linux/drivers/md/dm-writecache.c
<<
>>
Prefs
   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)                                  \
  44do {                                                            \
  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 IS_ENABLED(CONFIG_ARCH_HAS_COPY_MC) && 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
  59struct wc_memory_entry {
  60        __le64 original_sector;
  61        __le64 seq_count;
  62};
  63
  64struct 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[];
  77};
  78
  79struct 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
 109struct 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        sector_t data_device_sectors;
 152        void *block_start;
 153        struct wc_entry *entries;
 154        unsigned block_size;
 155        unsigned char block_size_bits;
 156
 157        bool pmem_mode:1;
 158        bool writeback_fua:1;
 159
 160        bool overwrote_committed:1;
 161        bool memory_vmapped:1;
 162
 163        bool start_sector_set:1;
 164        bool high_wm_percent_set:1;
 165        bool low_wm_percent_set:1;
 166        bool max_writeback_jobs_set:1;
 167        bool autocommit_blocks_set:1;
 168        bool autocommit_time_set:1;
 169        bool max_age_set:1;
 170        bool writeback_fua_set:1;
 171        bool flush_on_suspend:1;
 172        bool cleaner:1;
 173        bool cleaner_set:1;
 174
 175        unsigned high_wm_percent_value;
 176        unsigned low_wm_percent_value;
 177        unsigned autocommit_time_value;
 178        unsigned max_age_value;
 179
 180        unsigned writeback_all;
 181        struct workqueue_struct *writeback_wq;
 182        struct work_struct writeback_work;
 183        struct work_struct flush_work;
 184
 185        struct dm_io_client *dm_io;
 186
 187        raw_spinlock_t endio_list_lock;
 188        struct list_head endio_list;
 189        struct task_struct *endio_thread;
 190
 191        struct task_struct *flush_thread;
 192        struct bio_list flush_list;
 193
 194        struct dm_kcopyd_client *dm_kcopyd;
 195        unsigned long *dirty_bitmap;
 196        unsigned dirty_bitmap_size;
 197
 198        struct bio_set bio_set;
 199        mempool_t copy_pool;
 200};
 201
 202#define WB_LIST_INLINE          16
 203
 204struct writeback_struct {
 205        struct list_head endio_entry;
 206        struct dm_writecache *wc;
 207        struct wc_entry **wc_list;
 208        unsigned wc_list_n;
 209        struct wc_entry *wc_list_inline[WB_LIST_INLINE];
 210        struct bio bio;
 211};
 212
 213struct copy_struct {
 214        struct list_head endio_entry;
 215        struct dm_writecache *wc;
 216        struct wc_entry *e;
 217        unsigned n_entries;
 218        int error;
 219};
 220
 221DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(dm_writecache_throttle,
 222                                            "A percentage of time allocated for data copying");
 223
 224static void wc_lock(struct dm_writecache *wc)
 225{
 226        mutex_lock(&wc->lock);
 227}
 228
 229static void wc_unlock(struct dm_writecache *wc)
 230{
 231        mutex_unlock(&wc->lock);
 232}
 233
 234#ifdef DM_WRITECACHE_HAS_PMEM
 235static int persistent_memory_claim(struct dm_writecache *wc)
 236{
 237        int r;
 238        loff_t s;
 239        long p, da;
 240        pfn_t pfn;
 241        int id;
 242        struct page **pages;
 243        sector_t offset;
 244
 245        wc->memory_vmapped = false;
 246
 247        s = wc->memory_map_size;
 248        p = s >> PAGE_SHIFT;
 249        if (!p) {
 250                r = -EINVAL;
 251                goto err1;
 252        }
 253        if (p != s >> PAGE_SHIFT) {
 254                r = -EOVERFLOW;
 255                goto err1;
 256        }
 257
 258        offset = get_start_sect(wc->ssd_dev->bdev);
 259        if (offset & (PAGE_SIZE / 512 - 1)) {
 260                r = -EINVAL;
 261                goto err1;
 262        }
 263        offset >>= PAGE_SHIFT - 9;
 264
 265        id = dax_read_lock();
 266
 267        da = dax_direct_access(wc->ssd_dev->dax_dev, offset, p, &wc->memory_map, &pfn);
 268        if (da < 0) {
 269                wc->memory_map = NULL;
 270                r = da;
 271                goto err2;
 272        }
 273        if (!pfn_t_has_page(pfn)) {
 274                wc->memory_map = NULL;
 275                r = -EOPNOTSUPP;
 276                goto err2;
 277        }
 278        if (da != p) {
 279                long i;
 280                wc->memory_map = NULL;
 281                pages = kvmalloc_array(p, sizeof(struct page *), GFP_KERNEL);
 282                if (!pages) {
 283                        r = -ENOMEM;
 284                        goto err2;
 285                }
 286                i = 0;
 287                do {
 288                        long daa;
 289                        daa = dax_direct_access(wc->ssd_dev->dax_dev, offset + i, p - i,
 290                                                NULL, &pfn);
 291                        if (daa <= 0) {
 292                                r = daa ? daa : -EINVAL;
 293                                goto err3;
 294                        }
 295                        if (!pfn_t_has_page(pfn)) {
 296                                r = -EOPNOTSUPP;
 297                                goto err3;
 298                        }
 299                        while (daa-- && i < p) {
 300                                pages[i++] = pfn_t_to_page(pfn);
 301                                pfn.val++;
 302                                if (!(i & 15))
 303                                        cond_resched();
 304                        }
 305                } while (i < p);
 306                wc->memory_map = vmap(pages, p, VM_MAP, PAGE_KERNEL);
 307                if (!wc->memory_map) {
 308                        r = -ENOMEM;
 309                        goto err3;
 310                }
 311                kvfree(pages);
 312                wc->memory_vmapped = true;
 313        }
 314
 315        dax_read_unlock(id);
 316
 317        wc->memory_map += (size_t)wc->start_sector << SECTOR_SHIFT;
 318        wc->memory_map_size -= (size_t)wc->start_sector << SECTOR_SHIFT;
 319
 320        return 0;
 321err3:
 322        kvfree(pages);
 323err2:
 324        dax_read_unlock(id);
 325err1:
 326        return r;
 327}
 328#else
 329static int persistent_memory_claim(struct dm_writecache *wc)
 330{
 331        return -EOPNOTSUPP;
 332}
 333#endif
 334
 335static void persistent_memory_release(struct dm_writecache *wc)
 336{
 337        if (wc->memory_vmapped)
 338                vunmap(wc->memory_map - ((size_t)wc->start_sector << SECTOR_SHIFT));
 339}
 340
 341static struct page *persistent_memory_page(void *addr)
 342{
 343        if (is_vmalloc_addr(addr))
 344                return vmalloc_to_page(addr);
 345        else
 346                return virt_to_page(addr);
 347}
 348
 349static unsigned persistent_memory_page_offset(void *addr)
 350{
 351        return (unsigned long)addr & (PAGE_SIZE - 1);
 352}
 353
 354static void persistent_memory_flush_cache(void *ptr, size_t size)
 355{
 356        if (is_vmalloc_addr(ptr))
 357                flush_kernel_vmap_range(ptr, size);
 358}
 359
 360static void persistent_memory_invalidate_cache(void *ptr, size_t size)
 361{
 362        if (is_vmalloc_addr(ptr))
 363                invalidate_kernel_vmap_range(ptr, size);
 364}
 365
 366static struct wc_memory_superblock *sb(struct dm_writecache *wc)
 367{
 368        return wc->memory_map;
 369}
 370
 371static struct wc_memory_entry *memory_entry(struct dm_writecache *wc, struct wc_entry *e)
 372{
 373        return &sb(wc)->entries[e->index];
 374}
 375
 376static void *memory_data(struct dm_writecache *wc, struct wc_entry *e)
 377{
 378        return (char *)wc->block_start + (e->index << wc->block_size_bits);
 379}
 380
 381static sector_t cache_sector(struct dm_writecache *wc, struct wc_entry *e)
 382{
 383        return wc->start_sector + wc->metadata_sectors +
 384                ((sector_t)e->index << (wc->block_size_bits - SECTOR_SHIFT));
 385}
 386
 387static uint64_t read_original_sector(struct dm_writecache *wc, struct wc_entry *e)
 388{
 389#ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 390        return e->original_sector;
 391#else
 392        return le64_to_cpu(memory_entry(wc, e)->original_sector);
 393#endif
 394}
 395
 396static uint64_t read_seq_count(struct dm_writecache *wc, struct wc_entry *e)
 397{
 398#ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 399        return e->seq_count;
 400#else
 401        return le64_to_cpu(memory_entry(wc, e)->seq_count);
 402#endif
 403}
 404
 405static void clear_seq_count(struct dm_writecache *wc, struct wc_entry *e)
 406{
 407#ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 408        e->seq_count = -1;
 409#endif
 410        pmem_assign(memory_entry(wc, e)->seq_count, cpu_to_le64(-1));
 411}
 412
 413static void write_original_sector_seq_count(struct dm_writecache *wc, struct wc_entry *e,
 414                                            uint64_t original_sector, uint64_t seq_count)
 415{
 416        struct wc_memory_entry me;
 417#ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
 418        e->original_sector = original_sector;
 419        e->seq_count = seq_count;
 420#endif
 421        me.original_sector = cpu_to_le64(original_sector);
 422        me.seq_count = cpu_to_le64(seq_count);
 423        pmem_assign(*memory_entry(wc, e), me);
 424}
 425
 426#define writecache_error(wc, err, msg, arg...)                          \
 427do {                                                                    \
 428        if (!cmpxchg(&(wc)->error, 0, err))                             \
 429                DMERR(msg, ##arg);                                      \
 430        wake_up(&(wc)->freelist_wait);                                  \
 431} while (0)
 432
 433#define writecache_has_error(wc)        (unlikely(READ_ONCE((wc)->error)))
 434
 435static void writecache_flush_all_metadata(struct dm_writecache *wc)
 436{
 437        if (!WC_MODE_PMEM(wc))
 438                memset(wc->dirty_bitmap, -1, wc->dirty_bitmap_size);
 439}
 440
 441static void writecache_flush_region(struct dm_writecache *wc, void *ptr, size_t size)
 442{
 443        if (!WC_MODE_PMEM(wc))
 444                __set_bit(((char *)ptr - (char *)wc->memory_map) / BITMAP_GRANULARITY,
 445                          wc->dirty_bitmap);
 446}
 447
 448static void writecache_disk_flush(struct dm_writecache *wc, struct dm_dev *dev);
 449
 450struct io_notify {
 451        struct dm_writecache *wc;
 452        struct completion c;
 453        atomic_t count;
 454};
 455
 456static void writecache_notify_io(unsigned long error, void *context)
 457{
 458        struct io_notify *endio = context;
 459
 460        if (unlikely(error != 0))
 461                writecache_error(endio->wc, -EIO, "error writing metadata");
 462        BUG_ON(atomic_read(&endio->count) <= 0);
 463        if (atomic_dec_and_test(&endio->count))
 464                complete(&endio->c);
 465}
