigned short wc_list_contiguous;
  83         bool write_in_progress
  84 #if BITS_PER_LONG == 64
  85                 :1
  86 #endif
  87         ;
  88         unsigned long index
  89 #if BITS_PER_LONG == 64
  90                 :47
  91 #endif
  92         ;
  93         unsigned long age;
  94 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
  95         uint64_t original_sector;
  96         uint64_t seq_count;
  97 #endif
  98 };
  99
 100 #ifdef DM_WRITECACHE_HAS_PMEM
 101 #define WC_MODE_PMEM(wc)                        ((wc)->pmem_mode)
 102 #define WC_MODE_FUA(wc)                         ((wc)->writeback_fua)
 103 #else
 104 #define WC_MODE_PMEM(wc)                        false
 105 #define WC_MODE_FUA(wc)                         false
 106 #endif
 107 #define WC_MODE_SORT_FREELIST(wc)               (!WC_MODE_PMEM(wc))
 108
 109 struct dm_writecache {
 110         struct mutex lock;
 111         struct list_head lru;
 112         union {
 113                 struct list_head freelist;
 114                 struct {
 115                         struct rb_root freetree;
 116                         struct wc_entry *current_free;
 117                 };
 118         };
 119         struct rb_root tree;
 120
 121         size_t freelist_size;
 122         size_t writeback_size;
 123         size_t freelist_high_watermark;
 124         size_t freelist_low_watermark;
 125         unsigned long max_age;
 126
 127         unsigned uncommitted_blocks;
 128         unsigned autocommit_blocks;
 129         unsigned max_writeback_jobs;
 130
 131         int error;
 132
 133         unsigned long autocommit_jiffies;
 134         struct timer_list autocommit_timer;
 135         struct wait_queue_head freelist_wait;
 136
 137         struct timer_list max_age_timer;
 138
 139         atomic_t bio_in_progress[2];
 140         struct wait_queue_head bio_in_progress_wait[2];
 141
 142         struct dm_target *ti;
 143         struct dm_dev *dev;
 144         struct dm_dev *ssd_dev;
 145         sector_t start_sector;
 146         void *memory_map;
 147         uint64_t memory_map_size;
 148         size_t metadata_sectors;
 149         size_t n_blocks;
 150         uint64_t seq_count;
 151         void *block_start;
 152         struct wc_entry *entries;
 153         unsigned block_size;
 154         unsigned char block_size_bits;
 155
 156         bool pmem_mode:1;
 157         bool writeback_fua:1;
 158
 159         bool overwrote_committed:1;
 160         bool memory_vmapped:1;
 161
 162         bool high_wm_percent_set:1;
 163         bool low_wm_percent_set:1;
 164         bool max_writeback_jobs_set:1;
 165         bool autocommit_blocks_set:1;
 166         bool autocommit_time_set:1;
 167         bool writeback_fua_set:1;
 168         bool flush_on_suspend:1;
 169         bool cleaner:1;
 170
 171         unsigned writeback_all;
 172         struct workqueue_struct *writeback_wq;
 173         struct work_struct writeback_work;
 174         struct work_struct flush_work;
 175
 176         struct dm_io_client *dm_io;
 177
 178         raw_spinlock_t endio_list_lock;
 179         struct list_head endio_list;
 180         struct task_struct *endio_thread;
 181
 182         struct task_struct *flush_thread;
 183         struct bio_list flush_list;
 184
 185         struct dm_kcopyd_client *dm_kcopyd;
 186         unsigned long *dirty_bitmap;
 187         unsigned dirty_bitmap_size;
 188
 189         struct bio_set bio_set;
 190         mempool_t copy_pool;
 191 };
 192
 193 #define WB_LIST_INLINE          16
 194
 195 struct writeback_struct {
 196         struct list_head endio_entry;
 197         struct dm_writecache *wc;
 198         struct wc_entry **wc_list;
 199         unsigned wc_list_n;
 200         struct wc_entry *wc_list_inline[WB_LIST_INLINE];
 201         struct bio bio;
 202 };
 203
 204 struct copy_struct {
 205         struct list_head endio_entry;
 206         struct dm_writecache *wc;
 207         struct wc_entry *e;
 208         unsigned n_entries;
 209         int error;
 210 };
 211
 212 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(dm_writecache_throttle,
 213                                             "A percentage of time allocated for data copying");
 214
 215 static void wc_lock(struct dm_writecache *wc)
 216 {
 217         mutex_lock(&wc->lock);
 218 }
 219
 220 static void wc_unlock(struct dm_writecache *wc)
 221 {
 222         mutex_unlock(&wc->lock);
 223 }
 224
 225 #ifdef DM_WRITECACHE_HAS_PMEM
 226 static int persistent_memory_claim(struct dm_writecache *wc)
 227 {
 228         int r;
 229         loff_t s;
 230         long p, da;
 231         pfn_t pfn;
 232         int id;
 233         struct page **pages;
 234
 235         wc->memory_vmapped = false;
 236
 237         s = wc->memory_map_size;
 238         p = s >> PAGE_SHIFT;
 239         if (!p) {
 240                 r = -EINVAL;
 241                 goto err1;
 242         }
 243         if (p != s >> PAGE_SHIFT) {
 244                 r = -EOVERFLOW;
 245                 goto err1;
 246         }
 247
 248         id = dax_read_lock();
 249
 250         da = dax_direct_access(wc->ssd_dev->dax_dev, 0, p, &wc->memory_map, &pfn);
 251         if (da < 0) {
 252                 wc->memory_map = NULL;
 253                 r = da;
 254                 goto err2;
 255         }
 256         if (!pfn_t_has_page(pfn)) {
 257                 wc->memory_map = NULL;
 258                 r = -EOPNOTSUPP;
 259                 goto err2;
 260         }
 261         if (da != p) {
 262                 long i;
 263                 wc->memory_map = NULL;
 264                 pages = kvmalloc_array(p, sizeof(struct page *), GFP_KERNEL);
 265                 if (!pages) {
 266                         r = -ENOMEM;
 267                         goto err2;
 268                 }
 269                 i = 0;
 270                 do {
 271                         long daa;
 272                         daa = dax_direct_access(wc->ssd_dev->dax_dev, i, p - i,
 273                                                 NULL, &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         return &sb(wc)->entries[e->index];
 355 }
 356
 357 static void *memory_data(struct dm_writecache *wc, struct wc_entry *e)
 358 {
 359         return (char *)wc->block_start + (e->index << wc->block_size_bits);
 360 }
 361
 362 static sector_t cache_sector(struct dm_writecache *wc, struct wc_entry *e)
 363 {
 364         return wc->start_sector + wc->metadata_sectors +
 365                 ((sector_t)e->index << (wc->block_size_bits - SECTOR_SHIFT));
 366 }
 367
 368 static uint64_t read_original_sector(struct dm_writecache *wc, struct wc_entry *e)
 369 {
 370 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 371         return e->original_sector;
 372 #else
 373         return le64_to_cpu(memory_entry(wc, e)->original_sector);
 374 #endif
 375 }
 376
 377 static uint64_t read_seq_count(struct dm_writecache *wc, struct wc_entry *e)
 378 {
 379 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 380         return e->seq_count;
 381 #else
 382         return le64_to_cpu(memory_entry(wc, e)->seq_count);
 383 #endif
 384 }
 385
 386 static void clear_seq_count(struct dm_writecache *wc, struct wc_entry *e)
 387 {
 388 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 389         e->seq_count = -1;
 390 #endif
 391         pmem_assign(memory_entry(wc, e)->seq_count, cpu_to_le64(-1));
 392 }
 393
 394 static void write_original_sector_seq_count(struct dm_writecache *wc, struct wc_entry *e,
 395                                             uint64_t original_sector, uint64_t seq_count)
 396 {
 397         struct wc_memory_entry me;
 398 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 399         e->original_sector = original_sector;
 400         e->seq_count = seq_count;
 401 #endif
 402         me.original_sector = cpu_to_le64(original_sector);
 403         me.seq_count = cpu_to_le64(seq_count);
 404         pmem_assign(*memory_entry(wc, e), me);
 405 }
 406
 407 #define writecache_error(wc, err, msg, arg...)                          \
 408 do {                                                                    \
 409         if (!cmpxchg(&(wc)->error, 0, err))                             \
 410                 DMERR(msg, ##arg);                                      \
 411         wake_up(&(wc)->freelist_wait);                                  \
 412 } while (0)
 413
 414 #define writecache_has_error(wc)        (unlikely(READ_ONCE((wc)->error)))
 415
 416 static void writecache_flush_all_metadata(struct dm_writecache *wc)
 417 {
 418         if (!WC_MODE_PMEM(wc))
 419                 memset(wc->dirty_bitmap, -1, wc->dirty_bitmap_size);
 420 }
 421
 422 static void writecache_flush_region(struct dm_writecache *wc, void *ptr, size_t size)
 423 {
 424         if (!WC_MODE_PMEM(wc))
 425                 __set_bit(((char *)ptr - (char *)wc->memory_map) / BITMAP_GRANULARITY,
 426                           wc->dirty_bitmap);
 427 }
 428
 429 static void writecache_disk_flush(struct dm_writecache *wc, struct dm_dev *dev);
 430
 431 struct io_notify {
 432         struct dm_writecache *wc;
 433         struct completion c;
 434         atomic_t count;
 435 };
 436
 437 static void writecache_notify_io(unsigned long error, void *context)
 438 {
 439         struct io_notify *endio = context;
 440
 441         if (unlikely(error != 0))
 442                 writecache_error(endio->wc, -EIO, "error writing metadata");
 443         BUG_ON(atomic_read(&endio->count) <= 0);
 444         if (atomic_dec_and_test(&endio->count))
 445                 complete(&endio->c);
 446 }
 447
 448 static void writecache_wait_for_ios(struct dm_writecache *wc, int direction)
 449 {
 450         wait_event(wc->bio_in_progress_wait[direction],
 451                    !atomic_read(&wc->bio_in_progress[direction]));
 452 }
 453
 454 static void ssd_commit_flushed(struct dm_writecache *wc, bool wait_for_ios)
 455 {
 456         struct dm_io_region region;
 457         struct dm_io_request req;
 458         struct io_notify endio = {
 459                 wc,
 460                 COMPLETION_INITIALIZER_ONSTACK(endio.c),
 461                 ATOMIC_INIT(1),
 462         };
 463         unsigned bitmap_bits = wc->dirty_bitmap_size * 8;
 464         unsigned i = 0;
 465
 466         while (1) {
 467                 unsigned j;
 468                 i = find_next_bit(wc->dirty_bitmap, bitmap_bits, i);
 469                 if (unlikely(i == bitmap_bits))
 470                         break;
 471                 j = find_next_zero_bit(wc->dirty_bitmap, bitmap_bits, i);
 472
 473                 region.bdev = wc->ssd_dev->bdev;
 474                 region.sector = (sector_t)i * (BITMAP_GRANULARITY >> SECTOR_SHIFT);
 475                 region.count = (sector_t)(j - i) * (BITMAP_GRANULARITY >> SECTOR_SHIFT);
 476
 477                 if (unlikely(region.sector >= wc->metadata_sectors))
 478                         break;
 479                 if (unlikely(region.sector + region.count > wc->metadata_sectors))
 480                         region.count = wc->metadata_sectors - region.sector;
 481
 482                 region.sector += wc->start_sector;
 483                 atomic_inc(&endio.count);
 484                 req.bi_op = REQ_OP_WRITE;
 485                 req.bi_op_flags = REQ_SYNC;
 486                 req.mem.type = DM_IO_VMA;
 487                 req.mem.ptr.vma = (char *)wc->memory_map + (size_t)i * BITMAP_GRANULARITY;
 488                 req.client = wc->dm_io;
 489                 req.notify.fn = writecache_notify_io;
 490                 req.notify.context = &endio;
 491
 492                 /* writing via async dm-io (implied by notify.fn above) won't return an error */
 493                 (void) dm_io(&req, 1, &region, NULL);
 494                 i = j;
 495         }
 496
 497         writecache_notify_io(0, &endio);
 498         wait_for_completion_io(&endio.c);
 499
 500         if (wait_for_ios)
 501                 writecache_wait_for_ios(wc, WRITE);
 502
 503         writecache_disk_flush(wc, wc->ssd_dev);
 504
 505         memset(wc->dirty_bitmap, 0, wc->dirty_bitmap_size);
 506 }
 507
 508 static void ssd_commit_superblock(struct dm_writecache *wc)
 509 {
 510         int r;
 511         struct dm_io_region region;
 512         struct dm_io_request req;
 513
 514         region.bdev = wc->ssd_dev->bdev;
 515         region.sector = 0;
 516         region.count = PAGE_SIZE;
 517
 518         if (unlikely(region.sector + region.count > wc->metadata_sectors))
 519                 region.count = wc->metadata_sectors - region.sector;
 520
 521         region.sector += wc->start_sector;
 522
 523         req.bi_op = REQ_OP_WRITE;
 524         req.bi_op_flags = REQ_SYNC | REQ_FUA;
 525         req.mem.type = DM_IO_VMA;
 526         req.mem.ptr.vma = (char *)wc->memory_map;
 527         req.client = wc->dm_io;
 528         req.notify.fn = NULL;
 529         req.notify.context = NULL;
 530
 531         r = dm_io(&req, 1, &region, NULL);
 532         if (unlikely(r))
 533                 writecache_error(wc, r, "error writing superblock");
 534 }
 535
 536 static void writecache_commit_flushed(struct dm_writecache *wc, bool wait_for_ios)
 537 {
 538         if (WC_MODE_PMEM(wc))
 539                 wmb();
 540         else
 541                 ssd_commit_flushed(wc, wait_for_ios);
 542 }
 543
 544 static void writecache_disk_flush(struct dm_writecache *wc, struct dm_dev *dev)
 545 {
 546         int r;
 547         struct dm_io_region region;
 548         struct dm_io_request req;
 549
 550         region.bdev = dev->bdev;
 551         region.sector = 0;
 552         region.count = 0;
 553         req.bi_op = REQ_OP_WRITE;
 554         req.bi_op_flags = REQ_PREFLUSH;
 555         req.mem.type = DM_IO_KMEM;
 556         req.mem.ptr.addr = NULL;
 557         req.client = wc->dm_io;
 558         req.notify.fn = NULL;
 559
 560         r = dm_io(&req, 1, &region, NULL);
 561         if (unlikely(r))
 562                 writecache_error(wc, r, "error flushing metadata: %d", r);
 563 }
 564
 565 #define WFE_RETURN_FOLLOWING    1
 566 #define WFE_LOWEST_SEQ          2
 567
 568 static struct wc_entry *writecache_find_entry(struct dm_writecache *wc,
 569                                               uint64_t block, int flags)
 570 {
 571         struct wc_entry *e;
 572         struct rb_node *node = wc->tree.rb_node;
 573
 574         if (unlikely(!node))
 575                 return NULL;
 576
 577         while (1) {
 578                 e = container_of(node, struct wc_entry, rb_node);
 579                 if (read_original_sector(wc, e) == block)
 580                         break;
 581
 582                 node = (read_original_sector(wc, e) >= block ?