 466
 467static void writecache_wait_for_ios(struct dm_writecache *wc, int direction)
 468{
 469        wait_event(wc->bio_in_progress_wait[direction],
 470                   !atomic_read(&wc->bio_in_progress[direction]));
 471}
 472
 473static void ssd_commit_flushed(struct dm_writecache *wc, bool wait_for_ios)
 474{
 475        struct dm_io_region region;
 476        struct dm_io_request req;
 477        struct io_notify endio = {
 478                wc,
 479                COMPLETION_INITIALIZER_ONSTACK(endio.c),
 480                ATOMIC_INIT(1),
 481        };
 482        unsigned bitmap_bits = wc->dirty_bitmap_size * 8;
 483        unsigned i = 0;
 484
 485        while (1) {
 486                unsigned j;
 487                i = find_next_bit(wc->dirty_bitmap, bitmap_bits, i);
 488                if (unlikely(i == bitmap_bits))
 489                        break;
 490                j = find_next_zero_bit(wc->dirty_bitmap, bitmap_bits, i);
 491
 492                region.bdev = wc->ssd_dev->bdev;
 493                region.sector = (sector_t)i * (BITMAP_GRANULARITY >> SECTOR_SHIFT);
 494                region.count = (sector_t)(j - i) * (BITMAP_GRANULARITY >> SECTOR_SHIFT);
 495
 496                if (unlikely(region.sector >= wc->metadata_sectors))
 497                        break;
 498                if (unlikely(region.sector + region.count > wc->metadata_sectors))
 499                        region.count = wc->metadata_sectors - region.sector;
 500
 501                region.sector += wc->start_sector;
 502                atomic_inc(&endio.count);
 503                req.bi_op = REQ_OP_WRITE;
 504                req.bi_op_flags = REQ_SYNC;
 505                req.mem.type = DM_IO_VMA;
 506                req.mem.ptr.vma = (char *)wc->memory_map + (size_t)i * BITMAP_GRANULARITY;
 507                req.client = wc->dm_io;
 508                req.notify.fn = writecache_notify_io;
 509                req.notify.context = &endio;
 510
 511                /* writing via async dm-io (implied by notify.fn above) won't return an error */
 512                (void) dm_io(&req, 1, &region, NULL);
 513                i = j;
 514        }
 515
 516        writecache_notify_io(0, &endio);
 517        wait_for_completion_io(&endio.c);
 518
 519        if (wait_for_ios)
 520                writecache_wait_for_ios(wc, WRITE);
 521
 522        writecache_disk_flush(wc, wc->ssd_dev);
 523
 524        memset(wc->dirty_bitmap, 0, wc->dirty_bitmap_size);
 525}
 526
 527static void ssd_commit_superblock(struct dm_writecache *wc)
 528{
 529        int r;
 530        struct dm_io_region region;
 531        struct dm_io_request req;
 532
 533        region.bdev = wc->ssd_dev->bdev;
 534        region.sector = 0;
 535        region.count = PAGE_SIZE >> SECTOR_SHIFT;
 536
 537        if (unlikely(region.sector + region.count > wc->metadata_sectors))
 538                region.count = wc->metadata_sectors - region.sector;
 539
 540        region.sector += wc->start_sector;
 541
 542        req.bi_op = REQ_OP_WRITE;
 543        req.bi_op_flags = REQ_SYNC | REQ_FUA;
 544        req.mem.type = DM_IO_VMA;
 545        req.mem.ptr.vma = (char *)wc->memory_map;
 546        req.client = wc->dm_io;
 547        req.notify.fn = NULL;
 548        req.notify.context = NULL;
 549
 550        r = dm_io(&req, 1, &region, NULL);
 551        if (unlikely(r))
 552                writecache_error(wc, r, "error writing superblock");
 553}
 554
 555static void writecache_commit_flushed(struct dm_writecache *wc, bool wait_for_ios)
 556{
 557        if (WC_MODE_PMEM(wc))
 558                pmem_wmb();
 559        else
 560                ssd_commit_flushed(wc, wait_for_ios);
 561}
 562
 563static void writecache_disk_flush(struct dm_writecache *wc, struct dm_dev *dev)
 564{
 565        int r;
 566        struct dm_io_region region;
 567        struct dm_io_request req;
 568
 569        region.bdev = dev->bdev;
 570        region.sector = 0;
 571        region.count = 0;
 572        req.bi_op = REQ_OP_WRITE;
 573        req.bi_op_flags = REQ_PREFLUSH;
 574        req.mem.type = DM_IO_KMEM;
 575        req.mem.ptr.addr = NULL;
 576        req.client = wc->dm_io;
 577        req.notify.fn = NULL;
 578
 579        r = dm_io(&req, 1, &region, NULL);
 580        if (unlikely(r))
 581                writecache_error(wc, r, "error flushing metadata: %d", r);
 582}
 583
 584#define WFE_RETURN_FOLLOWING    1
 585#define WFE_LOWEST_SEQ          2
 586
 587static struct wc_entry *writecache_find_entry(struct dm_writecache *wc,
 588                                              uint64_t block, int flags)
 589{
 590        struct wc_entry *e;
 591        struct rb_node *node = wc->tree.rb_node;
 592
 593        if (unlikely(!node))
 594                return NULL;
 595
 596        while (1) {
 597                e = container_of(node, struct wc_entry, rb_node);
 598                if (read_original_sector(wc, e) == block)
 599                        break;
 600
 601                node = (read_original_sector(wc, e) >= block ?
 602                        e->rb_node.rb_left : e->rb_node.rb_right);
 603                if (unlikely(!node)) {
 604                        if (!(flags & WFE_RETURN_FOLLOWING))
 605                                return NULL;
 606                        if (read_original_sector(wc, e) >= block) {
 607                                return e;
 608                        } else {
 609                                node = rb_next(&e->rb_node);
 610                                if (unlikely(!node))
 611                                        return NULL;
 612                                e = container_of(node, struct wc_entry, rb_node);
 613                                return e;
 614                        }
 615                }
 616        }
 617
 618        while (1) {
 619                struct wc_entry *e2;
 620                if (flags & WFE_LOWEST_SEQ)
 621                        node = rb_prev(&e->rb_node);
 622                else
 623                        node = rb_next(&e->rb_node);
 624                if (unlikely(!node))
 625                        return e;
 626                e2 = container_of(node, struct wc_entry, rb_node);
 627                if (read_original_sector(wc, e2) != block)
 628                        return e;
 629                e = e2;
 630        }
 631}
 632
 633static void writecache_insert_entry(struct dm_writecache *wc, struct wc_entry *ins)
 634{
 635        struct wc_entry *e;
 636        struct rb_node **node = &wc->tree.rb_node, *parent = NULL;
 637
 638        while (*node) {
 639                e = container_of(*node, struct wc_entry, rb_node);
 640                parent = &e->rb_node;
 641                if (read_original_sector(wc, e) > read_original_sector(wc, ins))
 642                        node = &parent->rb_left;
 643                else
 644                        node = &parent->rb_right;
 645        }
 646        rb_link_node(&ins->rb_node, parent, node);
 647        rb_insert_color(&ins->rb_node, &wc->tree);
 648        list_add(&ins->lru, &wc->lru);
 649        ins->age = jiffies;
 650}
 651
 652static void writecache_unlink(struct dm_writecache *wc, struct wc_entry *e)
 653{
 654        list_del(&e->lru);
 655        rb_erase(&e->rb_node, &wc->tree);
 656}
 657
 658static void writecache_add_to_freelist(struct dm_writecache *wc, struct wc_entry *e)
 659{
 660        if (WC_MODE_SORT_FREELIST(wc)) {
 661                struct rb_node **node = &wc->freetree.rb_node, *parent = NULL;
 662                if (unlikely(!*node))
 663                        wc->current_free = e;
 664                while (*node) {
 665                        parent = *node;
 666                        if (&e->rb_node < *node)
 667                                node = &parent->rb_left;
 668                        else
 669                                node = &parent->rb_right;
 670                }
 671                rb_link_node(&e->rb_node, parent, node);
 672                rb_insert_color(&e->rb_node, &wc->freetree);
 673        } else {
 674                list_add_tail(&e->lru, &wc->freelist);
 675        }
 676        wc->freelist_size++;
 677}
 678
 679static inline void writecache_verify_watermark(struct dm_writecache *wc)
 680{
 681        if (unlikely(wc->freelist_size + wc->writeback_size <= wc->freelist_high_watermark))
 682                queue_work(wc->writeback_wq, &wc->writeback_work);
 683}
 684
 685static void writecache_max_age_timer(struct timer_list *t)
 686{
 687        struct dm_writecache *wc = from_timer(wc, t, max_age_timer);
 688
 689        if (!dm_suspended(wc->ti) && !writecache_has_error(wc)) {
 690                queue_work(wc->writeback_wq, &wc->writeback_work);
 691                mod_timer(&wc->max_age_timer, jiffies + wc->max_age / MAX_AGE_DIV);
 692        }
 693}
 694
 695static struct wc_entry *writecache_pop_from_freelist(struct dm_writecache *wc, sector_t expected_sector)
 696{
 697        struct wc_entry *e;
 698
 699        if (WC_MODE_SORT_FREELIST(wc)) {
 700                struct rb_node *next;
 701                if (unlikely(!wc->current_free))
 702                        return NULL;
 703                e = wc->current_free;
 704                if (expected_sector != (sector_t)-1 && unlikely(cache_sector(wc, e) != expected_sector))
 705                        return NULL;
 706                next = rb_next(&e->rb_node);
 707                rb_erase(&e->rb_node, &wc->freetree);
 708                if (unlikely(!next))
 709                        next = rb_first(&wc->freetree);
 710                wc->current_free = next ? container_of(next, struct wc_entry, rb_node) : NULL;
 711        } else {
 712                if (unlikely(list_empty(&wc->freelist)))
 713                        return NULL;
 714                e = container_of(wc->freelist.next, struct wc_entry, lru);
 715                if (expected_sector != (sector_t)-1 && unlikely(cache_sector(wc, e) != expected_sector))
 716                        return NULL;
 717                list_del(&e->lru);
 718        }
 719        wc->freelist_size--;
 720
 721        writecache_verify_watermark(wc);
 722
 723        return e;
 724}
 725
 726static void writecache_free_entry(struct dm_writecache *wc, struct wc_entry *e)
 727{
 728        writecache_unlink(wc, e);
 729        writecache_add_to_freelist(wc, e);
 730        clear_seq_count(wc, e);
 731        writecache_flush_region(wc, memory_entry(wc, e), sizeof(struct wc_memory_entry));
 732        if (unlikely(waitqueue_active(&wc->freelist_wait)))
 733                wake_up(&wc->freelist_wait);
 734}
 735
 736static void writecache_wait_on_freelist(struct dm_writecache *wc)
 737{
 738        DEFINE_WAIT(wait);
 739
 740        prepare_to_wait(&wc->freelist_wait, &wait, TASK_UNINTERRUPTIBLE);
 741        wc_unlock(wc);
 742        io_schedule();
 743        finish_wait(&wc->freelist_wait, &wait);
 744        wc_lock(wc);
 745}
 746
 747static void writecache_poison_lists(struct dm_writecache *wc)
 748{
 749        /*
 750         * Catch incorrect access to these values while the device is suspended.
 751         */
 752        memset(&wc->tree, -1, sizeof wc->tree);
 753        wc->lru.next = LIST_POISON1;
 754        wc->lru.prev = LIST_POISON2;
 755        wc->freelist.next = LIST_POISON1;
 756        wc->freelist.prev = LIST_POISON2;
 757}
 758
 759static void writecache_flush_entry(struct dm_writecache *wc, struct wc_entry *e)
 760{
 761        writecache_flush_region(wc, memory_entry(wc, e), sizeof(struct wc_memory_entry));
 762        if (WC_MODE_PMEM(wc))
 763                writecache_flush_region(wc, memory_data(wc, e), wc->block_size);
 764}
 765
 766static bool writecache_entry_is_committed(struct dm_writecache *wc, struct wc_entry *e)
 767{
 768        return read_seq_count(wc, e) < wc->seq_count;
 769}
 770
 771static void writecache_flush(struct dm_writecache *wc)
 772{
 773        struct wc_entry *e, *e2;
 774        bool need_flush_after_free;
 775
 776        wc->uncommitted_blocks = 0;
 777        del_timer(&wc->autocommit_timer);
 778
 779        if (list_empty(&wc->lru))
 780                return;
 781
 782        e = container_of(wc->lru.next, struct wc_entry, lru);
 783        if (writecache_entry_is_committed(wc, e)) {
 784                if (wc->overwrote_committed) {
 785                        writecache_wait_for_ios(wc, WRITE);
 786                        writecache_disk_flush(wc, wc->ssd_dev);
 787                        wc->overwrote_committed = false;
 788                }
 789                return;
 790        }
 791        while (1) {
 792                writecache_flush_entry(wc, e);
 793                if (unlikely(e->lru.next == &wc->lru))
 794                        break;
 795                e2 = container_of(e->lru.next, struct wc_entry, lru);
 796                if (writecache_entry_is_committed(wc, e2))
 797                        break;
 798                e = e2;
 799                cond_resched();
 800        }
 801        writecache_commit_flushed(wc, true);
 802
 803        wc->seq_count++;
 804        pmem_assign(sb(wc)->seq_count, cpu_to_le64(wc->seq_count));
 805        if (WC_MODE_PMEM(wc))
 806                writecache_commit_flushed(wc, false);
 807        else
 808                ssd_commit_superblock(wc);
 809
 810        wc->overwrote_committed = false;
 811
 812        need_flush_after_free = false;
 813        while (1) {
 814                /* Free another committed entry with lower seq-count */
 815                struct rb_node *rb_node = rb_prev(&e->rb_node);
 816
 817                if (rb_node) {
 818                        e2 = container_of(rb_node, struct wc_entry, rb_node);
 819                        if (read_original_sector(wc, e2) == read_original_sector(wc, e) &&
 820                            likely(!e2->write_in_progress)) {
 821                                writecache_free_entry(wc, e2);
 822                                need_flush_after_free = true;
 823                        }
 824                }
 825                if (unlikely(e->lru.prev == &wc->lru))
 826                        break;
 827                e = container_of(e->lru.prev, struct wc_entry, lru);
 828                cond_resched();
 829        }
 830
 831        if (need_flush_after_free)
 832                writecache_commit_flushed(wc, false);
 833}
 834
 835static void writecache_flush_work(struct work_struct *work)
 836{
 837        struct dm_writecache *wc = container_of(work, struct dm_writecache, flush_work);
 838
 839        wc_lock(wc);
 840        writecache_flush(wc);
 841        wc_unlock(wc);
 842}
 843
 844static void writecache_autocommit_timer(struct timer_list *t)
 845{
 846        struct dm_writecache *wc = from_timer(wc, t, autocommit_timer);
 847        if (!writecache_has_error(wc))
 848                queue_work(wc->writeback_wq, &wc->flush_work);
 849}
 850
 851static void writecache_schedule_autocommit(struct dm_writecache *wc)
 852{
 853        if (!timer_pending(&wc->autocommit_timer))
 854                mod_timer(&wc->autocommit_timer, jiffies + wc->autocommit_jiffies);
 855}
 856
 857static void writecache_discard(struct dm_writecache *wc, sector_t start, sector_t end)
 858{
 859        struct wc_entry *e;
 860        bool discarded_something = false;
 861
 862        e = writecache_find_entry(wc, start, WFE_RETURN_FOLLOWING | WFE_LOWEST_SEQ);
 863        if (unlikely(!e))
 864                return;
 865
 866        while (read_original_sector(wc, e) < end) {
 867                struct rb_node *node = rb_next(&e->rb_node);
 868
 869                if (likely(!e->write_in_progress)) {
 870                        if (!discarded_something) {
 871                                if (!WC_MODE_PMEM(wc)) {
 872                                        writecache_wait_for_ios(wc, READ);
 873                                        writecache_wait_for_ios(wc, WRITE);
 874                                }
 875                                discarded_something = true;
 876                        }
 877                        if (!writecache_entry_is_committed(wc, e))
 878                                wc->uncommitted_blocks--;
 879                        writecache_free_entry(wc, e);
 880                }
 881
 882                if (unlikely(!node))
 883                        break;
 884
 885                e = container_of(node, struct wc_entry, rb_node);
 886        }
 887
 888        if (discarded_something)
 889                writecache_commit_flushed(wc, false);
 890}
 891
 892static bool writecache_wait_for_writeback(struct dm_writecache *wc)
 893{
 894        if (wc->writeback_size) {
 895                writecache_wait_on_freelist(wc);
 896                return true;
 897        }
 898        return false;
 899}
 900
 901static void writecache_suspend(struct dm_target *ti)
 902{
 903        struct dm_writecache *wc = ti->private;
 904        bool flush_on_suspend;
 905
 906        del_timer_sync(&wc->autocommit_timer);
 907        del_timer_sync(&wc->max_age_timer);
 908
 909        wc_lock(wc);
 910        writecache_flush(wc);
 911        flush_on_suspend = wc->flush_on_suspend;
 912        if (flush_on_suspend) {
 913                wc->flush_on_suspend = false;
 914                wc->writeback_all++;
 915                queue_work(wc->writeback_wq, &wc->writeback_work);
 916        }
 917        wc_unlock(wc);
 918
 919        drain_workqueue(wc->writeback_wq);
 920
 921        wc_lock(wc);
 922        if (flush_on_suspend)
 923                wc->writeback_all--;
 924        while (writecache_wait_for_writeback(wc));
 925
 926        if (WC_MODE_PMEM(wc))
 927                persistent_memory_flush_cache(wc->memory_map, wc->memory_map_size);
 928
 929        writecache_poison_lists(wc);
 930
 931        wc_unlock(wc);
 932}
 933
 934static int writecache_alloc_entries(struct dm_writecache *wc)
 935{
 936        size_t b;
 937
 938        if (wc->entries)
 939                return 0;
 940        wc->entries = vmalloc(array_size(sizeof(struct wc_entry), wc->n_blocks));
 941        if (!wc->entries)
 942                return -ENOMEM;
 943        for (b = 0; b < wc->n_blocks; b++) {
 944                struct wc_entry *e = &wc->entries[b];
 945                e->index = b;
 946                e->write_in_progress = false;
 947                cond_resched();
 948        }
 949
 950        return 0;
 951}
 952
 953static int writecache_read_metadata(struct dm_writecache *wc, sector_t n_sectors)
 954{
 955        struct dm_io_region region;
 956        struct dm_io_request req;
 957
 958        region.bdev = wc->ssd_dev->bdev;
 959        region.sector = wc->start_sector;
 960        region.count = n_sectors;
 961        req.bi_op = REQ_OP_READ;
 962        req.bi_op_flags = REQ_SYNC;
 963        req.mem.type = DM_IO_VMA;
 964        req.mem.ptr.vma = (char *)wc->memory_map;
 965        req.client = wc->dm_io;
 966        req.notify.fn = NULL;
 967
 968        return dm_io(&req, 1, &region, NULL);
 969}
 970
 971static void writecache_resume(struct dm_target *ti)
 972{
 973        struct dm_writecache *wc = ti->private;
 974        size_t b;
 975        bool need_flush = false;
 976        __le64 sb_seq_count;
 977        int r;
 978
 979        wc_lock(wc);
 980
 981        wc->data_device_sectors = bdev_nr_sectors(wc->dev->bdev);
 982
 983        if (WC_MODE_PMEM(wc)) {
 984                persistent_memory_invalidate_cache(wc->memory_map, wc->memory_map_size);
 985        } else {
 986                r = writecache_read_metadata(wc, wc->metadata_sectors);
 987                if (r) {
 988                        size_t sb_entries_offset;
 989                        writecache_error(wc, r, "unable to read metadata: %d", r);
 990                        sb_entries_offset = offsetof(struct wc_memory_superblock, entries);
 991                        memset((char *)wc->memory_map + sb_entries_offset, -1,
 992                               (wc->metadata_sectors << SECTOR_SHIFT) - sb_entries_offset);
 993                }
 994        }
 995
 996        wc->tree = RB_ROOT;
 997        INIT_LIST_HEAD(&wc->lru);
 998        if (WC_MODE_SORT_FREELIST(wc)) {
 999                wc->freetree = RB_ROOT;
1000                wc->current_free = NULL;
1001        } else {
1002                INIT_LIST_HEAD(&wc->freelist);
1003        }
1004        wc->freelist_size = 0;
1005
1006        r = copy_mc_to_kernel(&sb_seq_count, &sb(wc)->seq_count,
1007                              sizeof(uint64_t));
1008        if (r) {
1009                writecache_error(wc, r, "hardware memory error when reading superblock: %d", r);
1010                sb_seq_count = cpu_to_le64(0);
1011        }
1012        wc->seq_count = le64_to_cpu(sb_seq_count);
1013
1014#ifdef DM_WRITECACHE_HANDLE_HARDWARE_ERRORS
1015        for (b = 0; b < wc->n_blocks; b++) {
1016                struct wc_entry *e = &wc->entries[b];
1017                struct wc_memory_entry wme;
1018                if (writecache_has_error(wc)) {
1019                        e->original_sector = -1;
1020                        e->seq_count = -1;
1021                        continue;
1022                }
1023                r = copy_mc_to_kernel(&wme, memory_entry(wc, e),
1024                                      sizeof(struct wc_memory_entry));
1025                if (r) {
1026                        writecache_error(wc, r, "hardware memory error when reading metadata entry %lu: %d",
1027                                         (unsigned long)b, r);
1028                        e->original_sector = -1;
1029                        e->seq_count = -1;
1030                } else {
1031                        e->original_sector = le64_to_cpu(wme.original_sector);
1032                        e->seq_count = le64_to_cpu(wme.seq_count);
1033                }
1034                cond_resched();
1035        }
1036#endif
1037        for (b = 0; b < wc->n_blocks; b++) {
1038                struct wc_entry *e = &wc->entries[b];
1039                if (!writecache_entry_is_committed(wc, e)) {
1040                        if (read_seq_count(wc, e) != -1) {
1041erase_this:
1042                                clear_seq_count(wc, e);
1043                                need_flush = true;
1044                        }
1045                        writecache_add_to_freelist(wc, e);
1046                } else {
1047                        struct wc_entry *old;
1048
1049                        old = writecache_find_entry(wc, read_original_sector(wc, e), 0);
1050                        if (!old) {
1051                                writecache_insert_entry(wc, e);
1052                        } else {
1053                                if (read_seq_count(wc, old) == read_seq_count(wc, e)) {
1054                                        writecache_error(wc, -EINVAL,
1055                                                 "two identical entries, position %llu, sector %llu, sequence %llu",
1056                                                 (unsigned long long)b, (unsigned long long)read_original_sector(wc, e),
1057                                                 (unsigned long long)read_seq_count(wc, e));
1058                                }
1059                                if (read_seq_count(wc, old) > read_seq_count(wc, e)) {
1060                                        goto erase_this;
1061                                } else {
1062                                        writecache_free_entry(wc, old);
1063                                        writecache_insert_entry(wc, e);
1064                                        need_flush = true;
1065                                }
1066                        }
1067                }
1068                cond_resched();
1069        }
1070
1071        if (need_flush) {
1072                writecache_flush_all_metadata(wc);
1073                writecache_commit_flushed(wc, false);
1074        }
1075
1076        writecache_verify_watermark(wc);
1077
1078        if (wc->max_age != MAX_AGE_UNSPECIFIED)
1079                mod_timer(&wc->max_age_timer, jiffies + wc->max_age / MAX_AGE_DIV);
1080
1081        wc_unlock(wc);
1082}
1083
1084static int process_flush_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1085{
1086        if (argc != 1)
1087                return -EINVAL;
1088
1089        wc_lock(wc);
1090        if (dm_suspended(wc->ti)) {
1091                wc_unlock(wc);
1092                return -EBUSY;
1093        }
1094        if (writecache_has_error(wc)) {
1095                wc_unlock(wc);
1096                return -EIO;
1097        }
1098
1099        writecache_flush(wc);
1100        wc->writeback_all++;
1101        queue_work(wc->writeback_wq, &wc->writeback_work);
1102        wc_unlock(wc);
1103
1104        flush_workqueue(wc->writeback_wq);
1105
1106        wc_lock(wc);
1107        wc->writeback_all--;
1108        if (writecache_has_error(wc)) {
1109                wc_unlock(wc);
1110                return -EIO;
1111        }
1112        wc_unlock(wc);
1113
1114        return 0;
1115}
1116
1117static int process_flush_on_suspend_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1118{
1119        if (argc != 1)
1120                return -EINVAL;
1121
1122        wc_lock(wc);
1123        wc->flush_on_suspend = true;
1124        wc_unlock(wc);
1125
1126        return 0;
1127}
1128
1129static void activate_cleaner(struct dm_writecache *wc)
1130{
1131        wc->flush_on_suspend = true;
1132        wc->cleaner = true;
1133        wc->freelist_high_watermark = wc->n_blocks;
1134        wc->freelist_low_watermark = wc->n_blocks;
1135}
1136
1137static int process_cleaner_mesg(unsigned argc, char **argv, struct dm_writecache *wc)
1138{
1139        if (argc != 1)
1140                return -EINVAL;
1141
1142        wc_lock(wc);
1143        activate_cleaner(wc);
1144        if (!dm_suspended(wc->ti))
1145                writecache_verify_watermark(wc);
1146        wc_unlock(wc);
1147
1148        return 0;
1149}
1150
1151static int writecache_message(struct dm_target *ti, unsigned argc, char **argv,
1152                              char *result, unsigned maxlen)
1153{
1154        int r = -EINVAL;
1155        struct dm_writecache *wc = ti->private;
1156
1157        if (!strcasecmp(argv[0], "flush"))
1158                r = process_flush_mesg(argc, argv, wc);
1159        else if (!strcasecmp(argv[0], "flush_on_suspend"))
1160                r = process_flush_on_suspend_mesg(argc, argv, wc);
1161        else if (!strcasecmp(argv[0], "cleaner"))
1162                r = process_cleaner_mesg(argc, argv, wc);
1163        else
1164                DMERR("unrecognised message received: %s", argv[0]);
1165
1166        return r;
1167}
1168
1169static void memcpy_flushcache_optimized(void *dest, void *source, size_t size)
1170{
1171        /*
1172         * clflushopt performs better with block size 1024, 2048, 4096
1173         * non-temporal stores perform better with block size 512
1174         *
1175         * block size   512             1024            2048            4096
1176         * movnti       496 MB/s        642 MB/s        725 MB/s        744 MB/s
1177         * clflushopt   373 MB/s        688 MB/s        1.1 GB/s        1.2 GB/s
1178         *
1179         * We see that movnti performs better for 512-byte blocks, and
1180         * clflushopt performs better for 1024-byte and larger blocks. So, we
1181         * prefer clflushopt for sizes >= 768.