 583                         e->rb_node.rb_left : e->rb_node.rb_right);
 584                 if (unlikely(!node)) {
 585                         if (!(flags & WFE_RETURN_FOLLOWING))
 586                                 return NULL;
 587                         if (read_original_sector(wc, e) >= block) {
 588                                 return e;
 589                         } else {
 590                                 node = rb_next(&e->rb_node);
 591                                 if (unlikely(!node))
 592                                         return NULL;
 593                                 e = container_of(node, struct wc_entry, rb_node);
 594                                 return e;
 595                         }
 596                 }
 597         }
 598
 599         while (1) {
 600                 struct wc_entry *e2;
 601                 if (flags & WFE_LOWEST_SEQ)
 602                         node = rb_prev(&e->rb_node);
 603                 else
 604                         node = rb_next(&e->rb_node);
 605                 if (unlikely(!node))
 606                         return e;
 607                 e2 = container_of(node, struct wc_entry, rb_node);
 608                 if (read_original_sector(wc, e2) != block)
 609                         return e;
 610                 e = e2;
 611         }
 612 }
 613
 614 static void writecache_insert_entry(struct dm_writecache *wc, struct wc_entry *ins)
 615 {
 616         struct wc_entry *e;
 617         struct rb_node **node = &wc->tree.rb_node, *parent = NULL;
 618
 619         while (*node) {
 620                 e = container_of(*node, struct wc_entry, rb_node);
 621                 parent = &e->rb_node;
 622                 if (read_original_sector(wc, e) > read_original_sector(wc, ins))
 623                         node = &parent->rb_left;
 624                 else
 625                         node = &parent->rb_right;
 626         }
 627         rb_link_node(&ins->rb_node, parent, node);
 628         rb_insert_color(&ins->rb_node, &wc->tree);
 629         list_add(&ins->lru, &wc->lru);
 630         ins->age = jiffies;
 631 }
 632
 633 static void writecache_unlink(struct dm_writecache *wc, struct wc_entry *e)
 634 {
 635         list_del(&e->lru);
 636         rb_erase(&e->rb_node, &wc->tree);
 637 }
 638
 639 static void writecache_add_to_freelist(struct dm_writecache *wc, struct wc_entry *e)
 640 {
 641         if (WC_MODE_SORT_FREELIST(wc)) {
 642                 struct rb_node **node = &wc->freetree.rb_node, *parent = NULL;
 643                 if (unlikely(!*node))
 644                         wc->current_free = e;
 645                 while (*node) {
 646                         parent = *node;
 647                         if (&e->rb_node < *node)
 648                                 node = &parent->rb_left;
 649                         else
 650                                 node = &parent->rb_right;
 651                 }
 652                 rb_link_node(&e->rb_node, parent, node);
 653                 rb_insert_color(&e->rb_node, &wc->freetree);
 654         } else {
 655                 list_add_tail(&e->lru, &wc->freelist);
 656         }
 657         wc->freelist_size++;
 658 }
 659
 660 static inline void writecache_verify_watermark(struct dm_writecache *wc)
 661 {
 662         if (unlikely(wc->freelist_size + wc->writeback_size <= wc->freelist_high_watermark))
 663                 queue_work(wc->writeback_wq, &wc->writeback_work);
 664 }
 665
 666 static void writecache_max_age_timer(struct timer_list *t)
 667 {
 668         struct dm_writecache *wc = from_timer(wc, t, max_age_timer);
 669
 670         if (!dm_suspended(wc->ti) && !writecache_has_error(wc)) {
 671                 queue_work(wc->writeback_wq, &wc->writeback_work);
 672                 mod_timer(&wc->max_age_timer, jiffies + wc->max_age / MAX_AGE_DIV);
 673         }
 674 }
 675
 676 static struct wc_entry *writecache_pop_from_freelist(struct dm_writecache *wc, sector_t expected_sector)
 677 {
 678         struct wc_entry *e;
 679
 680         if (WC_MODE_SORT_FREELIST(wc)) {
 681                 struct rb_node *next;
 682                 if (unlikely(!wc->current_free))
 683                         return NULL;
 684                 e = wc->current_free;
 685                 if (expected_sector != (sector_t)-1 && unlikely(cache_sector(wc, e) != expected_sector))
 686                         return NULL;
 687                 next = rb_next(&e->rb_node);
 688                 rb_erase(&e->rb_node, &wc->freetree);
 689                 if (unlikely(!next))
 690                         next = rb_first(&wc->freetree);
 691                 wc->current_free = next ? container_of(next, struct wc_entry, rb_node) : NULL;
 692         } else {
 693                 if (unlikely(list_empty(&wc->freelist)))
 694                         return NULL;
 695                 e = container_of(wc->freelist.next, struct wc_entry, lru);
 696                 if (expected_sector != (sector_t)-1 && unlikely(cache_sector(wc, e) != expected_sector))
 697                         return NULL;
 698                 list_del(&e->lru);
 699         }
 700         wc->freelist_size--;
 701
 702         writecache_verify_watermark(wc);
 703
 704         return e;
 705 }
 706
 707 static void writecache_free_entry(struct dm_writecache *wc, struct wc_entry *e)
 708 {
 709         writecache_unlink(wc, e);
 710         writecache_add_to_freelist(wc, e);
 711         clear_seq_count(wc, e);
 712         writecache_flush_region(wc, memory_entry(wc, e), sizeof(struct wc_memory_entry));
 713         if (unlikely(waitqueue_active(&wc->freelist_wait)))
 714                 wake_up(&wc->freelist_wait);
 715 }
 716
 717 static void writecache_wait_on_freelist(struct dm_writecache *wc)
 718 {
 719         DEFINE_WAIT(wait);
 720
 721         prepare_to_wait(&wc->freelist_wait, &wait, TASK_UNINTERRUPTIBLE);
 722         wc_unlock(wc);
 723         io_schedule();
 724         finish_wait(&wc->freelist_wait, &wait);
 725         wc_lock(wc);
 726 }
 727
 728 static void writecache_poison_lists(struct dm_writecache *wc)
 729 {
 730         /*
 731          * Catch incorrect access to these values while the device is suspended.
 732          */
 733         memset(&wc->tree, -1, sizeof wc->tree);
 734         wc->lru.next = LIST_POISON1;
 735         wc->lru.prev = LIST_POISON2;
 736         wc->freelist.next = LIST_POISON1;
 737         wc->freelist.prev = LIST_POISON2;
 738 }
 739
 740 static void writecache_flush_entry(struct dm_writecache *wc, struct wc_entry *e)
 741 {
 742         writecache_flush_region(wc, memory_entry(wc, e), sizeof(struct wc_memory_entry));
 743         if (WC_MODE_PMEM(wc))
 744                 writecache_flush_region(wc, memory_data(wc, e), wc->block_size);
 745 }
 746
 747 static bool writecache_entry_is_committed(struct dm_writecache *wc, struct wc_entry *e)
 748 {
 749         return read_seq_count(wc, e) < wc->seq_count;
 750 }
 751
 752 static void writecache_flush(struct dm_writecache *wc)
 753 {
 754         struct wc_entry *e, *e2;
 755         bool need_flush_after_free;
 756
 757         wc->uncommitted_blocks = 0;
 758         del_timer(&wc->autocommit_timer);
 759
 760         if (list_empty(&wc->lru))
 761                 return;
 762
 763         e = container_of(wc->lru.next, struct wc_entry, lru);
 764         if (writecache_entry_is_committed(wc, e)) {
 765                 if (wc->overwrote_committed) {
 766                         writecache_wait_for_ios(wc, WRITE);
 767                         writecache_disk_flush(wc, wc->ssd_dev);
 768                         wc->overwrote_committed = false;
 769                 }
 770                 return;
 771         }
 772         while (1) {
 773                 writecache_flush_entry(wc, e);
 774                 if (unlikely(e->lru.next == &wc->lru))
 775                         break;
 776                 e2 = container_of(e->lru.next, struct wc_entry, lru);
 777                 if (writecache_entry_is_committed(wc, e2))
 778                         break;
 779                 e = e2;
 780                 cond_resched();
 781         }
 782         writecache_commit_flushed(wc, true);
 783
 784         wc->seq_count++;
 785         pmem_assign(sb(wc)->seq_count, cpu_to_le64(wc->seq_count));
 786         if (WC_MODE_PMEM(wc))
 787                 writecache_commit_flushed(wc, false);
 788         else
 789                 ssd_commit_superblock(wc);
 790
 791         wc->overwrote_committed = false;
 792
 793         need_flush_after_free = false;
 794         while (1) {
 795                 /* Free another committed entry with lower seq-count */
 796                 struct rb_node *rb_node = rb_prev(&e->rb_node);
 797
 798                 if (rb_node) {
 799                         e2 = container_of(rb_node, struct wc_entry, rb_node);
 800                         if (read_original_sector(wc, e2) == read_original_sector(wc, e) &&
 801                             likely(!e2->write_in_progress)) {
 802                                 writecache_free_entry(wc, e2);
 803                                 need_flush_after_free = true;
 804                         }
 805                 }
 806                 if (unlikely(e->lru.prev == &wc->lru))
 807                         break;
 808                 e = container_of(e->lru.prev, struct wc_entry, lru);
 809                 cond_resched();
 810         }
 811
 812         if (need_flush_after_free)
 813                 writecache_commit_flushed(wc, false);
 814 }
 815
 816 static void writecache_flush_work(struct work_struct *work)
 817 {
 818         struct dm_writecache *wc = container_of(work, struct dm_writecache, flush_work);
 819
 820         wc_lock(wc);
 821         writecache_flush(wc);
 822         wc_unlock(wc);
 823 }
 824
 825 static void writecache_autocommit_timer(struct timer_list *t)
 826 {
 827         struct dm_writecache *wc = from_timer(wc, t, autocommit_timer);
 828         if (!writecache_has_error(wc))
 829                 queue_work(wc->writeback_wq, &wc->flush_work);
 830 }
 831
 832 static void writecache_schedule_autocommit(struct dm_writecache *wc)
 833 {
 834         if (!timer_pending(&wc->autocommit_timer))
 835                 mod_timer(&wc->autocommit_timer, jiffies + wc->autocommit_jiffies);
 836 }
 837
 838 static void writecache_discard(struct dm_writecache *wc, sector_t start, sector_t end)
 839 {
 840         struct wc_entry *e;
 841         bool discarded_something = false;
 842
 843         e = writecache_find_entry(wc, start, WFE_RETURN_FOLLOWING | WFE_LOWEST_SEQ);
 844         if (unlikely(!e))
 845                 return;
 846
 847         while (read_original_sector(wc, e) < end) {
 848                 struct rb_node *node = rb_next(&e->rb_node);
 849
 850                 if (likely(!e->write_in_progress)) {
 851                         if (!discarded_something) {
 852                                 writecache_wait_for_ios(wc, READ);
 853                                 writecache_wait_for_ios(wc, WRITE);
 854                                 discarded_something = true;
 855                         }
 856                         writecache_free_entry(wc, e);
 857                 }
 858
 859                 if (unlikely(!node))
 860                         break;
 861
 862                 e = container_of(node, struct wc_entry, rb_node);
 863         }
 864
 865         if (discarded_something)
 866                 writecache_commit_flushed(wc, false);
 867 }
 868
 869 static bool writecache_wait_for_writeback(struct dm_writecache *wc)
 870 {
 871         if (wc->writeback_size) {
 872                 writecache_wait_on_freelist(wc);
 873                 return true;
 874         }
 875         return false;
 876 }
 877
 878 static void writecache_suspend(struct dm_target *ti)
 879 {
 880         struct dm_writecache *wc = ti->private;
 881         bool flush_on_suspend;
 882
 883         del_timer_sync(&wc->autocommit_timer);
 884         del_timer_sync(&wc->max_age_timer);
 885
 886         wc_lock(wc);
 887         writecache_flush(wc);
 888         flush_on_suspend = wc->flush_on_suspend;
 889         if (flush_on_suspend) {
 890                 wc->flush_on_suspend = false;
 891                 wc->writeback_all++;
 892                 queue_work(wc->writeback_wq, &wc->writeback_work);
 893         }
 894         wc_unlock(wc);
 895
 896         drain_workqueue(wc->writeback_wq);
 897
 898         wc_lock(wc);
 899         if (flush_on_suspend)
 900                 wc->writeback_all--;
 901         while (writecache_wait_for_writeback(wc));
 902
 903         if (WC_MODE_PMEM(wc))
 904                 persistent_memory_flush_cache(wc->memory_map, wc->memory_map_size);
 905
 906         writecache_poison_lists(wc);
 907
 908         wc_unlock(wc);
 909 }
 910
 911 static int writecache_alloc_entries(struct dm_writecache *wc)
 912 {
 913         size_t b;
 914
 915         if (wc->entries)
 916                 return 0;
 917         wc->entries = vmalloc(array_size(sizeof(struct wc_entry), wc->n_blocks));
 918         if (!wc->entries)
 919                 return -ENOMEM;
 920         for (b = 0; b < wc->n_blocks; b++) {
 921                 struct wc_entry *e = &wc->entries[b];
 922                 e->index = b;
 923                 e->write_in_progress = false;
 924                 cond_resched();
 925         }
 926
 927         return 0;
 928 }
 929
 930 static int writecache_read_metadata(struct dm_writecache *wc, sector_t n_sectors)
 931 {
 932         struct dm_io_region region;
 933         struct dm_io_request req;
 934
 935         region.bdev = wc->ssd_dev->bdev;
 936         region.sector = wc->start_sector;
 937         region.count = n_sectors;
 938         req.bi_op = REQ_OP_READ;
 939         req.bi_op_flags = REQ_SYNC;
 940         req.mem.type = DM_IO_VMA;
 941         req.mem.ptr.vma = (char *)wc->memory_map;
 942         req.client = wc->dm_io;
 943         req.notify.fn = NULL;
 944
 945         return dm_io(&req, 1, &region, NULL);
 946 }
 947
 948 static void writecache_resume(struct dm_target *ti)
 949 {
 950         struct dm_writecache *wc = ti->private;
 951         size_t b;
 952         bool need_flush = false;
 953         __le64 sb_seq_count;
 954         int r;
 955
 956         wc_lock(wc);
 957
 958         if (WC_MODE_PMEM(wc)) {
 959                 persistent_memory_invalidate_cache(wc->memory_map, wc->memory_map_size);
 960         } else {
 961                 r = writecache_read_metadata(wc, wc->metadata_sectors);
 962                 if (r) {
 963                         size_t sb_entries_offset;
 964                         writecache_error(wc, r, "unable to read metadata: %d", r);
 965                         sb_entries_offset = offsetof(struct wc_memory_superblock, entries);
 966                         memset((char *)wc->memory_map + sb_entries_offset, -1,
 967                                (wc->metadata_sectors << SECTOR_SHIFT) - sb_entries_offset);
 968                 }
 969         }
 970
 971         wc->tree = RB_ROOT;
 972         INIT_LIST_HEAD(&wc->lru);
 973         if (WC_MODE_SORT_FREELIST(wc)) {
 974                 wc->freetree = RB_ROOT;
 975                 wc->current_free = NULL;
 976         } else {
 977                 INIT_LIST_HEAD(&wc->freelist);
 978         }
 979         wc->freelist_size = 0;
 980
 981         r = memcpy_mcsafe(&sb_seq_count, &sb(wc)->seq_count, sizeof(uint64_t));
 982         if (r) {
 983                 writecache_error(wc, r, "hardware memory error when reading superblock: %d", r);
 984                 sb_seq_count = cpu_to_le64(0);
 985         }
 986         wc->seq_count = le64_to_cpu(sb_seq_count);
 987
 988 #ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 989         for (b = 0; b < wc->n_blocks; b++) {
 990                 struct wc_entry *e = &wc->entries[b];
 991                 struct wc_memory_entry wme;
 992                 if (writecache_has_error(wc)) {
 993                         e->original_sector = -1;
 994                         e->seq_count = -1;
 995                         continue;
 996                 }
 997                 r = memcpy_mcsafe(&wme, memory_entry(wc, e), sizeof(struct wc_memory_entry));
 998                 if (r) {
 999                         writecache_error(wc, r, "hardware memory error when reading metadata entry %lu: %d",
1000                                          (unsigned long)b, r);
1001                         e->original_sector = -1;
1002                         e->seq_count = -1;
1003                 } else {
1004                         e->original_sector = le64_to_cpu(wme.original_sector);
1005                         e->seq_count = le64_to_cpu(wme.seq_count);
1006                 }
1007                 cond_resched();
1008         }
1009 #endif
1010         for (b = 0; b < wc->n_blocks; b++) {
1011                 struct wc_entry *e = &wc->entries[b];
1012                 if (!writecache_entry_is_committed(wc, e)) {
1013                         if (read_seq_count(wc, e) != -1) {
1014 erase_this:
1015                                 clear_seq_count(wc, e);
1016                                 need_flush = true;
1017                         }
1018                         writecache_add_to_freelist(wc, e);
1019                 } else {
1020                         struct wc_entry *old;
1021
1022                         old = writecache_find_entry(wc, read_original_sector(wc, e), 0);
1023                         if (!old) {
1024                                 writecache_insert_entry(wc, e);
1025                         } else {
1026                                 if (read_seq_count(wc, old) == read_seq_count(wc, e)) {
1027                                         writecache_error(wc, -EINVAL,
1028                                                  "two identical entries, position %llu, sector %llu, sequence %llu",
1029                                                  (unsigned long long)b, (unsigned long long)read_original_sector(wc, e),
1030                                                  (unsigned long long)read_seq_count(wc, e));
1031                                 }
1032                                 if (read_seq_count(wc, old) > read_seq_count(wc, e)) {
1033                                         goto erase_this;
1034                                 } else {
1035                                         writecache_free_entry(wc, old);
1036                                         writecache_insert_entry(wc, e);
1037                                         need_flush = true;
1038                                 }
1039                         }
1040                 }
1041                 cond_resched();
1042         }
1043
1044         if (need_flush) {
1045                 writecache_flush_all_metadata(wc);
1046                 writecache_commit_flushed(wc, false);
1047         }
1048
1049         writecache_verify_watermark(wc);
1050
1051         if (wc->max_age != MAX_AGE_UNSPECIFIED)
1052                 mod_timer(&wc->max_age_timer, jiffies + wc->max_age / MAX_AGE_DIV);
1053
1054         wc_unlock(wc);
1055 }
1056
1057 static int process_flush_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1058 {
1059         if (argc != 1)
1060                 return -EINVAL;
1061
1062         wc_lock(wc);
1063         if (dm_suspended(wc->ti)) {
1064                 wc_unlock(wc);
1065                 return -EBUSY;
1066         }
1067         if (writecache_has_error(wc)) {
1068                 wc_unlock(wc);
1069                 return -EIO;
1070         }
1071
1072         writecache_flush(wc);
1073         wc->writeback_all++;
1074         queue_work(wc->writeback_wq, &wc->writeback_work);
1075         wc_unlock(wc);
1076
1077         flush_workqueue(wc->writeback_wq);
1078
1079         wc_lock(wc);
1080         wc->writeback_all--;
1081         if (writecache_has_error(wc)) {
1082                 wc_unlock(wc);
1083                 return -EIO;
1084         }
1085         wc_unlock(wc);
1086
1087         return 0;
1088 }
1089
1090 static int process_flush_on_suspend_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1091 {
1092         if (argc != 1)
1093                 return -EINVAL;
1094
1095         wc_lock(wc);
1096         wc->flush_on_suspend = true;
1097         wc_unlock(wc);
1098
1099         return 0;
1100 }
1101
1102 static void activate_cleaner(struct dm_writecache *wc)
1103 {
1104         wc->flush_on_suspend = true;
1105         wc->cleaner = true;
1106         wc->freelist_high_watermark = wc->n_blocks;
1107         wc->freelist_low_watermark = wc->n_blocks;
1108 }
1109
1110 static int process_cleaner_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1111 {
1112         if (argc != 1)
1113                 return -EINVAL;
1114
1115         wc_lock(wc);
1116         activate_cleaner(wc);
1117         if (!dm_suspended(wc->ti))
1118                 writecache_verify_watermark(wc);
1119         wc_unlock(wc);
1120
1121         return 0;
1122 }
1123
1124 static int writecache_message(struct dm_target *ti, unsigned argc, char **argv,
1125                               char *result, unsigned maxlen)
1126 {
1127         int r = -EINVAL;
1128         struct dm_writecache *wc = ti->private;
1129
1130         if (!strcasecmp(argv[0], "flush"))
1131                 r = process_flush_mesg(argc, argv, wc);
1132         else if (!strcasecmp(argv[0], "flush_on_suspend"))
1133                 r = process_flush_on_suspend_mesg(argc, argv, wc);
1134         else if (!strcasecmp(argv[0], "cleaner"))
1135                 r = process_cleaner_mesg(argc, argv, wc);
1136         else
1137                 DMERR("unrecognised message received: %s", argv[0]);
1138
1139         return r;
1140 }
1141
1142 static void memcpy_flushcache_optimized(void *dest, void *source, size_t size)
1143 {
1144         /*
1145          * clflushopt performs better with block size 1024, 2048, 4096
1146          * non-temporal stores perform better with block size 512
1147          *
1148          * block size   512             1024            2048            4096
1149          * movnti       496 MB/s        642 MB/s        725 MB/s        744 MB/s
1150          * clflushopt   373 MB/s        688 MB/s        1.1 GB/s        1.2 GB/s
1151          *
1152          * We see that movnti performs better for 512-byte blocks, and
1153          * clflushopt performs better for 1024-byte and larger blocks. So, we
1154          * prefer clflushopt for sizes >= 768.