1182         *
1183         * NOTE: this happens to be the case now (with dm-writecache's single
1184         * threaded model) but re-evaluate this once memcpy_flushcache() is
1185         * enabled to use movdir64b which might invalidate this performance
1186         * advantage seen with cache-allocating-writes plus flushing.
1187         */
1188#ifdef CONFIG_X86
1189        if (static_cpu_has(X86_FEATURE_CLFLUSHOPT) &&
1190            likely(boot_cpu_data.x86_clflush_size == 64) &&
1191            likely(size >= 768)) {
1192                do {
1193                        memcpy((void *)dest, (void *)source, 64);
1194                        clflushopt((void *)dest);
1195                        dest += 64;
1196                        source += 64;
1197                        size -= 64;
1198                } while (size >= 64);
1199                return;
1200        }
1201#endif
1202        memcpy_flushcache(dest, source, size);
1203}
1204
1205static void bio_copy_block(struct dm_writecache *wc, struct bio *bio, void *data)
1206{
1207        void *buf;
1208        unsigned long flags;
1209        unsigned size;
1210        int rw = bio_data_dir(bio);
1211        unsigned remaining_size = wc->block_size;
1212
1213        do {
1214                struct bio_vec bv = bio_iter_iovec(bio, bio->bi_iter);
1215                buf = bvec_kmap_irq(&bv, &flags);
1216                size = bv.bv_len;
1217                if (unlikely(size > remaining_size))
1218                        size = remaining_size;
1219
1220                if (rw == READ) {
1221                        int r;
1222                        r = copy_mc_to_kernel(buf, data, size);
1223                        flush_dcache_page(bio_page(bio));
1224                        if (unlikely(r)) {
1225                                writecache_error(wc, r, "hardware memory error when reading data: %d", r);
1226                                bio->bi_status = BLK_STS_IOERR;
1227                        }
1228                } else {
1229                        flush_dcache_page(bio_page(bio));
1230                        memcpy_flushcache_optimized(data, buf, size);
1231                }
1232
1233                bvec_kunmap_irq(buf, &flags);
1234
1235                data = (char *)data + size;
1236                remaining_size -= size;
1237                bio_advance(bio, size);
1238        } while (unlikely(remaining_size));
1239}
1240
1241static int writecache_flush_thread(void *data)
1242{
1243        struct dm_writecache *wc = data;
1244
1245        while (1) {
1246                struct bio *bio;
1247
1248                wc_lock(wc);
1249                bio = bio_list_pop(&wc->flush_list);
1250                if (!bio) {
1251                        set_current_state(TASK_INTERRUPTIBLE);
1252                        wc_unlock(wc);
1253
1254                        if (unlikely(kthread_should_stop())) {
1255                                set_current_state(TASK_RUNNING);
1256                                break;
1257                        }
1258
1259                        schedule();
1260                        continue;
1261                }
1262
1263                if (bio_op(bio) == REQ_OP_DISCARD) {
1264                        writecache_discard(wc, bio->bi_iter.bi_sector,
1265                                           bio_end_sector(bio));
1266                        wc_unlock(wc);
1267                        bio_set_dev(bio, wc->dev->bdev);
1268                        submit_bio_noacct(bio);
1269                } else {
1270                        writecache_flush(wc);
1271                        wc_unlock(wc);
1272                        if (writecache_has_error(wc))
1273                                bio->bi_status = BLK_STS_IOERR;
1274                        bio_endio(bio);
1275                }
1276        }
1277
1278        return 0;
1279}
1280
1281static void writecache_offload_bio(struct dm_writecache *wc, struct bio *bio)
1282{
1283        if (bio_list_empty(&wc->flush_list))
1284                wake_up_process(wc->flush_thread);
1285        bio_list_add(&wc->flush_list, bio);
1286}
1287
1288static int writecache_map(struct dm_target *ti, struct bio *bio)
1289{
1290        struct wc_entry *e;
1291        struct dm_writecache *wc = ti->private;
1292
1293        bio->bi_private = NULL;
1294
1295        wc_lock(wc);
1296
1297        if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1298                if (writecache_has_error(wc))
1299                        goto unlock_error;
1300                if (WC_MODE_PMEM(wc)) {
1301                        writecache_flush(wc);
1302                        if (writecache_has_error(wc))
1303                                goto unlock_error;
1304                        goto unlock_submit;
1305                } else {
1306                        writecache_offload_bio(wc, bio);
1307                        goto unlock_return;
1308                }
1309        }
1310
1311        bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
1312
1313        if (unlikely((((unsigned)bio->bi_iter.bi_sector | bio_sectors(bio)) &
1314                                (wc->block_size / 512 - 1)) != 0)) {
1315                DMERR("I/O is not aligned, sector %llu, size %u, block size %u",
1316                      (unsigned long long)bio->bi_iter.bi_sector,
1317                      bio->bi_iter.bi_size, wc->block_size);
1318                goto unlock_error;
1319        }
1320
1321        if (unlikely(bio_op(bio) == REQ_OP_DISCARD)) {
1322                if (writecache_has_error(wc))
1323                        goto unlock_error;
1324                if (WC_MODE_PMEM(wc)) {
1325                        writecache_discard(wc, bio->bi_iter.bi_sector, bio_end_sector(bio));
1326                        goto unlock_remap_origin;
1327                } else {
1328                        writecache_offload_bio(wc, bio);
1329                        goto unlock_return;
1330                }
1331        }
1332
1333        if (bio_data_dir(bio) == READ) {
1334read_next_block:
1335                e = writecache_find_entry(wc, bio->bi_iter.bi_sector, WFE_RETURN_FOLLOWING);
1336                if (e && read_original_sector(wc, e) == bio->bi_iter.bi_sector) {
1337                        if (WC_MODE_PMEM(wc)) {
1338                                bio_copy_block(wc, bio, memory_data(wc, e));
1339                                if (bio->bi_iter.bi_size)
1340                                        goto read_next_block;
1341                                goto unlock_submit;
1342                        } else {
1343                                dm_accept_partial_bio(bio, wc->block_size >> SECTOR_SHIFT);
1344                                bio_set_dev(bio, wc->ssd_dev->bdev);
1345                                bio->bi_iter.bi_sector = cache_sector(wc, e);
1346                                if (!writecache_entry_is_committed(wc, e))
1347                                        writecache_wait_for_ios(wc, WRITE);
1348                                goto unlock_remap;
1349                        }
1350                } else {
1351                        if (e) {
1352                                sector_t next_boundary =
1353                                        read_original_sector(wc, e) - bio->bi_iter.bi_sector;
1354                                if (next_boundary < bio->bi_iter.bi_size >> SECTOR_SHIFT) {
1355                                        dm_accept_partial_bio(bio, next_boundary);
1356                                }
1357                        }
1358                        goto unlock_remap_origin;
1359                }
1360        } else {
1361                do {
1362                        bool found_entry = false;
1363                        if (writecache_has_error(wc))
1364                                goto unlock_error;
1365                        e = writecache_find_entry(wc, bio->bi_iter.bi_sector, 0);
1366                        if (e) {
1367                                if (!writecache_entry_is_committed(wc, e))
1368                                        goto bio_copy;
1369                                if (!WC_MODE_PMEM(wc) && !e->write_in_progress) {
1370                                        wc->overwrote_committed = true;
1371                                        goto bio_copy;
1372                                }
1373                                found_entry = true;
1374                        } else {
1375                                if (unlikely(wc->cleaner))
1376                                        goto direct_write;
1377                        }
1378                        e = writecache_pop_from_freelist(wc, (sector_t)-1);
1379                        if (unlikely(!e)) {
1380                                if (!found_entry) {
1381direct_write:
1382                                        e = writecache_find_entry(wc, bio->bi_iter.bi_sector, WFE_RETURN_FOLLOWING);
1383                                        if (e) {
1384                                                sector_t next_boundary = read_original_sector(wc, e) - bio->bi_iter.bi_sector;
1385                                                BUG_ON(!next_boundary);
1386                                                if (next_boundary < bio->bi_iter.bi_size >> SECTOR_SHIFT) {
1387                                                        dm_accept_partial_bio(bio, next_boundary);
1388                                                }
1389                                        }
1390                                        goto unlock_remap_origin;
1391                                }
1392                                writecache_wait_on_freelist(wc);
1393                                continue;
1394                        }
1395                        write_original_sector_seq_count(wc, e, bio->bi_iter.bi_sector, wc->seq_count);
1396                        writecache_insert_entry(wc, e);
1397                        wc->uncommitted_blocks++;
1398bio_copy:
1399                        if (WC_MODE_PMEM(wc)) {
1400                                bio_copy_block(wc, bio, memory_data(wc, e));
1401                        } else {
1402                                unsigned bio_size = wc->block_size;
1403                                sector_t start_cache_sec = cache_sector(wc, e);
1404                                sector_t current_cache_sec = start_cache_sec + (bio_size >> SECTOR_SHIFT);
1405
1406                                while (bio_size < bio->bi_iter.bi_size) {
1407                                        struct wc_entry *f = writecache_pop_from_freelist(wc, current_cache_sec);
1408                                        if (!f)
1409                                                break;
1410                                        write_original_sector_seq_count(wc, f, bio->bi_iter.bi_sector +
1411                                                                        (bio_size >> SECTOR_SHIFT), wc->seq_count);
1412                                        writecache_insert_entry(wc, f);
1413                                        wc->uncommitted_blocks++;
1414                                        bio_size += wc->block_size;
1415                                        current_cache_sec += wc->block_size >> SECTOR_SHIFT;
1416                                }
1417
1418                                bio_set_dev(bio, wc->ssd_dev->bdev);
1419                                bio->bi_iter.bi_sector = start_cache_sec;
1420                                dm_accept_partial_bio(bio, bio_size >> SECTOR_SHIFT);
1421
1422                                if (unlikely(wc->uncommitted_blocks >= wc->autocommit_blocks)) {
1423                                        wc->uncommitted_blocks = 0;
1424                                        queue_work(wc->writeback_wq, &wc->flush_work);
1425                                } else {
1426                                        writecache_schedule_autocommit(wc);