1155          *
1156          * NOTE: this happens to be the case now (with dm-writecache's single
1157          * threaded model) but re-evaluate this once memcpy_flushcache() is
1158          * enabled to use movdir64b which might invalidate this performance
1159          * advantage seen with cache-allocating-writes plus flushing.
1160          */
1161 #ifdef CONFIG_X86
1162         if (static_cpu_has(X86_FEATURE_CLFLUSHOPT) &&
1163             likely(boot_cpu_data.x86_clflush_size == 64) &&
1164             likely(size >= 768)) {
1165                 do {
1166                         memcpy((void *)dest, (void *)source, 64);
1167                         clflushopt((void *)dest);
1168                         dest += 64;
1169                         source += 64;
1170                         size -= 64;
1171                 } while (size >= 64);
1172                 return;
1173         }
1174 #endif
1175         memcpy_flushcache(dest, source, size);
1176 }
1177
1178 static void bio_copy_block(struct dm_writecache *wc, struct bio *bio, void *data)
1179 {
1180         void *buf;
1181         unsigned long flags;
1182         unsigned size;
1183         int rw = bio_data_dir(bio);
1184         unsigned remaining_size = wc->block_size;
1185
1186         do {
1187                 struct bio_vec bv = bio_iter_iovec(bio, bio->bi_iter);
1188                 buf = bvec_kmap_irq(&bv, &flags);
1189                 size = bv.bv_len;
1190                 if (unlikely(size > remaining_size))
1191                         size = remaining_size;
1192
1193                 if (rw == READ) {
1194                         int r;
1195                         r = memcpy_mcsafe(buf, data, size);
1196                         flush_dcache_page(bio_page(bio));
1197                         if (unlikely(r)) {
1198                                 writecache_error(wc, r, "hardware memory error when reading data: %d", r);
1199                                 bio->bi_status = BLK_STS_IOERR;
1200                         }
1201                 } else {
1202                         flush_dcache_page(bio_page(bio));
1203                         memcpy_flushcache_optimized(data, buf, size);
1204                 }
1205
1206                 bvec_kunmap_irq(buf, &flags);
1207
1208                 data = (char *)data + size;
1209                 remaining_size -= size;
1210                 bio_advance(bio, size);
1211         } while (unlikely(remaining_size));
1212 }
1213
1214 static int writecache_flush_thread(void *data)
1215 {
1216         struct dm_writecache *wc = data;
1217
1218         while (1) {
1219                 struct bio *bio;
1220
1221                 wc_lock(wc);
1222                 bio = bio_list_pop(&wc->flush_list);
1223                 if (!bio) {
1224                         set_current_state(TASK_INTERRUPTIBLE);
1225                         wc_unlock(wc);
1226
1227                         if (unlikely(kthread_should_stop())) {
1228                                 set_current_state(TASK_RUNNING);
1229                                 break;
1230                         }
1231
1232                         schedule();
1233                         continue;
1234                 }
1235
1236                 if (bio_op(bio) == REQ_OP_DISCARD) {
1237                         writecache_discard(wc, bio->bi_iter.bi_sector,
1238                                            bio_end_sector(bio));
1239                         wc_unlock(wc);
1240                         bio_set_dev(bio, wc->dev->bdev);
1241                         generic_make_request(bio);
1242                 } else {
1243                         writecache_flush(wc);
1244                         wc_unlock(wc);
1245                         if (writecache_has_error(wc))
1246                                 bio->bi_status = BLK_STS_IOERR;
1247                         bio_endio(bio);
1248                 }
1249         }
1250
1251         return 0;
1252 }
1253
1254 static void writecache_offload_bio(struct dm_writecache *wc, struct bio *bio)
1255 {
1256         if (bio_list_empty(&wc->flush_list))
1257                 wake_up_process(wc->flush_thread);
1258         bio_list_add(&wc->flush_list, bio);
1259 }
1260
1261 static int writecache_map(struct dm_target *ti, struct bio *bio)
1262 {
1263         struct wc_entry *e;
1264         struct dm_writecache *wc = ti->private;
1265
1266         bio->bi_private = NULL;
1267
1268         wc_lock(wc);
1269
1270         if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1271                 if (writecache_has_error(wc))
1272                         goto unlock_error;
1273                 if (WC_MODE_PMEM(wc)) {
1274                         writecache_flush(wc);
1275                         if (writecache_has_error(wc))
1276                                 goto unlock_error;
1277                         goto unlock_submit;
1278                 } else {
1279                         writecache_offload_bio(wc, bio);
1280                         goto unlock_return;
1281                 }
1282         }
1283
1284         bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
1285
1286         if (unlikely((((unsigned)bio->bi_iter.bi_sector | bio_sectors(bio)) &
1287                                 (wc->block_size / 512 - 1)) != 0)) {
1288                 DMERR("I/O is not aligned, sector %llu, size %u, block size %u",
1289                       (unsigned long long)bio->bi_iter.bi_sector,
1290                       bio->bi_iter.bi_size, wc->block_size);
1291                 goto unlock_error;
1292         }
1293
1294         if (unlikely(bio_op(bio) == REQ_OP_DISCARD)) {
1295                 if (writecache_has_error(wc))
1296                         goto unlock_error;
1297                 if (WC_MODE_PMEM(wc)) {
1298                         writecache_discard(wc, bio->bi_iter.bi_sector, bio_end_sector(bio));
1299                         goto unlock_remap_origin;
1300                 } else {
1301                         writecache_offload_bio(wc, bio);
1302                         goto unlock_return;
1303                 }
1304         }
1305
1306         if (bio_data_dir(bio) == READ) {
1307 read_next_block:
1308                 e = writecache_find_entry(wc, bio->bi_iter.bi_sector, WFE_RETURN_FOLLOWING);
1309                 if (e && read_original_sector(wc, e) == bio->bi_iter.bi_sector) {
1310                         if (WC_MODE_PMEM(wc)) {
1311                                 bio_copy_block(wc, bio, memory_data(wc, e));
1312                                 if (bio->bi_iter.bi_size)
1313                                         goto read_next_block;
1314                                 goto unlock_submit;
1315                         } else {
1316                                 dm_accept_partial_bio(bio, wc->block_size >> SECTOR_SHIFT);
1317                                 bio_set_dev(bio, wc->ssd_dev->bdev);
1318                                 bio->bi_iter.bi_sector = cache_sector(wc, e);
1319                                 if (!writecache_entry_is_committed(wc, e))
1320                                         writecache_wait_for_ios(wc, WRITE);
1321                                 goto unlock_remap;
1322                         }
1323                 } else {
1324                         if (e) {
1325                                 sector_t next_boundary =
1326                                         read_original_sector(wc, e) - bio->bi_iter.bi_sector;
1327                                 if (next_boundary < bio->bi_iter.bi_size >> SECTOR_SHIFT) {
1328                                         dm_accept_partial_bio(bio, next_boundary);
1329                                 }
1330                         }
1331                         goto unlock_remap_origin;
1332                 }
1333         } else {
1334                 do {
1335                         bool found_entry = false;
1336                         if (writecache_has_error(wc))
1337                                 goto unlock_error;
1338                         e = writecache_find_entry(wc, bio->bi_iter.bi_sector, 0);
1339                         if (e) {
1340                                 if (!writecache_entry_is_committed(wc, e))
1341                                         goto bio_copy;
1342                                 if (!WC_MODE_PMEM(wc) && !e->write_in_progress) {
1343                                         wc->overwrote_committed = true;
1344                                         goto bio_copy;
1345                                 }
1346                                 found_entry = true;
1347                         } else {
1348                                 if (unlikely(wc->cleaner))
1349                                         goto direct_write;
1350                         }
1351                         e = writecache_pop_from_freelist(wc, (sector_t)-1);
1352                         if (unlikely(!e)) {
1353                                 if (!found_entry) {
1354 direct_write:
1355                                         e = writecache_find_entry(wc, bio->bi_iter.bi_sector, WFE_RETURN_FOLLOWING);
1356                                         if (e) {
1357                                                 sector_t next_boundary = read_original_sector(wc, e) - bio->bi_iter.bi_sector;
1358                                                 BUG_ON(!next_boundary);
1359                                                 if (next_boundary < bio->bi_iter.bi_size >> SECTOR_SHIFT) {
1360                                                         dm_accept_partial_bio(bio, next_boundary);
1361                                                 }
1362                                         }
1363                                         goto unlock_remap_origin;
1364                                 }
1365                                 writecache_wait_on_freelist(wc);
1366                                 continue;
1367                         }
1368                         write_original_sector_seq_count(wc, e, bio->bi_iter.bi_sector, wc->seq_count);
1369                         writecache_insert_entry(wc, e);
1370                         wc->uncommitted_blocks++;
1371 bio_copy:
1372                         if (WC_MODE_PMEM(wc)) {
1373                                 bio_copy_block(wc, bio, memory_data(wc, e));
1374                         } else {
1375                                 unsigned bio_size = wc->block_size;
1376                                 sector_t start_cache_sec = cache_sector(wc, e);
1377                                 sector_t current_cache_sec = start_cache_sec + (bio_size >> SECTOR_SHIFT);
1378
1379                                 while (bio_size < bio->bi_iter.bi_size) {
1380                                         struct wc_entry *f = writecache_pop_from_freelist(wc, current_cache_sec);
1381                                         if (!f)
1382                                                 break;
1383                                         write_original_sector_seq_count(wc, f, bio->bi_iter.bi_sector +
1384                                                                         (bio_size >> SECTOR_SHIFT), wc->seq_count);
1385                                         writecache_insert_entry(wc, f);
1386                                         wc->uncommitted_blocks++;
1387                                         bio_size += wc->block_size;
1388                                         current_cache_sec += wc->block_size >> SECTOR_SHIFT;
1389                                 }
1390
1391                                 bio_set_dev(bio, wc->ssd_dev->bdev);
1392                                 bio->bi_iter.bi_sector = start_cache_sec;
1393                                 dm_accept_partial_bio(bio, bio_size >> SECTOR_SHIFT);
1394
1395                                 if (unlikely(wc->uncommitted_blocks >= wc->autocommit_blocks)) {
1396                                         wc->uncommitted_blocks = 0;
1397                                         queue_work(wc->writeback_wq, &wc->flush_work);
1398                                 } else {
1399                                         writecache_schedule_autocommit(wc);
1400                                 }
1401                                 goto unlock_remap;
1402                         }
1403                 } while (bio->bi_iter.bi_size);
1404
1405                 if (unlikely(bio->bi_opf & REQ_FUA ||
1406                              wc->uncommitted_blocks >= wc->autocommit_blocks))
1407                         writecache_flush(wc);
1408                 else
1409                         writecache_schedule_autocommit(wc);
1410                 goto unlock_submit;
1411         }
1412
1413 unlock_remap_origin:
1414         bio_set_dev(bio, wc->dev->bdev);
1415         wc_unlock(wc);
1416         return DM_MAPIO_REMAPPED;
1417
1418 unlock_remap:
1419         /* make sure that writecache_end_io decrements bio_in_progress: */
1420         bio->bi_private = (void *)1;
1421         atomic_inc(&wc->bio_in_progress[bio_data_dir(bio)]);
1422         wc_unlock(wc);
1423         return DM_MAPIO_REMAPPED;
1424
1425 unlock_submit:
1426         wc_unlock(wc);
1427         bio_endio(bio);
1428         return DM_MAPIO_SUBMITTED;
1429
1430 unlock_return:
1431         wc_unlock(wc);
1432         return DM_MAPIO_SUBMITTED;
1433
1434 unlock_error:
1435         wc_unlock(wc);
1436         bio_io_error(bio);
1437         return DM_MAPIO_SUBMITTED;
1438 }
1439
1440 static int writecache_end_io(struct dm_target *ti, struct bio *bio, blk_status_t *status)
1441 {
1442         struct dm_writecache *wc = ti->private;
1443
1444         if (bio->bi_private != NULL) {
1445                 int dir = bio_data_dir(bio);
1446                 if (atomic_dec_and_test(&wc->bio_in_progress[dir]))
1447                         if (unlikely(waitqueue_active(&wc->bio_in_progress_wait[dir])))
1448                                 wake_up(&wc->bio_in_progress_wait[dir]);
1449         }
1450         return 0;
1451 }
1452
1453 static int writecache_iterate_devices(struct dm_target *ti,
1454                                       iterate_devices_callout_fn fn, void *data)
1455 {
1456         struct dm_writecache *wc = ti->private;
1457
1458         return fn(ti, wc->dev, 0, ti->len, data);
1459 }
1460
1461 static void writecache_io_hints(struct dm_target *ti, struct queue_limits *limits)
1462 {
1463         struct dm_writecache *wc = ti->private;
1464
1465         if (limits->logical_block_size < wc->block_size)
1466                 limits->logical_block_size = wc->block_size;
1467
1468         if (limits->physical_block_size < wc->block_size)
1469                 limits->physical_block_size = wc->block_size;
1470
1471         if (limits->io_min < wc->block_size)
1472                 limits->io_min = wc->block_size;
1473 }
1474
1475
1476 static void writecache_writeback_endio(struct bio *bio)
1477 {
1478         struct writeback_struct *wb = container_of(bio, struct writeback_struct, bio);
1479         struct dm_writecache *wc = wb->wc;
1480         unsigned long flags;
1481
1482         raw_spin_lock_irqsave(&wc->endio_list_lock, flags);
1483         if (unlikely(list_empty(&wc->endio_list)))
1484                 wake_up_process(wc->endio_thread);
1485         list_add_tail(&wb->endio_entry, &wc->endio_list);
1486         raw_spin_unlock_irqrestore(&wc->endio_list_lock, flags);
1487 }
1488
1489 static void writecache_copy_endio(int read_err, unsigned long write_err, void *ptr)
1490 {
1491         struct copy_struct *c = ptr;
1492         struct dm_writecache *wc = c->wc;
1493
1494         c->error = likely(!(read_err | write_err)) ? 0 : -EIO;
1495
1496         raw_spin_lock_irq(&wc->endio_list_lock);
1497         if (unlikely(list_empty(&wc->endio_list)))
1498                 wake_up_process(wc->endio_thread);
1499         list_add_tail(&c->endio_entry, &wc->endio_list);
1500         raw_spin_unlock_irq(&wc->endio_list_lock);
1501 }
1502
1503 static void __writecache_endio_pmem(struct dm_writecache *wc, struct list_head *list)
1504 {
1505         unsigned i;
1506         struct writeback_struct *wb;
1507         struct wc_entry *e;
1508         unsigned long n_walked = 0;
1509
1510         do {
1511                 wb = list_entry(list->next, struct writeback_struct, endio_entry);
1512                 list_del(&wb->endio_entry);
1513
1514                 if (unlikely(wb->bio.bi_status != BLK_STS_OK))
1515                         writecache_error(wc, blk_status_to_errno(wb->bio.bi_status),
1516                                         "write error %d", wb->bio.bi_status);
1517                 i = 0;
1518                 do {
1519                         e = wb->wc_list[i];
1520                         BUG_ON(!e->write_in_progress);
1521                         e->write_in_progress = false;
1522                         INIT_LIST_HEAD(&e->lru);
1523                         if (!writecache_has_error(wc))
1524                                 writecache_free_entry(wc, e);
1525                         BUG_ON(!wc->writeback_size);
1526                         wc->writeback_size--;
1527                         n_walked++;
1528                         if (unlikely(n_walked >= ENDIO_LATENCY)) {
1529                                 writecache_commit_flushed(wc, false);
1530                                 wc_unlock(wc);
1531                                 wc_lock(wc);
1532                                 n_walked = 0;
1533                         }
1534                 } while (++i < wb->wc_list_n);
1535
1536                 if (wb->wc_list != wb->wc_list_inline)
1537                         kfree(wb->wc_list);
1538                 bio_put(&wb->bio);
1539         } while (!list_empty(list));
1540 }
1541
1542 static void __writecache_endio_ssd(struct dm_writecache *wc, struct list_head *list)
1543 {
1544         struct copy_struct *c;
1545         struct wc_entry *e;
1546
1547         do {
1548                 c = list_entry(list->next, struct copy_struct, endio_entry);
1549                 list_del(&c->endio_entry);
1550
1551                 if (unlikely(c->error))
1552                         writecache_error(wc, c->error, "copy error");
1553
1554                 e = c->e;
1555                 do {
1556                         BUG_ON(!e->write_in_progress);
1557                         e->write_in_progress = false;
1558                         INIT_LIST_HEAD(&e->lru);
1559                         if (!writecache_has_error(wc))
1560                                 writecache_free_entry(wc, e);
1561
1562                         BUG_ON(!wc->writeback_size);
1563                         wc->writeback_size--;
1564                         e++;
1565                 } while (--c->n_entries);
1566                 mempool_free(c, &wc->copy_pool);
1567         } while (!list_empty(list));
1568 }
1569
1570 static int writecache_endio_thread(void *data)
1571 {
1572         struct dm_writecache *wc = data;
1573
1574         while (1) {
1575                 struct list_head list;
1576
1577                 raw_spin_lock_irq(&wc->endio_list_lock);
1578                 if (!list_empty(&wc->endio_list))
1579                         goto pop_from_list;
1580                 set_current_state(TASK_INTERRUPTIBLE);
1581                 raw_spin_unlock_irq(&wc->endio_list_lock);
1582
1583                 if (unlikely(kthread_should_stop())) {
1584                         set_current_state(TASK_RUNNING);
1585                         break;
1586                 }
1587
1588                 schedule();
1589
1590                 continue;
1591
1592 pop_from_list:
1593                 list = wc->endio_list;
1594                 list.next->prev = list.prev->next = &list;
1595                 INIT_LIST_HEAD(&wc->endio_list);
1596                 raw_spin_unlock_irq(&wc->endio_list_lock);
1597
1598                 if (!WC_MODE_FUA(wc))
1599                         writecache_disk_flush(wc, wc->dev);
1600
1601                 wc_lock(wc);
1602
1603                 if (WC_MODE_PMEM(wc)) {
1604                         __writecache_endio_pmem(wc, &list);
1605                 } else {
1606                         __writecache_endio_ssd(wc, &list);
1607                         writecache_wait_for_ios(wc, READ);
1608                 }
1609
1610                 writecache_commit_flushed(wc, false);
1611
1612                 wc_unlock(wc);
1613         }
1614
1615         return 0;
1616 }
1617
1618 static bool wc_add_block(struct writeback_struct *wb, struct wc_entry *e, gfp_t gfp)
1619 {
1620         struct dm_writecache *wc = wb->wc;
1621         unsigned block_size = wc->block_size;
1622         void *address = memory_data(wc, e);
1623
1624         persistent_memory_flush_cache(address, block_size);
1625         return bio_add_page(&wb->bio, persistent_memory_page(address),
1626                             block_size, persistent_memory_page_offset(address)) != 0;
1627 }
1628
1629 struct writeback_list {
1630         struct list_head list;
1631         size_t size;
1632 };
1633
1634 static void __writeback_throttle(struct dm_writecache *wc, struct writeback_list *wbl)
1635 {
1636         if (unlikely(wc->max_writeback_jobs)) {
1637                 if (READ_ONCE(wc->writeback_size) - wbl->size >= wc->max_writeback_jobs) {
1638                         wc_lock(wc);
1639                         while (wc->writeback_size - wbl->size >= wc->max_writeback_jobs)
1640                                 writecache_wait_on_freelist(wc);
1641                         wc_unlock(wc);
1642                 }
1643         }
1644         cond_resched();
1645 }
1646
1647 static void __writecache_writeback_pmem(struct dm_writecache *wc, struct writeback_list *wbl)
1648 {
1649         struct wc_entry *e, *f;
1650         struct bio *bio;
1651         struct writeback_struct *wb;
1652         unsigned max_pages;
1653
1654         while (wbl->size) {
1655                 wbl->size--;
1656                 e = container_of(wbl->list.prev, struct wc_entry, lru);
1657                 list_del(&e->lru);
1658
1659                 max_pages = e->wc_list_contiguous;
1660
1661                 bio = bio_alloc_bioset(GFP_NOIO, max_pages, &wc->bio_set);
1662                 wb = container_of(bio, struct writeback_struct, bio);
1663                 wb->wc = wc;
1664                 bio->bi_end_io = writecache_writeback_endio;
1665                 bio_set_dev(bio, wc->dev->bdev);
1666                 bio->bi_iter.bi_sector = read_original_sector(wc, e);
1667                 if (max_pages <= WB_LIST_INLINE ||
1668                     unlikely(!(wb->wc_list = kmalloc_array(max_pages, sizeof(struct wc_entry *),
1669                                                            GFP_NOIO | __GFP_NORETRY |
1670                                                            __GFP_NOMEMALLOC | __GFP_NOWARN)))) {
1671                         wb->wc_list = wb->wc_list_inline;
1672                         max_pages = WB_LIST_INLINE;
1673                 }
1674
1675                 BUG_ON(!wc_add_block(wb, e, GFP_NOIO));
1676
1677                 wb->wc_list[0] = e;
1678                 wb->wc_list_n = 1;
1679
1680                 while (wbl->size && wb->wc_list_n < max_pages) {
1681                         f = container_of(wbl->list.prev, struct wc_entry, lru);
1682                         if (read_original_sector(wc, f) !=
1683                             read_original_sector(wc, e) + (wc->block_size >> SECTOR_SHIFT))
1684                                 break;
1685                         if (!wc_add_block(wb, f, GFP_NOWAIT | __GFP_NOWARN))
1686                                 break;
1687                         wbl->size--;
1688                         list_del(&f->lru);
1689                         wb->wc_list[wb->wc_list_n++] = f;
1690                         e = f;
1691                 }
1692                 bio_set_op_attrs(bio, REQ_OP_WRITE, WC_MODE_FUA(wc) * REQ_FUA);
1693                 if (writecache_has_error(wc)) {
1694                         bio->bi_status = BLK_STS_IOERR;