1427                                }
1428                                goto unlock_remap;
1429                        }
1430                } while (bio->bi_iter.bi_size);
1431
1432                if (unlikely(bio->bi_opf & REQ_FUA ||
1433                             wc->uncommitted_blocks >= wc->autocommit_blocks))
1434                        writecache_flush(wc);
1435                else
1436                        writecache_schedule_autocommit(wc);
1437                goto unlock_submit;
1438        }
1439
1440unlock_remap_origin:
1441        bio_set_dev(bio, wc->dev->bdev);
1442        wc_unlock(wc);
1443        return DM_MAPIO_REMAPPED;
1444
1445unlock_remap:
1446        /* make sure that writecache_end_io decrements bio_in_progress: */
1447        bio->bi_private = (void *)1;
1448        atomic_inc(&wc->bio_in_progress[bio_data_dir(bio)]);
1449        wc_unlock(wc);
1450        return DM_MAPIO_REMAPPED;
1451
1452unlock_submit:
1453        wc_unlock(wc);
1454        bio_endio(bio);
1455        return DM_MAPIO_SUBMITTED;
1456
1457unlock_return:
1458        wc_unlock(wc);
1459        return DM_MAPIO_SUBMITTED;
1460
1461unlock_error:
1462        wc_unlock(wc);
1463        bio_io_error(bio);
1464        return DM_MAPIO_SUBMITTED;
1465}
1466
1467static int writecache_end_io(struct dm_target *ti, struct bio *bio, blk_status_t *status)
1468{
1469        struct dm_writecache *wc = ti->private;
1470
1471        if (bio->bi_private != NULL) {
1472                int dir = bio_data_dir(bio);
1473                if (atomic_dec_and_test(&wc->bio_in_progress[dir]))
1474                        if (unlikely(waitqueue_active(&wc->bio_in_progress_wait[dir])))
1475                                wake_up(&wc->bio_in_progress_wait[dir]);
1476        }
1477        return 0;
1478}
1479
1480static int writecache_iterate_devices(struct dm_target *ti,
1481                                      iterate_devices_callout_fn fn, void *data)
1482{
1483        struct dm_writecache *wc = ti->private;
1484
1485        return fn(ti, wc->dev, 0, ti->len, data);
1486}
1487
1488static void writecache_io_hints(struct dm_target *ti, struct queue_limits *limits)
1489{
1490        struct dm_writecache *wc = ti->private;
1491
1492        if (limits->logical_block_size < wc->block_size)
1493                limits->logical_block_size = wc->block_size;
1494
1495        if (limits->physical_block_size < wc->block_size)
1496                limits->physical_block_size = wc->block_size;
1497
1498        if (limits->io_min < wc->block_size)
1499                limits->io_min = wc->block_size;
1500}
1501
1502
1503static void writecache_writeback_endio(struct bio *bio)
1504{
1505        struct writeback_struct *wb = container_of(bio, struct writeback_struct, bio);
1506        struct dm_writecache *wc = wb->wc;
1507        unsigned long flags;
1508
1509        raw_spin_lock_irqsave(&wc->endio_list_lock, flags);
1510        if (unlikely(list_empty(&wc->endio_list)))
1511                wake_up_process(wc->endio_thread);
1512        list_add_tail(&wb->endio_entry, &wc->endio_list);
1513        raw_spin_unlock_irqrestore(&wc->endio_list_lock, flags);
1514}
1515
1516static void writecache_copy_endio(int read_err, unsigned long write_err, void *ptr)
1517{
1518        struct copy_struct *c = ptr;
1519        struct dm_writecache *wc = c->wc;
1520
1521        c->error = likely(!(read_err | write_err)) ? 0 : -EIO;
1522
1523        raw_spin_lock_irq(&wc->endio_list_lock);
1524        if (unlikely(list_empty(&wc->endio_list)))
1525                wake_up_process(wc->endio_thread);
1526        list_add_tail(&c->endio_entry, &wc->endio_list);
1527        raw_spin_unlock_irq(&wc->endio_list_lock);
1528}
1529
1530static void __writecache_endio_pmem(struct dm_writecache *wc, struct list_head *list)
1531{
1532        unsigned i;
1533        struct writeback_struct *wb;
1534        struct wc_entry *e;
1535        unsigned long n_walked = 0;
1536
1537        do {
1538                wb = list_entry(list->next, struct writeback_struct, endio_entry);
1539                list_del(&wb->endio_entry);
1540
1541                if (unlikely(wb->bio.bi_status != BLK_STS_OK))
1542                        writecache_error(wc, blk_status_to_errno(wb->bio.bi_status),
1543                                        "write error %d", wb->bio.bi_status);
1544                i = 0;
1545                do {
1546                        e = wb->wc_list[i];
1547                        BUG_ON(!e->write_in_progress);
1548                        e->write_in_progress = false;
1549                        INIT_LIST_HEAD(&e->lru);
1550                        if (!writecache_has_error(wc))
1551                                writecache_free_entry(wc, e);
1552                        BUG_ON(!wc->writeback_size);
1553                        wc->writeback_size--;
1554                        n_walked++;
1555                        if (unlikely(n_walked >= ENDIO_LATENCY)) {
1556                                writecache_commit_flushed(wc, false);
1557                                wc_unlock(wc);
1558                                wc_lock(wc);
1559                                n_walked = 0;
1560                        }
1561                } while (++i < wb->wc_list_n);
1562
1563                if (wb->wc_list != wb->wc_list_inline)
1564                        kfree(wb->wc_list);
1565                bio_put(&wb->bio);
1566        } while (!list_empty(list));
1567}
1568
1569static void __writecache_endio_ssd(struct dm_writecache *wc, struct list_head *list)
1570{
1571        struct copy_struct *c;
1572        struct wc_entry *e;
1573
1574        do {
1575                c = list_entry(list->next, struct copy_struct, endio_entry);
1576                list_del(&c->endio_entry);
1577
1578                if (unlikely(c->error))
1579                        writecache_error(wc, c->error, "copy error");
1580
1581                e = c->e;
1582                do {
1583                        BUG_ON(!e->write_in_progress);
1584                        e->write_in_progress = false;
1585                        INIT_LIST_HEAD(&e->lru);
1586                        if (!writecache_has_error(wc))
1587                                writecache_free_entry(wc, e);
1588
1589                        BUG_ON(!wc->writeback_size);
1590                        wc->writeback_size--;
1591                        e++;
1592                } while (--c->n_entries);
1593                mempool_free(c, &wc->copy_pool);
1594        } while (!list_empty(list));
1595}
1596
1597static int writecache_endio_thread(void *data)
1598{
1599        struct dm_writecache *wc = data;
1600
1601        while (1) {
1602                struct list_head list;
1603
1604                raw_spin_lock_irq(&wc->endio_list_lock);
1605                if (!list_empty(&wc->endio_list))
1606                        goto pop_from_list;
1607                set_current_state(TASK_INTERRUPTIBLE);
1608                raw_spin_unlock_irq(&wc->endio_list_lock);
1609
1610                if (unlikely(kthread_should_stop())) {
1611                        set_current_state(TASK_RUNNING);
1612                        break;
1613                }
1614
1615                schedule();
1616
1617                continue;
1618
1619pop_from_list:
1620                list = wc->endio_list;
1621                list.next->prev = list.prev->next = &list;
1622                INIT_LIST_HEAD(&wc->endio_list);
1623                raw_spin_unlock_irq(&wc->endio_list_lock);
1624
1625                if (!WC_MODE_FUA(wc))
1626                        writecache_disk_flush(wc, wc->dev);
1627
1628                wc_lock(wc);
1629
1630                if (WC_MODE_PMEM(wc)) {
1631                        __writecache_endio_pmem(wc, &list);
1632                } else {
1633                        __writecache_endio_ssd(wc, &list);
1634                        writecache_wait_for_ios(wc, READ);
1635                }
1636
1637                writecache_commit_flushed(wc, false);
1638
1639                wc_unlock(wc);
1640        }
1641
1642        return 0;
1643}
1644
1645static bool wc_add_block(struct writeback_struct *wb, struct wc_entry *e, gfp_t gfp)
1646{
1647        struct dm_writecache *wc = wb->wc;
1648        unsigned block_size = wc->block_size;
1649        void *address = memory_data(wc, e);
1650
1651        persistent_memory_flush_cache(address, block_size);
1652
1653        if (unlikely(bio_end_sector(&wb->bio) >= wc->data_device_sectors))
1654                return true;
1655
1656        return bio_add_page(&wb->bio, persistent_memory_page(address),
1657                            block_size, persistent_memory_page_offset(address)) != 0;
1658}
1659
1660struct writeback_list {
1661        struct list_head list;
1662        size_t size;
1663};
1664
1665static void __writeback_throttle(struct dm_writecache *wc, struct writeback_list *wbl)
1666{
1667        if (unlikely(wc->max_writeback_jobs)) {
1668                if (READ_ONCE(wc->writeback_size) - wbl->size >= wc->max_writeback_jobs) {
1669                        wc_lock(wc);
1670                        while (wc->writeback_size - wbl->size >= wc->max_writeback_jobs)
1671                                writecache_wait_on_freelist(wc);
1672                        wc_unlock(wc);
1673                }
1674        }
1675        cond_resched();
1676}
1677
1678static void __writecache_writeback_pmem(struct dm_writecache *wc, struct writeback_list *wbl)
1679{
1680        struct wc_entry *e, *f;
1681        struct bio *bio;
1682        struct writeback_struct *wb;
1683        unsigned max_pages;
1684
1685        while (wbl->size) {
1686                wbl->size--;
1687                e = container_of(wbl->list.prev, struct wc_entry, lru);
1688                list_del(&e->lru);
1689
1690                max_pages = e->wc_list_contiguous;
1691
1692                bio = bio_alloc_bioset(GFP_NOIO, max_pages, &wc->bio_set);
1693                wb = container_of(bio, struct writeback_struct, bio);
1694                wb->wc = wc;
1695                bio->bi_end_io = writecache_writeback_endio;
1696                bio_set_dev(bio, wc->dev->bdev);
1697                bio->bi_iter.bi_sector = read_original_sector(wc, e);
1698                if (max_pages <= WB_LIST_INLINE ||
1699                    unlikely(!(wb->wc_list = kmalloc_array(max_pages, sizeof(struct wc_entry *),
1700                                                           GFP_NOIO | __GFP_NORETRY |
1701                                                           __GFP_NOMEMALLOC | __GFP_NOWARN)))) {
1702                        wb->wc_list = wb->wc_list_inline;
1703                        max_pages = WB_LIST_INLINE;
1704                }
1705
1706                BUG_ON(!wc_add_block(wb, e, GFP_NOIO));
1707
1708                wb->wc_list[0] = e;
1709                wb->wc_list_n = 1;
1710
1711                while (wbl->size && wb->wc_list_n < max_pages) {
1712                        f = container_of(wbl->list.prev, struct wc_entry, lru);
1713                        if (read_original_sector(wc, f) !=
1714                            read_original_sector(wc, e) + (wc->block_size >> SECTOR_SHIFT))
1715                                break;
1716                        if (!wc_add_block(wb, f, GFP_NOWAIT | __GFP_NOWARN))
1717                                break;
1718                        wbl->size--;
1719                        list_del(&f->lru);
1720                        wb->wc_list[wb->wc_list_n++] = f;
1721                        e = f;
1722                }
1723                bio_set_op_attrs(bio, REQ_OP_WRITE, WC_MODE_FUA(wc) * REQ_FUA);
1724                if (writecache_has_error(wc)) {
1725                        bio->bi_status = BLK_STS_IOERR;
1726                        bio_endio(bio);
1727                } else if (unlikely(!bio_sectors(bio))) {
1728                        bio->bi_status = BLK_STS_OK;
1729                        bio_endio(bio);
1730                } else {
1731                        submit_bio(bio);
1732                }
1733
1734                __writeback_throttle(wc, wbl);
1735        }
1736}
1737
1738static void __writecache_writeback_ssd(struct dm_writecache *wc, struct writeback_list *wbl)
1739{
1740        struct wc_entry *e, *f;
1741        struct dm_io_region from, to;
1742        struct copy_struct *c;
1743
1744        while (wbl->size) {
1745                unsigned n_sectors;
1746
1747                wbl->size--;
1748                e = container_of(wbl->list.prev, struct wc_entry, lru);
1749                list_del(&e->lru);
1750
1751                n_sectors = e->wc_list_contiguous << (wc->block_size_bits - SECTOR_SHIFT);
1752
1753                from.bdev = wc->ssd_dev->bdev;
1754                from.sector = cache_sector(wc, e);
1755                from.count = n_sectors;
1756                to.bdev = wc->dev->bdev;
1757                to.sector = read_original_sector(wc, e);
1758                to.count = n_sectors;
1759
1760                c = mempool_alloc(&wc->copy_pool, GFP_NOIO);
1761                c->wc = wc;
1762                c->e = e;
1763                c->n_entries = e->wc_list_contiguous;
1764
1765                while ((n_sectors -= wc->block_size >> SECTOR_SHIFT)) {
1766                        wbl->size--;
1767                        f = container_of(wbl->list.prev, struct wc_entry, lru);
1768                        BUG_ON(f != e + 1);
1769                        list_del(&f->lru);
1770                        e = f;
1771                }
1772
1773                if (unlikely(to.sector + to.count > wc->data_device_sectors)) {
1774                        if (to.sector >= wc->data_device_sectors) {
1775                                writecache_copy_endio(0, 0, c);
1776                                continue;
1777                        }
1778                        from.count = to.count = wc->data_device_sectors - to.sector;
1779                }
1780
1781                dm_kcopyd_copy(wc->dm_kcopyd, &from, 1, &to, 0, writecache_copy_endio, c);
1782
1783                __writeback_throttle(wc, wbl);
1784        }
1785}
1786
1787static void writecache_writeback(struct work_struct *work)
1788{
1789        struct dm_writecache *wc = container_of(work, struct dm_writecache, writeback_work);
1790        struct blk_plug plug;
1791        struct wc_entry *f, *g, *e = NULL;
1792        struct rb_node *node, *next_node;
1793        struct list_head skipped;
1794        struct writeback_list wbl;
1795        unsigned long n_walked;
1796
1797        wc_lock(wc);
1798restart:
1799        if (writecache_has_error(wc)) {
1800                wc_unlock(wc);
1801                return;
1802        }
1803
1804        if (unlikely(wc->writeback_all)) {
1805                if (writecache_wait_for_writeback(wc))
1806                        goto restart;
1807        }
1808
1809        if (wc->overwrote_committed) {
1810                writecache_wait_for_ios(wc, WRITE);
1811        }
1812
1813        n_walked = 0;
1814        INIT_LIST_HEAD(&skipped);
1815        INIT_LIST_HEAD(&wbl.list);
1816        wbl.size = 0;
1817        while (!list_empty(&wc->lru) &&
1818               (wc->writeback_all ||
1819                wc->freelist_size + wc->writeback_size <= wc->freelist_low_watermark ||
1820                (jiffies - container_of(wc->lru.prev, struct wc_entry, lru)->age >=
1821                 wc->max_age - wc->max_age / MAX_AGE_DIV))) {
1822
1823                n_walked++;
1824                if (unlikely(n_walked > WRITEBACK_LATENCY) &&
1825                    likely(!wc->writeback_all) && likely(!dm_suspended(wc->ti))) {
1826                        queue_work(wc->writeback_wq, &wc->writeback_work);
1827                        break;
1828                }
1829
1830                if (unlikely(wc->writeback_all)) {
1831                        if (unlikely(!e)) {
1832                                writecache_flush(wc);
1833                                e = container_of(rb_first(&wc->tree), struct wc_entry, rb_node);
1834                        } else
1835                                e = g;
1836                } else
1837                        e = container_of(wc->lru.prev, struct wc_entry, lru);
1838                BUG_ON(e->write_in_progress);
1839                if (unlikely(!writecache_entry_is_committed(wc, e))) {
1840                        writecache_flush(wc);
1841                }
1842                node = rb_prev(&e->rb_node);
1843                if (node) {
1844                        f = container_of(node, struct wc_entry, rb_node);
1845                        if (unlikely(read_original_sector(wc, f) ==
1846                                     read_original_sector(wc, e))) {
1847                                BUG_ON(!f->write_in_progress);
1848                                list_del(&e->lru);
1849                                list_add(&e->lru, &skipped);
1850                                cond_resched();
1851                                continue;
1852                        }
1853                }
1854                wc->writeback_size++;
1855                list_del(&e->lru);
1856                list_add(&e->lru, &wbl.list);
1857                wbl.size++;
1858                e->write_in_progress = true;
1859                e->wc_list_contiguous = 1;
1860
1861                f = e;
1862
1863                while (1) {
1864                        next_node = rb_next(&f->rb_node);
1865                        if (unlikely(!next_node))
1866                                break;
1867                        g = container_of(next_node, struct wc_entry, rb_node);
1868                        if (unlikely(read_original_sector(wc, g) ==
1869                            read_original_sector(wc, f))) {
1870                                f = g;
1871                                continue;
1872                        }
1873                        if (read_original_sector(wc, g) !=
1874                            read_original_sector(wc, f) + (wc->block_size >> SECTOR_SHIFT))
1875                                break;
1876                        if (unlikely(g->write_in_progress))
1877                                break;
1878                        if (unlikely(!writecache_entry_is_committed(wc, g)))
1879                                break;
1880
1881                        if (!WC_MODE_PMEM(wc)) {
1882                                if (g != f + 1)
1883                                        break;
1884                        }
1885
1886                        n_walked++;
1887                        //if (unlikely(n_walked > WRITEBACK_LATENCY) && likely(!wc->writeback_all))
1888                        //      break;
1889
1890                        wc->writeback_size++;
1891                        list_del(&g->lru);
1892                        list_add(&g->lru, &wbl.list);
1893                        wbl.size++;
1894                        g->write_in_progress = true;
1895                        g->wc_list_contiguous = BIO_MAX_VECS;
1896                        f = g;
1897                        e->wc_list_contiguous++;
1898                        if (unlikely(e->wc_list_contiguous == BIO_MAX_VECS)) {
1899                                if (unlikely(wc->writeback_all)) {
1900                                        next_node = rb_next(&f->rb_node);
1901                                        if (likely(next_node))
1902                                                g = container_of(next_node, struct wc_entry, rb_node);
1903                                }
1904                                break;
1905                        }
1906                }
1907                cond_resched();
1908        }
1909
1910        if (!list_empty(&skipped)) {
1911                list_splice_tail(&skipped, &wc->lru);
1912                /*
1913                 * If we didn't do any progress, we must wait until some
1914                 * writeback finishes to avoid burning CPU in a loop
1915                 */
1916                if (unlikely(!wbl.size))
1917                        writecache_wait_for_writeback(wc);
1918        }
1919
1920        wc_unlock(wc);
1921
1922        blk_start_plug(&plug);
1923
1924        if (WC_MODE_PMEM(wc))
1925                __writecache_writeback_pmem(wc, &wbl);
1926        else
1927                __writecache_writeback_ssd(wc, &wbl);
1928
1929        blk_finish_plug(&plug);
1930
1931        if (unlikely(wc->writeback_all)) {
1932                wc_lock(wc);
1933                while (writecache_wait_for_writeback(wc));
1934                wc_unlock(wc);
1935        }
1936}
1937
1938static int calculate_memory_size(uint64_t device_size, unsigned block_size,
1939                                 size_t *n_blocks_p, size_t *n_metadata_blocks_p)
1940{
1941        uint64_t n_blocks, offset;
1942        struct wc_entry e;
1943
1944        n_blocks = device_size;
1945        do_div(n_blocks, block_size + sizeof(struct wc_memory_entry));
1946
1947        while (1) {
1948                if (!n_blocks)
1949                        return -ENOSPC;
1950                /* Verify the following entries[n_blocks] won't overflow */
1951                if (n_blocks >= ((size_t)-sizeof(struct wc_memory_superblock) /
1952                                 sizeof(struct wc_memory_entry)))
1953                        return -EFBIG;
1954                offset = offsetof(struct wc_memory_superblock, entries[n_blocks]);
1955                offset = (offset + block_size - 1) & ~(uint64_t)(block_size - 1);
1956                if (offset + n_blocks * block_size <= device_size)
1957                        break;
1958                n_blocks--;
1959        }
1960
1961        /* check if the bit field overflows */
1962        e.index = n_blocks;
1963        if (e.index != n_blocks)
1964                return -EFBIG;
1965
1966        if (n_blocks_p)
1967                *n_blocks_p = n_blocks;
1968        if (n_metadata_blocks_p)
1969                *n_metadata_blocks_p = offset >> __ffs(block_size);
1970        return 0;
1971}
1972
1973static int init_memory(struct dm_writecache *wc)
1974{
1975        size_t b;
1976        int r;
1977
1978        r = calculate_memory_size(wc->memory_map_size, wc->block_size, &wc->n_blocks, NULL);
1979        if (r)
1980                return r;
1981
1982        r = writecache_alloc_entries(wc);
1983        if (r)
1984                return r;
1985
1986        for (b = 0; b < ARRAY_SIZE(sb(wc)->padding); b++)
1987                pmem_assign(sb(wc)->padding[b], cpu_to_le64(0));
1988        pmem_assign(sb(wc)->version, cpu_to_le32(MEMORY_SUPERBLOCK_VERSION));
1989        pmem_assign(sb(wc)->block_size, cpu_to_le32(wc->block_size));
1990        pmem_assign(sb(wc)->n_blocks, cpu_to_le64(wc->n_blocks));
1991        pmem_assign(sb(wc)->seq_count, cpu_to_le64(0));
1992
1993        for (b = 0; b < wc->n_blocks; b++) {
1994                write_original_sector_seq_count(wc, &wc->entries[b], -1, -1);
1995                cond_resched();
1996        }
1997
1998        writecache_flush_all_metadata(wc);
1999        writecache_commit_flushed(wc, false);
2000        pmem_assign(sb(wc)->magic, cpu_to_le32(MEMORY_SUPERBLOCK_MAGIC));
2001        writecache_flush_region(wc, &sb(wc)->magic, sizeof sb(wc)->magic);
2002        writecache_commit_flushed(wc, false);
2003
2004        return 0;
2005}
2006
2007static void writecache_dtr(struct dm_target *ti)
2008{
2009        struct dm_writecache *wc = ti->private;
2010
2011        if (!wc)
2012                return;
2013
2014        if (wc->endio_thread)
2015                kthread_stop(wc->endio_thread);
2016
2017        if (wc->flush_thread)
2018                kthread_stop(wc->flush_thread);
2019
2020        bioset_exit(&wc->bio_set);
2021
2022        mempool_exit(&wc->copy_pool);
2023
2024        if (wc->writeback_wq)
2025                destroy_workqueue(wc->writeback_wq);
2026
2027        if (wc->dev)
2028                dm_put_device(ti, wc->dev);
2029