1695                         bio_endio(bio);
1696                 } else {
1697                         submit_bio(bio);
1698                 }
1699
1700                 __writeback_throttle(wc, wbl);
1701         }
1702 }
1703
1704 static void __writecache_writeback_ssd(struct dm_writecache *wc, struct writeback_list *wbl)
1705 {
1706         struct wc_entry *e, *f;
1707         struct dm_io_region from, to;
1708         struct copy_struct *c;
1709
1710         while (wbl->size) {
1711                 unsigned n_sectors;
1712
1713                 wbl->size--;
1714                 e = container_of(wbl->list.prev, struct wc_entry, lru);
1715                 list_del(&e->lru);
1716
1717                 n_sectors = e->wc_list_contiguous << (wc->block_size_bits - SECTOR_SHIFT);
1718
1719                 from.bdev = wc->ssd_dev->bdev;
1720                 from.sector = cache_sector(wc, e);
1721                 from.count = n_sectors;
1722                 to.bdev = wc->dev->bdev;
1723                 to.sector = read_original_sector(wc, e);
1724                 to.count = n_sectors;
1725
1726                 c = mempool_alloc(&wc->copy_pool, GFP_NOIO);
1727                 c->wc = wc;
1728                 c->e = e;
1729                 c->n_entries = e->wc_list_contiguous;
1730
1731                 while ((n_sectors -= wc->block_size >> SECTOR_SHIFT)) {
1732                         wbl->size--;
1733                         f = container_of(wbl->list.prev, struct wc_entry, lru);
1734                         BUG_ON(f != e + 1);
1735                         list_del(&f->lru);
1736                         e = f;
1737                 }
1738
1739                 dm_kcopyd_copy(wc->dm_kcopyd, &from, 1, &to, 0, writecache_copy_endio, c);
1740
1741                 __writeback_throttle(wc, wbl);
1742         }
1743 }
1744
1745 static void writecache_writeback(struct work_struct *work)
1746 {
1747         struct dm_writecache *wc = container_of(work, struct dm_writecache, writeback_work);
1748         struct blk_plug plug;
1749         struct wc_entry *f, *uninitialized_var(g), *e = NULL;
1750         struct rb_node *node, *next_node;
1751         struct list_head skipped;
1752         struct writeback_list wbl;
1753         unsigned long n_walked;
1754
1755         wc_lock(wc);
1756 restart:
1757         if (writecache_has_error(wc)) {
1758                 wc_unlock(wc);
1759                 return;
1760         }
1761
1762         if (unlikely(wc->writeback_all)) {
1763                 if (writecache_wait_for_writeback(wc))
1764                         goto restart;
1765         }
1766
1767         if (wc->overwrote_committed) {
1768                 writecache_wait_for_ios(wc, WRITE);
1769         }
1770
1771         n_walked = 0;
1772         INIT_LIST_HEAD(&skipped);
1773         INIT_LIST_HEAD(&wbl.list);
1774         wbl.size = 0;
1775         while (!list_empty(&wc->lru) &&
1776                (wc->writeback_all ||
1777                 wc->freelist_size + wc->writeback_size <= wc->freelist_low_watermark ||
1778                 (jiffies - container_of(wc->lru.prev, struct wc_entry, lru)->age >=
1779                  wc->max_age - wc->max_age / MAX_AGE_DIV))) {
1780
1781                 n_walked++;
1782                 if (unlikely(n_walked > WRITEBACK_LATENCY) &&
1783                     likely(!wc->writeback_all) && likely(!dm_suspended(wc->ti))) {
1784                         queue_work(wc->writeback_wq, &wc->writeback_work);
1785                         break;
1786                 }
1787
1788                 if (unlikely(wc->writeback_all)) {
1789                         if (unlikely(!e)) {
1790                                 writecache_flush(wc);
1791                                 e = container_of(rb_first(&wc->tree), struct wc_entry, rb_node);
1792                         } else
1793                                 e = g;
1794                 } else
1795                         e = container_of(wc->lru.prev, struct wc_entry, lru);
1796                 BUG_ON(e->write_in_progress);
1797                 if (unlikely(!writecache_entry_is_committed(wc, e))) {
1798                         writecache_flush(wc);
1799                 }
1800                 node = rb_prev(&e->rb_node);
1801                 if (node) {
1802                         f = container_of(node, struct wc_entry, rb_node);
1803                         if (unlikely(read_original_sector(wc, f) ==
1804                                      read_original_sector(wc, e))) {
1805                                 BUG_ON(!f->write_in_progress);
1806                                 list_del(&e->lru);
1807                                 list_add(&e->lru, &skipped);
1808                                 cond_resched();
1809                                 continue;
1810                         }
1811                 }
1812                 wc->writeback_size++;
1813                 list_del(&e->lru);
1814                 list_add(&e->lru, &wbl.list);
1815                 wbl.size++;
1816                 e->write_in_progress = true;
1817                 e->wc_list_contiguous = 1;
1818
1819                 f = e;
1820
1821                 while (1) {
1822                         next_node = rb_next(&f->rb_node);
1823                         if (unlikely(!next_node))
1824                                 break;
1825                         g = container_of(next_node, struct wc_entry, rb_node);
1826                         if (unlikely(read_original_sector(wc, g) ==
1827                             read_original_sector(wc, f))) {
1828                                 f = g;
1829                                 continue;
1830                         }
1831                         if (read_original_sector(wc, g) !=
1832                             read_original_sector(wc, f) + (wc->block_size >> SECTOR_SHIFT))
1833                                 break;
1834                         if (unlikely(g->write_in_progress))
1835                                 break;
1836                         if (unlikely(!writecache_entry_is_committed(wc, g)))
1837                                 break;
1838
1839                         if (!WC_MODE_PMEM(wc)) {
1840                                 if (g != f + 1)
1841                                         break;
1842                         }
1843
1844                         n_walked++;
1845                         //if (unlikely(n_walked > WRITEBACK_LATENCY) && likely(!wc->writeback_all))
1846                         //      break;
1847
1848                         wc->writeback_size++;
1849                         list_del(&g->lru);
1850                         list_add(&g->lru, &wbl.list);
1851                         wbl.size++;
1852                         g->write_in_progress = true;
1853                         g->wc_list_contiguous = BIO_MAX_PAGES;
1854                         f = g;
1855                         e->wc_list_contiguous++;
1856                         if (unlikely(e->wc_list_contiguous == BIO_MAX_PAGES)) {
1857                                 if (unlikely(wc->writeback_all)) {
1858                                         next_node = rb_next(&f->rb_node);
1859                                         if (likely(next_node))
1860                                                 g = container_of(next_node, struct wc_entry, rb_node);
1861                                 }
1862                                 break;
1863                         }
1864                 }
1865                 cond_resched();
1866         }
1867
1868         if (!list_empty(&skipped)) {
1869                 list_splice_tail(&skipped, &wc->lru);
1870                 /*
1871                  * If we didn't do any progress, we must wait until some
1872                  * writeback finishes to avoid burning CPU in a loop
1873                  */
1874                 if (unlikely(!wbl.size))
1875                         writecache_wait_for_writeback(wc);
1876         }
1877
1878         wc_unlock(wc);
1879
1880         blk_start_plug(&plug);
1881
1882         if (WC_MODE_PMEM(wc))
1883                 __writecache_writeback_pmem(wc, &wbl);
1884         else
1885                 __writecache_writeback_ssd(wc, &wbl);
1886
1887         blk_finish_plug(&plug);
1888
1889         if (unlikely(wc->writeback_all)) {
1890                 wc_lock(wc);
1891                 while (writecache_wait_for_writeback(wc));
1892                 wc_unlock(wc);
1893         }
1894 }
1895
1896 static int calculate_memory_size(uint64_t device_size, unsigned block_size,
1897                                  size_t *n_blocks_p, size_t *n_metadata_blocks_p)
1898 {
1899         uint64_t n_blocks, offset;
1900         struct wc_entry e;
1901
1902         n_blocks = device_size;
1903         do_div(n_blocks, block_size + sizeof(struct wc_memory_entry));
1904
1905         while (1) {
1906                 if (!n_blocks)
1907                         return -ENOSPC;
1908                 /* Verify the following entries[n_blocks] won't overflow */
1909                 if (n_blocks >= ((size_t)-sizeof(struct wc_memory_superblock) /
1910                                  sizeof(struct wc_memory_entry)))
1911                         return -EFBIG;
1912                 offset = offsetof(struct wc_memory_superblock, entries[n_blocks]);
1913                 offset = (offset + block_size - 1) & ~(uint64_t)(block_size - 1);
1914                 if (offset + n_blocks * block_size <= device_size)
1915                         break;
1916                 n_blocks--;
1917         }
1918
1919         /* check if the bit field overflows */
1920         e.index = n_blocks;
1921         if (e.index != n_blocks)
1922                 return -EFBIG;
1923
1924         if (n_blocks_p)
1925                 *n_blocks_p = n_blocks;
1926         if (n_metadata_blocks_p)
1927                 *n_metadata_blocks_p = offset >> __ffs(block_size);
1928         return 0;
1929 }
1930
1931 static int init_memory(struct dm_writecache *wc)
1932 {
1933         size_t b;
1934         int r;
1935
1936         r = calculate_memory_size(wc->memory_map_size, wc->block_size, &wc->n_blocks, NULL);
1937         if (r)
1938                 return r;
1939
1940         r = writecache_alloc_entries(wc);
1941         if (r)
1942                 return r;
1943
1944         for (b = 0; b < ARRAY_SIZE(sb(wc)->padding); b++)
1945                 pmem_assign(sb(wc)->padding[b], cpu_to_le64(0));
1946         pmem_assign(sb(wc)->version, cpu_to_le32(MEMORY_SUPERBLOCK_VERSION));
1947         pmem_assign(sb(wc)->block_size, cpu_to_le32(wc->block_size));
1948         pmem_assign(sb(wc)->n_blocks, cpu_to_le64(wc->n_blocks));
1949         pmem_assign(sb(wc)->seq_count, cpu_to_le64(0));
1950
1951         for (b = 0; b < wc->n_blocks; b++) {
1952                 write_original_sector_seq_count(wc, &wc->entries[b], -1, -1);
1953                 cond_resched();
1954         }
1955
1956         writecache_flush_all_metadata(wc);
1957         writecache_commit_flushed(wc, false);
1958         pmem_assign(sb(wc)->magic, cpu_to_le32(MEMORY_SUPERBLOCK_MAGIC));
1959         writecache_flush_region(wc, &sb(wc)->magic, sizeof sb(wc)->magic);
1960         writecache_commit_flushed(wc, false);
1961
1962         return 0;
1963 }
1964
1965 static void writecache_dtr(struct dm_target *ti)
1966 {
1967         struct dm_writecache *wc = ti->private;
1968
1969         if (!wc)
1970                 return;
1971
1972         if (wc->endio_thread)
1973                 kthread_stop(wc->endio_thread);
1974
1975         if (wc->flush_thread)
1976                 kthread_stop(wc->flush_thread);
1977
1978         bioset_exit(&wc->bio_set);
1979
1980         mempool_exit(&wc->copy_pool);
1981
1982         if (wc->writeback_wq)
1983                 destroy_workqueue(wc->writeback_wq);
1984
1985         if (wc->dev)
1986                 dm_put_device(ti, wc->dev);
1987
1988         if (wc->ssd_dev)
1989                 dm_put_device(ti, wc->ssd_dev);
1990
1991         if (wc->entries)