2030        if (wc->ssd_dev)
2031                dm_put_device(ti, wc->ssd_dev);
2032
2033        vfree(wc->entries);
2034
2035        if (wc->memory_map) {
2036                if (WC_MODE_PMEM(wc))
2037                        persistent_memory_release(wc);
2038                else
2039                        vfree(wc->memory_map);
2040        }
2041
2042        if (wc->dm_kcopyd)
2043                dm_kcopyd_client_destroy(wc->dm_kcopyd);
2044
2045        if (wc->dm_io)
2046                dm_io_client_destroy(wc->dm_io);
2047
2048        vfree(wc->dirty_bitmap);
2049
2050        kfree(wc);
2051}
2052
2053static int writecache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2054{
2055        struct dm_writecache *wc;
2056        struct dm_arg_set as;
2057        const char *string;
2058        unsigned opt_params;
2059        size_t offset, data_size;
2060        int i, r;
2061        char dummy;
2062        int high_wm_percent = HIGH_WATERMARK;
2063        int low_wm_percent = LOW_WATERMARK;
2064        uint64_t x;
2065        struct wc_memory_superblock s;
2066
2067        static struct dm_arg _args[] = {
2068                {0, 16, "Invalid number of feature args"},
2069        };
2070
2071        as.argc = argc;
2072        as.argv = argv;
2073
2074        wc = kzalloc(sizeof(struct dm_writecache), GFP_KERNEL);
2075        if (!wc) {
2076                ti->error = "Cannot allocate writecache structure";
2077                r = -ENOMEM;
2078                goto bad;
2079        }
2080        ti->private = wc;
2081        wc->ti = ti;
2082
2083        mutex_init(&wc->lock);
2084        wc->max_age = MAX_AGE_UNSPECIFIED;
2085        writecache_poison_lists(wc);
2086        init_waitqueue_head(&wc->freelist_wait);
2087        timer_setup(&wc->autocommit_timer, writecache_autocommit_timer, 0);
2088        timer_setup(&wc->max_age_timer, writecache_max_age_timer, 0);
2089
2090        for (i = 0; i < 2; i++) {
2091                atomic_set(&wc->bio_in_progress[i], 0);
2092                init_waitqueue_head(&wc->bio_in_progress_wait[i]);
2093        }
2094
2095        wc->dm_io = dm_io_client_create();
2096        if (IS_ERR(wc->dm_io)) {
2097                r = PTR_ERR(wc->dm_io);
2098                ti->error = "Unable to allocate dm-io client";
2099                wc->dm_io = NULL;
2100                goto bad;
2101        }
2102
2103        wc->writeback_wq = alloc_workqueue("writecache-writeback", WQ_MEM_RECLAIM, 1);
2104        if (!wc->writeback_wq) {
2105                r = -ENOMEM;
2106                ti->error = "Could not allocate writeback workqueue";
2107                goto bad;
2108        }
2109        INIT_WORK(&wc->writeback_work, writecache_writeback);
2110        INIT_WORK(&wc->flush_work, writecache_flush_work);
2111
2112        raw_spin_lock_init(&wc->endio_list_lock);
2113        INIT_LIST_HEAD(&wc->endio_list);
2114        wc->endio_thread = kthread_create(writecache_endio_thread, wc, "writecache_endio");
2115        if (IS_ERR(wc->endio_thread)) {
2116                r = PTR_ERR(wc->endio_thread);
2117                wc->endio_thread = NULL;
2118                ti->error = "Couldn't spawn endio thread";
2119                goto bad;
2120        }
2121        wake_up_process(wc->endio_thread);
2122
2123        /*
2124         * Parse the mode (pmem or ssd)
2125         */
2126        string = dm_shift_arg(&as);
2127        if (!string)
2128                goto bad_arguments;
2129
2130        if (!strcasecmp(string, "s")) {
2131                wc->pmem_mode = false;
2132        } else if (!strcasecmp(string, "p")) {
2133#ifdef DM_WRITECACHE_HAS_PMEM
2134                wc->pmem_mode = true;
2135                wc->writeback_fua = true;
2136#else
2137                /*
2138                 * If the architecture doesn't support persistent memory or
2139                 * the kernel doesn't support any DAX drivers, this driver can
2140                 * only be used in SSD-only mode.
2141                 */
2142                r = -EOPNOTSUPP;
2143                ti->error = "Persistent memory or DAX not supported on this system";
2144                goto bad;
2145#endif
2146        } else {
2147                goto bad_arguments;
2148        }
2149
2150        if (WC_MODE_PMEM(wc)) {
2151                r = bioset_init(&wc->bio_set, BIO_POOL_SIZE,
2152                                offsetof(struct writeback_struct, bio),
2153                                BIOSET_NEED_BVECS);
2154                if (r) {
2155                        ti->error = "Could not allocate bio set";
2156                        goto bad;
2157                }
2158        } else {
2159                r = mempool_init_kmalloc_pool(&wc->copy_pool, 1, sizeof(struct copy_struct));
2160                if (r) {
2161                        ti->error = "Could not allocate mempool";
2162                        goto bad;
2163                }
2164        }
2165
2166        /*
2167         * Parse the origin data device
2168         */
2169        string = dm_shift_arg(&as);
2170        if (!string)
2171                goto bad_arguments;
2172        r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->dev);
2173        if (r) {
2174                ti->error = "Origin data device lookup failed";
2175                goto bad;
2176        }
2177
2178        /*
2179         * Parse cache data device (be it pmem or ssd)
2180         */
2181        string = dm_shift_arg(&as);
2182        if (!string)
2183                goto bad_arguments;
2184
2185        r = dm_get_device(ti, string, dm_table_get_mode(ti->table), &wc->ssd_dev);
2186        if (r) {
2187                ti->error = "Cache data device lookup failed";
2188                goto bad;
2189        }
2190        wc->memory_map_size = i_size_read(wc->ssd_dev->bdev->bd_inode);
2191
2192        /*
2193         * Parse the cache block size
2194         */
2195        string = dm_shift_arg(&as);
2196        if (!string)
2197                goto bad_arguments;
2198        if (sscanf(string, "%u%c", &wc->block_size, &dummy) != 1 ||
2199            wc->block_size < 512 || wc->block_size > PAGE_SIZE ||
2200            (wc->block_size & (wc->block_size - 1))) {
2201                r = -EINVAL;
2202                ti->error = "Invalid block size";
2203                goto bad;
2204        }
2205        if (wc->block_size < bdev_logical_block_size(wc->dev->bdev) ||
2206            wc->block_size < bdev_logical_block_size(wc->ssd_dev->bdev)) {
2207                r = -EINVAL;
2208                ti->error = "Block size is smaller than device logical block size";
2209                goto bad;
2210        }
2211        wc->block_size_bits = __ffs(wc->block_size);
2212
2213        wc->max_writeback_jobs = MAX_WRITEBACK_JOBS;
2214        wc->autocommit_blocks = !WC_MODE_PMEM(wc) ? AUTOCOMMIT_BLOCKS_SSD : AUTOCOMMIT_BLOCKS_PMEM;
2215        wc->autocommit_jiffies = msecs_to_jiffies(AUTOCOMMIT_MSEC);
2216
2217        /*
2218         * Parse optional arguments
2219         */
2220        r = dm_read_arg_group(_args, &as, &opt_params, &ti->error);
2221        if (r)
2222                goto bad;
2223
2224        while (opt_params) {
2225                string = dm_shift_arg(&as), opt_params--;
2226                if (!strcasecmp(string, "start_sector") && opt_params >= 1) {
2227                        unsigned long long start_sector;
2228                        string = dm_shift_arg(&as), opt_params--;
2229                        if (sscanf(string, "%llu%c", &start_sector, &dummy) != 1)
2230                                goto invalid_optional;
2231                        wc->start_sector = start_sector;
2232                        wc->start_sector_set = true;
2233                        if (wc->start_sector != start_sector ||
2234                            wc->start_sector >= wc->memory_map_size >> SECTOR_SHIFT)
2235                                goto invalid_optional;
2236                } else if (!strcasecmp(string, "high_watermark") && opt_params >= 1) {
2237                        string = dm_shift_arg(&as), opt_params--;
2238                        if (sscanf(string, "%d%c", &high_wm_percent, &dummy) != 1)
2239                                goto invalid_optional;
2240                        if (high_wm_percent < 0 || high_wm_percent > 100)
2241                                goto invalid_optional;
2242                        wc->high_wm_percent_value = high_wm_percent;
2243                        wc->high_wm_percent_set = true;
2244                } else if (!strcasecmp(string, "low_watermark") && opt_params >= 1) {
2245                        string = dm_shift_arg(&as), opt_params--;
2246                        if (sscanf(string, "%d%c", &low_wm_percent, &dummy) != 1)
2247                                goto invalid_optional;
2248                        if (low_wm_percent < 0 || low_wm_percent > 100)
2249                                goto invalid_optional;
2250                        wc->low_wm_percent_value = low_wm_percent;
2251                        wc->low_wm_percent_set = true;
2252                } else if (!strcasecmp(string, "writeback_jobs") && opt_params >= 1) {
2253                        string = dm_shift_arg(&as), opt_params--;
2254                        if (sscanf(string, "%u%c", &wc->max_writeback_jobs, &dummy) != 1)
2255                                goto invalid_optional;
2256                        wc->max_writeback_jobs_set = true;
2257                } else if (!strcasecmp(string, "autocommit_blocks") && opt_params >= 1) {
2258                        string = dm_shift_arg(&as), opt_params--;
2259                        if (sscanf(string, "%u%c", &wc->autocommit_blocks, &dummy) != 1)
2260                                goto invalid_optional;
2261                        wc->autocommit_blocks_set = true;
2262                } else if (!strcasecmp(string, "autocommit_time") && opt_params >= 1) {
2263                        unsigned autocommit_msecs;
2264                        string = dm_shift_arg(&as), opt_params--;
2265                        if (sscanf(string, "%u%c", &autocommit_msecs, &dummy) != 1)
2266                                goto invalid_optional;
2267                        if (autocommit_msecs > 3600000)
2268                                goto invalid_optional;
2269                        wc->autocommit_jiffies = msecs_to_jiffies(autocommit_msecs);
2270                        wc->autocommit_time_value = autocommit_msecs;
2271                        wc->autocommit_time_set = true;
2272                } else if (!strcasecmp(string, "max_age") && opt_params >= 1) {
2273                        unsigned max_age_msecs;
2274                        string = dm_shift_arg(&as), opt_params--;
2275                        if (sscanf(string, "%u%c", &max_age_msecs, &dummy) != 1)
2276                                goto invalid_optional;
2277                        if (max_age_msecs > 86400000)
2278                                goto invalid_optional;
2279                        wc->max_age = msecs_to_jiffies(max_age_msecs);
2280                        wc->max_age_set = true;
2281                        wc->max_age_value = max_age_msecs;
2282                } else if (!strcasecmp(string, "cleaner")) {
2283                        wc->cleaner_set = true;
2284                        wc->cleaner = true;
2285                } else if (!strcasecmp(string, "fua")) {
2286                        if (WC_MODE_PMEM(wc)) {
2287                                wc->writeback_fua = true;
2288                                wc->writeback_fua_set = true;
2289                        } else goto invalid_optional;
2290                } else if (!strcasecmp(string, "nofua")) {
2291                        if (WC_MODE_PMEM(wc)) {
2292                                wc->writeback_fua = false;
2293                                wc->writeback_fua_set = true;
2294                        } else goto invalid_optional;
2295                } else {