1992                 vfree(wc->entries);
1993
1994         if (wc->memory_map) {
1995                 if (WC_MODE_PMEM(wc))
1996                         persistent_memory_release(wc);
1997                 else
1998                         vfree(wc->memory_map);
1999         }
2000
2001         if (wc->dm_kcopyd)
2002                 dm_kcopyd_client_destroy(wc->dm_kcopyd);
2003
2004         if (wc->dm_io)
2005                 dm_io_client_destroy(wc->dm_io);
2006
2007         if (wc->dirty_bitmap)
2008                 vfree(wc->dirty_bitmap);
2009
2010         kfree(wc);
2011 }
2012
2013 static int writecache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2014 {
2015         struct dm_writecache *wc;
2016         struct dm_arg_set as;
2017         const char *string;
2018         unsigned opt_params;
2019         size_t offset, data_size;
2020         int i, r;
2021         char dummy;
2022         int high_wm_percent = HIGH_WATERMARK;
2023         int low_wm_percent = LOW_WATERMARK;
2024         uint64_t x;
2025         struct wc_memory_superblock s;
2026
2027         static struct dm_arg _args[] = {
2028                 {0, 10, "Invalid number of feature args"},
2029         };
2030
2031         as.argc = argc;
2032         as.argv = argv;
2033
2034         wc = kzalloc(sizeof(struct dm_writecache), GFP_KERNEL);
2035         if (!wc) {
2036                 ti->error = "Cannot allocate writecache structure";
2037                 r = -ENOMEM;
2038                 goto bad;
2039         }
2040         ti->private = wc;
2041         wc->ti = ti;
2042
2043         mutex_init(&wc->lock);
2044         wc->max_age = MAX_AGE_UNSPECIFIED;
2045         writecache_poison_lists(wc);
2046         init_waitqueue_head(&wc->freelist_wait);
2047         timer_setup(&wc->autocommit_timer, writecache_autocommit_timer, 0);
2048         timer_setup(&wc->max_age_timer, writecache_max_age_timer, 0);
2049
2050         for (i = 0; i < 2; i++) {
2051                 atomic_set(&wc->bio_in_progress[i], 0);
2052                 init_waitqueue_head(&wc->bio_in_progress_wait[i]);
2053         }
2054
2055         wc->dm_io = dm_io_client_create();
2056         if (IS_ERR(wc->dm_io)) {
2057                 r = PTR_ERR(wc->dm_io);
2058                 ti->error = "Unable to allocate dm-io client";
2059                 wc->dm_io = NULL;
2060                 goto bad;
2061         }
2062
2063         wc->writeback_wq = alloc_workqueue("writecache-writeback", WQ_MEM_RECLAIM, 1);
2064         if (!wc->writeback_wq) {
2065                 r = -ENOMEM;
2066                 ti->error = "Could not allocate writeback workqueue";
2067                 goto bad;
2068         }
2069         INIT_WORK(&wc->writeback_work, writecache_writeback);
2070         INIT_WORK(&wc->flush_work, writecache_flush_work);
2071
2072         raw_spin_lock_init(&wc->endio_list_lock);
2073         INIT_LIST_HEAD(&wc->endio_list);
2074         wc->endio_thread = kthread_create(writecache_endio_thread, wc, "writecache_endio");
2075         if (IS_ERR(wc->endio_thread)) {
2076                 r = PTR_ERR(wc->endio_thread);
2077                 wc->endio_thread = NULL;
2078                 ti->error = "Couldn't spawn endio thread";
2079                 goto bad;
2080         }
2081         wake_up_process(wc->endio_thread);
2082
2083         /*
2084          * Parse the mode (pmem or ssd)
2085          */
2086         string = dm_shift_arg(&as);
2087         if (!string)
2088                 goto bad_arguments;
2089
2090         if (!strcasecmp(string, "s")) {
2091                 wc->pmem_mode = false;
2092         } else if (!strcasecmp(string, "p")) {
2093 #ifdef DM_WRITECACHE_HAS_PMEM
2094                 wc->pmem_mode = true;
2095                 wc->writeback_fua = true;
2096 #else
2097                 /*
2098                  * If the architecture doesn't support persistent memory or
2099                  * the kernel doesn't support any DAX drivers, this driver can
2100                  * only be used in SSD-only mode.
2101                  */
2102                 r = -EOPNOTSUPP;
2103                 ti->error = "Persistent memory or DAX not supported on this system";
2104                 goto bad;
2105 #endif
2106         } else {
2107                 goto bad_arguments;
2108         }
2109
2110         if (WC_MODE_PMEM(wc)) {
2111                 r = bioset_init(&wc->bio_set, BIO_POOL_SIZE,
2112                                 offsetof(struct writeback_struct, bio),
2113                                 BIOSET_NEED_BVECS);
2114                 if (r) {
2115                         ti->error = "Could not allocate bio set";
2116                         goto bad;
2117                 }
2118         } else {
2119                 r = mempool_init_kmalloc_pool(&wc->copy_pool, 1, sizeof(struct copy_struct));
2120                 if (r) {
2121                         ti->error = "Could not allocate mempool";
2122                         goto bad;
2123                 }
2124         }
2125
2126         /*
2127          * Parse the origin data device
2128          */
2129         string = dm_shift_arg(&as);
2130         if (!string)
2131                 goto bad_arguments;
2132         r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->dev);
2133         if (r) {
2134                 ti->error = "Origin data device lookup failed";
2135                 goto bad;
2136         }
2137
2138         /*
2139          * Parse cache data device (be it pmem or ssd)
2140          */
2141         string = dm_shift_arg(&as);
2142         if (!string)
2143                 goto bad_arguments;
2144
2145         r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->ssd_dev);
2146         if (r) {
2147                 ti->error = "Cache data device lookup failed";
2148                 goto bad;
2149         }
2150         wc->memory_map_size = i_size_read(wc->ssd_dev->bdev->bd_inode);
2151
2152         /*
2153          * Parse the cache block size
2154          */
2155         string = dm_shift_arg(&as);
2156         if (!string)
2157                 goto bad_arguments;
2158         if (sscanf(string, "%u%c", &wc->block_size, &dummy) != 1 ||
2159             wc->block_size < 512 || wc->block_size > PAGE_SIZE ||
2160             (wc->block_size & (wc->block_size - 1))) {
2161                 r = -EINVAL;
2162                 ti->error = "Invalid block size";
2163                 goto bad;
2164         }
2165         if (wc->block_size < bdev_logical_block_size(wc->dev->bdev) ||
2166             wc->block_size < bdev_logical_block_size(wc->ssd_dev->bdev)) {
2167                 r = -EINVAL;
2168                 ti->error = "Block size is smaller than device logical block size";
2169                 goto bad;
2170         }
2171         wc->block_size_bits = __ffs(wc->block_size);
2172
2173         wc->max_writeback_jobs = MAX_WRITEBACK_JOBS;
2174         wc->autocommit_blocks = !WC_MODE_PMEM(wc) ? AUTOCOMMIT_BLOCKS_SSD : AUTOCOMMIT_BLOCKS_PMEM;
2175         wc->autocommit_jiffies = msecs_to_jiffies(AUTOCOMMIT_MSEC);
2176
2177         /*
2178          * Parse optional arguments
2179          */
2180         r = dm_read_arg_group(_args, &as, &opt_params, &ti->error);
2181         if (r)
2182                 goto bad;
2183
2184         while (opt_params) {
2185                 string = dm_shift_arg(&as), opt_params--;
2186                 if (!strcasecmp(string, "start_sector") && opt_params >= 1) {
2187                         unsigned long long start_sector;
2188                         string = dm_shift_arg(&as), opt_params--;
2189                         if (sscanf(string, "%llu%c", &start_sector, &dummy) != 1)
2190                                 goto invalid_optional;
2191                         wc->start_sector = start_sector;
2192                         if (wc->start_sector != start_sector ||
2193                             wc->start_sector >= wc->memory_map_size >> SECTOR_SHIFT)
2194                                 goto invalid_optional;
2195                 } else if (!strcasecmp(string, "high_watermark") && opt_params >= 1) {
2196                         string = dm_shift_arg(&as), opt_params--;
2197                         if (sscanf(string, "%d%c", &high_wm_percent, &dummy) != 1)
2198                                 goto invalid_optional;
2199                         if (high_wm_percent < 0 || high_wm_percent > 100)
2200                                 goto invalid_optional;
2201                         wc->high_wm_percent_set = true;
2202                 } else if (!strcasecmp(string, "low_watermark") && opt_params >= 1) {
2203                         string = dm_shift_arg(&as), opt_params--;
2204                         if (sscanf(string, "%d%c", &low_wm_percent, &dummy) != 1)
2205                                 goto invalid_optional;
2206                         if (low_wm_percent < 0 || low_wm_percent > 100)
2207                                 goto invalid_optional;
2208                         wc->low_wm_percent_set = true;
2209                 } else if (!strcasecmp(string, "writeback_jobs") && opt_params >= 1) {
2210                         string = dm_shift_arg(&as), opt_params--;
2211                         if (sscanf(string, "%u%c", &wc->max_writeback_jobs, &dummy) != 1)
2212                                 goto invalid_optional;
2213                         wc->max_writeback_jobs_set = true;
2214                 } else if (!strcasecmp(string, "autocommit_blocks") && opt_params >= 1) {
2215                         string = dm_shift_arg(&as), opt_params--;
2216                         if (sscanf(string, "%u%c", &wc->autocommit_blocks, &dummy) != 1)
2217                                 goto invalid_optional;
2218                         wc->autocommit_blocks_set = true;
2219                 } else if (!strcasecmp(string, "autocommit_time") && opt_params >= 1) {
2220                         unsigned autocommit_msecs;
2221                         string = dm_shift_arg(&as), opt_params--;
2222                         if (sscanf(string, "%u%c", &autocommit_msecs, &dummy) != 1)
2223                                 goto invalid_optional;
2224                         if (autocommit_msecs > 3600000)
2225                                 goto invalid_optional;
2226                         wc->autocommit_jiffies = msecs_to_jiffies(autocommit_msecs);
2227                         wc->autocommit_time_set = true;
2228                 } else if (!strcasecmp(string, "max_age") && opt_params >= 1) {
2229                         unsigned max_age_msecs;
2230                         string = dm_shift_arg(&as), opt_params--;
2231                         if (sscanf(string, "%u%c", &max_age_msecs, &dummy) != 1)
2232                                 goto invalid_optional;
2233                         if (max_age_msecs > 86400000)
2234                                 goto invalid_optional;
2235                         wc->max_age = msecs_to_jiffies(max_age_msecs);
2236                 } else if (!strcasecmp(string, "cleaner")) {
2237                         wc->cleaner = true;
2238                 } else if (!strcasecmp(string, "fua")) {
2239                         if (WC_MODE_PMEM(wc)) {
2240                                 wc->writeback_fua = true;
2241                                 wc->writeback_fua_set = true;
2242                         } else goto invalid_optional;
2243                 } else if (!strcasecmp(string, "nofua")) {
2244                         if (WC_MODE_PMEM(wc)) {
2245                                 wc->writeback_fua = false;
2246                                 wc->writeback_fua_set = true;
2247                         } else goto invalid_optional;
2248                 } else {
2249 invalid_optional:
2250                         r = -EINVAL;
2251                         ti->error = "Invalid optional argument";
2252                         goto bad;
2253                 }
2254         }
2255
2256         if (high_wm_percent < low_wm_percent) {
2257                 r = -EINVAL;