2296invalid_optional:
2297                        r = -EINVAL;
2298                        ti->error = "Invalid optional argument";
2299                        goto bad;
2300                }
2301        }
2302
2303        if (high_wm_percent < low_wm_percent) {
2304                r = -EINVAL;
2305                ti->error = "High watermark must be greater than or equal to low watermark";
2306                goto bad;
2307        }
2308
2309        if (WC_MODE_PMEM(wc)) {
2310                if (!dax_synchronous(wc->ssd_dev->dax_dev)) {
2311                        r = -EOPNOTSUPP;
2312                        ti->error = "Asynchronous persistent memory not supported as pmem cache";
2313                        goto bad;
2314                }
2315
2316                r = persistent_memory_claim(wc);
2317                if (r) {
2318                        ti->error = "Unable to map persistent memory for cache";
2319                        goto bad;
2320                }
2321        } else {
2322                size_t n_blocks, n_metadata_blocks;
2323                uint64_t n_bitmap_bits;
2324
2325                wc->memory_map_size -= (uint64_t)wc->start_sector << SECTOR_SHIFT;
2326
2327                bio_list_init(&wc->flush_list);
2328                wc->flush_thread = kthread_create(writecache_flush_thread, wc, "dm_writecache_flush");
2329                if (IS_ERR(wc->flush_thread)) {
2330                        r = PTR_ERR(wc->flush_thread);
2331                        wc->flush_thread = NULL;
2332                        ti->error = "Couldn't spawn flush thread";
2333                        goto bad;
2334                }
2335                wake_up_process(wc->flush_thread);
2336
2337                r = calculate_memory_size(wc->memory_map_size, wc->block_size,
2338                                          &n_blocks, &n_metadata_blocks);
2339                if (r) {
2340                        ti->error = "Invalid device size";
2341                        goto bad;
2342                }
2343
2344                n_bitmap_bits = (((uint64_t)n_metadata_blocks << wc->block_size_bits) +
2345                                 BITMAP_GRANULARITY - 1) / BITMAP_GRANULARITY;
2346                /* this is limitation of test_bit functions */
2347                if (n_bitmap_bits > 1U << 31) {
2348                        r = -EFBIG;
2349                        ti->error = "Invalid device size";
2350                        goto bad;
2351                }
2352
2353                wc->memory_map = vmalloc(n_metadata_blocks << wc->block_size_bits);
2354                if (!wc->memory_map) {
2355                        r = -ENOMEM;
2356                        ti->error = "Unable to allocate memory for metadata";
2357                        goto bad;
2358                }
2359
2360                wc->dm_kcopyd = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2361                if (IS_ERR(wc->dm_kcopyd)) {
2362                        r = PTR_ERR(wc->dm_kcopyd);
2363                        ti->error = "Unable to allocate dm-kcopyd client";
2364                        wc->dm_kcopyd = NULL;
2365                        goto bad;
2366                }
2367
2368                wc->metadata_sectors = n_metadata_blocks << (wc->block_size_bits - SECTOR_SHIFT);
2369                wc->dirty_bitmap_size = (n_bitmap_bits + BITS_PER_LONG - 1) /
2370                        BITS_PER_LONG * sizeof(unsigned long);
2371                wc->dirty_bitmap = vzalloc(wc->dirty_bitmap_size);
2372                if (!wc->dirty_bitmap) {
2373                        r = -ENOMEM;
2374                        ti->error = "Unable to allocate dirty bitmap";
2375                        goto bad;
2376                }
2377
2378                r = writecache_read_metadata(wc, wc->block_size >> SECTOR_SHIFT);
2379                if (r) {
2380                        ti->error = "Unable to read first block of metadata";
2381                        goto bad;
2382                }
2383        }
2384
2385        r = copy_mc_to_kernel(&s, sb(wc), sizeof(struct wc_memory_superblock));
2386        if (r) {
2387                ti->error = "Hardware memory error when reading superblock";
2388                goto bad;
2389        }
2390        if (!le32_to_cpu(s.magic) && !le32_to_cpu(s.version)) {
2391                r = init_memory(wc);
2392                if (r) {
2393                        ti->error = "Unable to initialize device";
2394                        goto bad;
2395                }
2396                r = copy_mc_to_kernel(&s, sb(wc),
2397                                      sizeof(struct wc_memory_superblock));
2398                if (r) {
2399                        ti->error = "Hardware memory error when reading superblock";
2400                        goto bad;
2401                }
2402        }
2403
2404        if (le32_to_cpu(s.magic) != MEMORY_SUPERBLOCK_MAGIC) {
2405                ti->error = "Invalid magic in the superblock";
2406                r = -EINVAL;
2407                goto bad;
2408        }
2409
2410        if (le32_to_cpu(s.version) != MEMORY_SUPERBLOCK_VERSION) {
2411                ti->error = "Invalid version in the superblock";
2412                r = -EINVAL;
2413                goto bad;
2414        }
2415
2416        if (le32_to_cpu(s.block_size) != wc->block_size) {
2417                ti->error = "Block size does not match superblock";
2418                r = -EINVAL;
2419                goto bad;
2420        }
2421
2422        wc->n_blocks = le64_to_cpu(s.n_blocks);
2423
2424        offset = wc->n_blocks * sizeof(struct wc_memory_entry);
2425        if (offset / sizeof(struct wc_memory_entry) != le64_to_cpu(sb(wc)->n_blocks)) {
2426overflow:
2427                ti->error = "Overflow in size calculation";
2428                r = -EINVAL;
2429                goto bad;
2430        }
2431        offset += sizeof(struct wc_memory_superblock);
2432        if (offset < sizeof(struct wc_memory_superblock))
2433                goto overflow;
2434        offset = (offset + wc->block_size - 1) & ~(size_t)(wc->block_size - 1);
2435        data_size = wc->n_blocks * (size_t)wc->block_size;
2436        if (!offset || (data_size / wc->block_size != wc->n_blocks) ||
2437            (offset + data_size < offset))
2438                goto overflow;
2439        if (offset + data_size > wc->memory_map_size) {
2440                ti->error = "Memory area is too small";
2441                r = -EINVAL;
2442                goto bad;
2443        }
2444
2445        wc->metadata_sectors = offset >> SECTOR_SHIFT;
2446        wc->block_start = (char *)sb(wc) + offset;
2447
2448        x = (uint64_t)wc->n_blocks * (100 - high_wm_percent);
2449        x += 50;
2450        do_div(x, 100);
2451        wc->freelist_high_watermark = x;
2452        x = (uint64_t)wc->n_blocks * (100 - low_wm_percent);
2453        x += 50;
2454        do_div(x, 100);
2455        wc->freelist_low_watermark = x;
2456
2457        if (wc->cleaner)
2458                activate_cleaner(wc);
2459
2460        r = writecache_alloc_entries(wc);
2461        if (r) {
2462                ti->error = "Cannot allocate memory";
2463                goto bad;
2464        }
2465
2466        ti->num_flush_bios = 1;
2467        ti->flush_supported = true;
2468        ti->num_discard_bios = 1;
2469
2470        if (WC_MODE_PMEM(wc))
2471                persistent_memory_flush_cache(wc->memory_map, wc->memory_map_size);
2472
2473        return 0;
2474
2475bad_arguments:
2476        r = -EINVAL;
2477        ti->error = "Bad arguments";
2478bad:
2479        writecache_dtr(ti);
2480        return r;
2481}
2482
2483static void writecache_status(struct dm_target *ti, status_type_t type,
2484                              unsigned status_flags, char *result, unsigned maxlen)
2485{
2486        struct dm_writecache *wc = ti->private;
2487        unsigned extra_args;
2488        unsigned sz = 0;
2489
2490        switch (type) {
2491        case STATUSTYPE_INFO:
2492                DMEMIT("%ld %llu %llu %llu", writecache_has_error(wc),
2493                       (unsigned long long)wc->n_blocks, (unsigned long long)wc->freelist_size,
2494                       (unsigned long long)wc->writeback_size);
2495                break;
2496        case STATUSTYPE_TABLE:
2497                DMEMIT("%c %s %s %u ", WC_MODE_PMEM(wc) ? 'p' : 's',
2498                                wc->dev->name, wc->ssd_dev->name, wc->block_size);
2499                extra_args = 0;
2500                if (wc->start_sector_set)
2501                        extra_args += 2;
2502                if (wc->high_wm_percent_set)
2503                        extra_args += 2;
2504                if (wc->low_wm_percent_set)
2505                        extra_args += 2;
2506                if (wc->max_writeback_jobs_set)
2507                        extra_args += 2;
2508                if (wc->autocommit_blocks_set)
2509                        extra_args += 2;
2510                if (wc->autocommit_time_set)
2511                        extra_args += 2;
2512                if (wc->max_age_set)
2513                        extra_args += 2;
2514                if (wc->cleaner_set)
2515                        extra_args++;
2516                if (wc->writeback_fua_set)
2517                        extra_args++;
2518
2519                DMEMIT("%u", extra_args);
2520                if (wc->start_sector_set)
2521                        DMEMIT(" start_sector %llu", (unsigned long long)wc->start_sector);
2522                if (wc->high_wm_percent_set)
2523                        DMEMIT(" high_watermark %u", wc->high_wm_percent_value);
2524                if (wc->low_wm_percent_set)
2525                        DMEMIT(" low_watermark %u", wc->low_wm_percent_value);
2526                if (wc->max_writeback_jobs_set)
2527                        DMEMIT(" writeback_jobs %u", wc->max_writeback_jobs);
2528                if (wc->autocommit_blocks_set)
2529                        DMEMIT(" autocommit_blocks %u", wc->autocommit_blocks);
2530                if (wc->autocommit_time_set)
2531                        DMEMIT(" autocommit_time %u", wc->autocommit_time_value);
2532                if (wc->max_age_set)
2533                        DMEMIT(" max_age %u", wc->max_age_value);
2534                if (wc->cleaner_set)
2535                        DMEMIT(" cleaner");
2536                if (wc->writeback_fua_set)
2537                        DMEMIT(" %sfua", wc->writeback_fua ? "" : "no");
2538                break;
2539        }
2540}
2541
2542static struct target_type writecache_target = {
2543        .name                   = "writecache",
2544        .version                = {1, 4, 0},
2545        .module                 = THIS_MODULE,
2546        .ctr                    = writecache_ctr,
2547        .dtr                    = writecache_dtr,
2548        .status                 = writecache_status,
2549        .postsuspend            = writecache_suspend,
2550        .resume                 = writecache_resume,
2551        .message                = writecache_message,
2552        .map                    = writecache_map,
2553        .end_io                 = writecache_end_io,
2554        .iterate_devices        = writecache_iterate_devices,
2555        .io_hints               = writecache_io_hints,
2556};
2557
2558static int __init dm_writecache_init(void)
2559{
2560        int r;
2561
2562        r = dm_register_target(&writecache_target);
2563        if (r < 0) {
2564                DMERR("register failed %d", r);
2565                return r;
2566        }
2567
2568        return 0;
2569}
2570
2571static void __exit dm_writecache_exit(void)
2572{
2573        dm_unregister_target(&writecache_target);
2574}
2575
2576module_init(dm_writecache_init);
2577module_exit(dm_writecache_exit);
2578
2579MODULE_DESCRIPTION(DM_NAME " writecache target");
2580MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
2581MODULE_LICENSE("GPL");
2582