2258                 ti->error = "High watermark must be greater than or equal to low watermark";
2259                 goto bad;
2260         }
2261
2262         if (WC_MODE_PMEM(wc)) {
2263                 r = persistent_memory_claim(wc);
2264                 if (r) {
2265                         ti->error = "Unable to map persistent memory for cache";
2266                         goto bad;
2267                 }
2268         } else {
2269                 size_t n_blocks, n_metadata_blocks;
2270                 uint64_t n_bitmap_bits;
2271
2272                 wc->memory_map_size -= (uint64_t)wc->start_sector << SECTOR_SHIFT;
2273
2274                 bio_list_init(&wc->flush_list);
2275                 wc->flush_thread = kthread_create(writecache_flush_thread, wc, "dm_writecache_flush");
2276                 if (IS_ERR(wc->flush_thread)) {
2277                         r = PTR_ERR(wc->flush_thread);
2278                         wc->flush_thread = NULL;
2279                         ti->error = "Couldn't spawn flush thread";
2280                         goto bad;
2281                 }
2282                 wake_up_process(wc->flush_thread);
2283
2284                 r = calculate_memory_size(wc->memory_map_size, wc->block_size,
2285                                           &n_blocks, &n_metadata_blocks);
2286                 if (r) {
2287                         ti->error = "Invalid device size";
2288                         goto bad;
2289                 }
2290
2291                 n_bitmap_bits = (((uint64_t)n_metadata_blocks << wc->block_size_bits) +
2292                                  BITMAP_GRANULARITY - 1) / BITMAP_GRANULARITY;
2293                 /* this is limitation of test_bit functions */
2294                 if (n_bitmap_bits > 1U << 31) {
2295                         r = -EFBIG;
2296                         ti->error = "Invalid device size";
2297                         goto bad;
2298                 }
2299
2300                 wc->memory_map = vmalloc(n_metadata_blocks << wc->block_size_bits);
2301                 if (!wc->memory_map) {
2302                         r = -ENOMEM;
2303                         ti->error = "Unable to allocate memory for metadata";
2304                         goto bad;
2305                 }
2306
2307                 wc->dm_kcopyd = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2308                 if (IS_ERR(wc->dm_kcopyd)) {
2309                         r = PTR_ERR(wc->dm_kcopyd);
2310                         ti->error = "Unable to allocate dm-kcopyd client";
2311                         wc->dm_kcopyd = NULL;
2312                         goto bad;
2313                 }
2314
2315                 wc->metadata_sectors = n_metadata_blocks << (wc->block_size_bits - SECTOR_SHIFT);
2316                 wc->dirty_bitmap_size = (n_bitmap_bits + BITS_PER_LONG - 1) /
2317                         BITS_PER_LONG * sizeof(unsigned long);
2318                 wc->dirty_bitmap = vzalloc(wc->dirty_bitmap_size);
2319                 if (!wc->dirty_bitmap) {
2320                         r = -ENOMEM;
2321                         ti->error = "Unable to allocate dirty bitmap";
2322                         goto bad;
2323                 }
2324
2325                 r = writecache_read_metadata(wc, wc->block_size >> SECTOR_SHIFT);
2326                 if (r) {
2327                         ti->error = "Unable to read first block of metadata";
2328                         goto bad;
2329                 }
2330         }
2331
2332         r = memcpy_mcsafe(&s, sb(wc), sizeof(struct wc_memory_superblock));
2333         if (r) {
2334                 ti->error = "Hardware memory error when reading superblock";
2335                 goto bad;
2336         }
2337         if (!le32_to_cpu(s.magic) && !le32_to_cpu(s.version)) {
2338                 r = init_memory(wc);
2339                 if (r) {
2340                         ti->error = "Unable to initialize device";
2341                         goto bad;
2342                 }
2343                 r = memcpy_mcsafe(&s, sb(wc), sizeof(struct wc_memory_superblock));
2344                 if (r) {
2345                         ti->error = "Hardware memory error when reading superblock";
2346                         goto bad;
2347                 }
2348         }
2349
2350         if (le32_to_cpu(s.magic) != MEMORY_SUPERBLOCK_MAGIC) {
2351                 ti->error = "Invalid magic in the superblock";
2352                 r = -EINVAL;
2353                 goto bad;
2354         }
2355
2356         if (le32_to_cpu(s.version) != MEMORY_SUPERBLOCK_VERSION) {
2357                 ti->error = "Invalid version in the superblock";
2358                 r = -EINVAL;
2359                 goto bad;
2360         }
2361
2362         if (le32_to_cpu(s.block_size) != wc->block_size) {
2363                 ti->error = "Block size does not match superblock";
2364                 r = -EINVAL;
2365                 goto bad;
2366         }
2367
2368         wc->n_blocks = le64_to_cpu(s.n_blocks);
2369
2370         offset = wc->n_blocks * sizeof(struct wc_memory_entry);
2371         if (offset / sizeof(struct wc_memory_entry) != le64_to_cpu(sb(wc)->n_blocks)) {
2372 overflow:
2373                 ti->error = "Overflow in size calculation";
2374                 r = -EINVAL;
2375                 goto bad;
2376         }
2377         offset += sizeof(struct wc_memory_superblock);
2378         if (offset < sizeof(struct wc_memory_superblock))
2379                 goto overflow;
2380         offset = (offset + wc->block_size - 1) & ~(size_t)(wc->block_size - 1);
2381         data_size = wc->n_blocks * (size_t)wc->block_size;
2382         if (!offset || (data_size / wc->block_size != wc->n_blocks) ||
2383             (offset + data_size < offset))
2384                 goto overflow;
2385         if (offset + data_size > wc->memory_map_size) {
2386                 ti->error = "Memory area is too small";
2387                 r = -EINVAL;
2388                 goto bad;
2389         }
2390
2391         wc->metadata_sectors = offset >> SECTOR_SHIFT;
2392         wc->block_start = (char *)sb(wc) + offset;
2393
2394         x = (uint64_t)wc->n_blocks * (100 - high_wm_percent);
2395         x += 50;
2396         do_div(x, 100);
2397         wc->freelist_high_watermark = x;
2398         x = (uint64_t)wc->n_blocks * (100 - low_wm_percent);
2399         x += 50;
2400         do_div(x, 100);
2401         wc->freelist_low_watermark = x;
2402
2403         if (wc->cleaner)
2404                 activate_cleaner(wc);
2405
2406         r = writecache_alloc_entries(wc);
2407         if (r) {
2408                 ti->error = "Cannot allocate memory";
2409                 goto bad;
2410         }
2411
2412         ti->num_flush_bios = 1;
2413         ti->flush_supported = true;
2414         ti->num_discard_bios = 1;
2415
2416         if (WC_MODE_PMEM(wc))
2417                 persistent_memory_flush_cache(wc->memory_map, wc->memory_map_size);
2418
2419         return 0;
2420
2421 bad_arguments:
2422         r = -EINVAL;
2423         ti->error = "Bad arguments";
2424 bad:
2425         writecache_dtr(ti);
2426         return r;
2427 }
2428
2429 static void writecache_status(struct dm_target *ti, status_type_t type,
2430                               unsigned status_flags, char *result, unsigned maxlen)
2431 {
2432         struct dm_writecache *wc = ti->private;
2433         unsigned extra_args;
2434         unsigned sz = 0;
2435         uint64_t x;
2436
2437         switch (type) {
2438         case STATUSTYPE_INFO:
2439                 DMEMIT("%ld %llu %llu %llu", writecache_has_error(wc),
2440                        (unsigned long long)wc->n_blocks, (unsigned long long)wc->freelist_size,
2441                        (unsigned long long)wc->writeback_size);
2442                 break;
2443         case STATUSTYPE_TABLE:
2444                 DMEMIT("%c %s %s %u ", WC_MODE_PMEM(wc) ? 'p' : 's',
2445                                 wc->dev->name, wc->ssd_dev->name, wc->block_size);
2446                 extra_args = 0;
2447                 if (wc->start_sector)
2448                         extra_args += 2;
2449                 if (wc->high_wm_percent_set && !wc->cleaner)
2450                         extra_args += 2;
2451                 if (wc->low_wm_percent_set && !wc->cleaner)
2452                         extra_args += 2;
2453                 if (wc->max_writeback_jobs_set)
2454                         extra_args += 2;
2455                 if (wc->autocommit_blocks_set)
2456                         extra_args += 2;
2457                 if (wc->autocommit_time_set)
2458                         extra_args += 2;
2459                 if (wc->cleaner)
2460                         extra_args++;
2461                 if (wc->writeback_fua_set)
2462                         extra_args++;
2463
2464                 DMEMIT("%u", extra_args);
2465                 if (wc->start_sector)
2466                         DMEMIT(" start_sector %llu", (unsigned long long)wc->start_sector);
2467                 if (wc->high_wm_percent_set && !wc->cleaner) {
2468                         x = (uint64_t)wc->freelist_high_watermark * 100;
2469                         x += wc->n_blocks / 2;
2470                         do_div(x, (size_t)wc->n_blocks);
2471                         DMEMIT(" high_watermark %u", 100 - (unsigned)x);
2472                 }
2473                 if (wc->low_wm_percent_set && !wc->cleaner) {
2474                         x = (uint64_t)wc->freelist_low_watermark * 100;
2475                         x += wc->n_blocks / 2;
2476                         do_div(x, (size_t)wc->n_blocks);
2477                         DMEMIT(" low_watermark %u", 100 - (unsigned)x);
2478                 }
2479                 if (wc->max_writeback_jobs_set)
2480                         DMEMIT(" writeback_jobs %u", wc->max_writeback_jobs);
2481                 if (wc->autocommit_blocks_set)
2482                         DMEMIT(" autocommit_blocks %u", wc->autocommit_blocks);
2483                 if (wc->autocommit_time_set)
2484                         DMEMIT(" autocommit_time %u", jiffies_to_msecs(wc->autocommit_jiffies));
2485                 if (wc->max_age != MAX_AGE_UNSPECIFIED)
2486                         DMEMIT(" max_age %u", jiffies_to_msecs(wc->max_age));
2487                 if (wc->cleaner)
2488                         DMEMIT(" cleaner");
2489                 if (wc->writeback_fua_set)
2490                         DMEMIT(" %sfua", wc->writeback_fua ? "" : "no");
2491                 break;
2492         }
2493 }
2494
2495 static struct target_type writecache_target = {
2496         .name                   = "writecache",
2497         .version                = {1, 3, 0},
2498         .module                 = THIS_MODULE,
2499         .ctr                    = writecache_ctr,
2500         .dtr                    = writecache_dtr,
2501         .status                 = writecache_status,
2502         .postsuspend            = writecache_suspend,
2503         .resume                 = writecache_resume,
2504         .message                = writecache_message,
2505         .map                    = writecache_map,
2506         .end_io                 = writecache_end_io,
2507         .iterate_devices        = writecache_iterate_devices,
2508         .io_hints               = writecache_io_hints,
2509 };
2510
2511 static int __init dm_writecache_init(void)
2512 {
2513         int r;
2514
2515         r = dm_register_target(&writecache_target);
2516         if (r < 0) {
2517                 DMERR("register failed %d", r);
2518                 return r;
2519         }
2520
2521         return 0;
2522 }
2523
2524 static void __exit dm_writecache_exit(void)
2525 {
2526         dm_unregister_target(&writecache_target);
2527 }
2528
2529 module_init(dm_writecache_init);
2530 module_exit(dm_writecache_exit);
2531
2532 MODULE_DESCRIPTION(DM_NAME " writecache target");
2533 MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
2534 MODULE_LICENSE("GPL");