linux/drivers/md/dm-bufio.c
<<
>>
Prefs
   1/*
   2 * Copyright (C) 2009-2011 Red Hat, Inc.
   3 *
   4 * Author: Mikulas Patocka <mpatocka@redhat.com>
   5 *
   6 * This file is released under the GPL.
   7 */
   8
   9#include "dm-bufio.h"
  10
  11#include <linux/device-mapper.h>
  12#include <linux/dm-io.h>
  13#include <linux/slab.h>
  14#include <linux/vmalloc.h>
  15#include <linux/shrinker.h>
  16#include <linux/module.h>
  17
  18#define DM_MSG_PREFIX "bufio"
  19
  20/*
  21 * Memory management policy:
  22 *      Limit the number of buffers to DM_BUFIO_MEMORY_PERCENT of main memory
  23 *      or DM_BUFIO_VMALLOC_PERCENT of vmalloc memory (whichever is lower).
  24 *      Always allocate at least DM_BUFIO_MIN_BUFFERS buffers.
  25 *      Start background writeback when there are DM_BUFIO_WRITEBACK_PERCENT
  26 *      dirty buffers.
  27 */
  28#define DM_BUFIO_MIN_BUFFERS            8
  29
  30#define DM_BUFIO_MEMORY_PERCENT         2
  31#define DM_BUFIO_VMALLOC_PERCENT        25
  32#define DM_BUFIO_WRITEBACK_PERCENT      75
  33
  34/*
  35 * Check buffer ages in this interval (seconds)
  36 */
  37#define DM_BUFIO_WORK_TIMER_SECS        10
  38
  39/*
  40 * Free buffers when they are older than this (seconds)
  41 */
  42#define DM_BUFIO_DEFAULT_AGE_SECS       60
  43
  44/*
  45 * The number of bvec entries that are embedded directly in the buffer.
  46 * If the chunk size is larger, dm-io is used to do the io.
  47 */
  48#define DM_BUFIO_INLINE_VECS            16
  49
  50/*
  51 * Buffer hash
  52 */
  53#define DM_BUFIO_HASH_BITS      20
  54#define DM_BUFIO_HASH(block) \
  55        ((((block) >> DM_BUFIO_HASH_BITS) ^ (block)) & \
  56         ((1 << DM_BUFIO_HASH_BITS) - 1))
  57
  58/*
  59 * Don't try to use kmem_cache_alloc for blocks larger than this.
  60 * For explanation, see alloc_buffer_data below.
  61 */
  62#define DM_BUFIO_BLOCK_SIZE_SLAB_LIMIT  (PAGE_SIZE >> 1)
  63#define DM_BUFIO_BLOCK_SIZE_GFP_LIMIT   (PAGE_SIZE << (MAX_ORDER - 1))
  64
  65/*
  66 * dm_buffer->list_mode
  67 */
  68#define LIST_CLEAN      0
  69#define LIST_DIRTY      1
  70#define LIST_SIZE       2
  71
  72/*
  73 * Linking of buffers:
  74 *      All buffers are linked to cache_hash with their hash_list field.
  75 *
  76 *      Clean buffers that are not being written (B_WRITING not set)
  77 *      are linked to lru[LIST_CLEAN] with their lru_list field.
  78 *
  79 *      Dirty and clean buffers that are being written are linked to
  80 *      lru[LIST_DIRTY] with their lru_list field. When the write
  81 *      finishes, the buffer cannot be relinked immediately (because we
  82 *      are in an interrupt context and relinking requires process
  83 *      context), so some clean-not-writing buffers can be held on
  84 *      dirty_lru too.  They are later added to lru in the process
  85 *      context.
  86 */
  87struct dm_bufio_client {
  88        struct mutex lock;
  89
  90        struct list_head lru[LIST_SIZE];
  91        unsigned long n_buffers[LIST_SIZE];
  92
  93        struct block_device *bdev;
  94        unsigned block_size;
  95        unsigned char sectors_per_block_bits;
  96        unsigned char pages_per_block_bits;
  97        unsigned char blocks_per_page_bits;
  98        unsigned aux_size;
  99        void (*alloc_callback)(struct dm_buffer *);
 100        void (*write_callback)(struct dm_buffer *);
 101
 102        struct dm_io_client *dm_io;
 103
 104        struct list_head reserved_buffers;
 105        unsigned need_reserved_buffers;
 106
 107        struct hlist_head *cache_hash;
 108        wait_queue_head_t free_buffer_wait;
 109
 110        int async_write_error;
 111
 112        struct list_head client_list;
 113        struct shrinker shrinker;
 114};
 115
 116/*
 117 * Buffer state bits.
 118 */
 119#define B_READING       0
 120#define B_WRITING       1
 121#define B_DIRTY         2
 122
 123/*
 124 * Describes how the block was allocated:
 125 * kmem_cache_alloc(), __get_free_pages() or vmalloc().
 126 * See the comment at alloc_buffer_data.
 127 */
 128enum data_mode {
 129        DATA_MODE_SLAB = 0,
 130        DATA_MODE_GET_FREE_PAGES = 1,
 131        DATA_MODE_VMALLOC = 2,
 132        DATA_MODE_LIMIT = 3
 133};
 134
 135struct dm_buffer {
 136        struct hlist_node hash_list;
 137        struct list_head lru_list;
 138        sector_t block;
 139        void *data;
 140        enum data_mode data_mode;
 141        unsigned char list_mode;                /* LIST_* */
 142        unsigned hold_count;
 143        int read_error;
 144        int write_error;
 145        unsigned long state;
 146        unsigned long last_accessed;
 147        struct dm_bufio_client *c;
 148        struct bio bio;
 149        struct bio_vec bio_vec[DM_BUFIO_INLINE_VECS];
 150};
 151
 152/*----------------------------------------------------------------*/
 153
 154static struct kmem_cache *dm_bufio_caches[PAGE_SHIFT - SECTOR_SHIFT];
 155static char *dm_bufio_cache_names[PAGE_SHIFT - SECTOR_SHIFT];
 156
 157static inline int dm_bufio_cache_index(struct dm_bufio_client *c)
 158{
 159        unsigned ret = c->blocks_per_page_bits - 1;
 160
 161        BUG_ON(ret >= ARRAY_SIZE(dm_bufio_caches));
 162
 163        return ret;
 164}
 165
 166#define DM_BUFIO_CACHE(c)       (dm_bufio_caches[dm_bufio_cache_index(c)])
 167#define DM_BUFIO_CACHE_NAME(c)  (dm_bufio_cache_names[dm_bufio_cache_index(c)])
 168
 169#define dm_bufio_in_request()   (!!current->bio_list)
 170
 171static void dm_bufio_lock(struct dm_bufio_client *c)
 172{
 173        mutex_lock_nested(&c->lock, dm_bufio_in_request());
 174}
 175
 176static int dm_bufio_trylock(struct dm_bufio_client *c)
 177{
 178        return mutex_trylock(&c->lock);
 179}
 180
 181static void dm_bufio_unlock(struct dm_bufio_client *c)
 182{
 183        mutex_unlock(&c->lock);
 184}
 185
 186/*
 187 * FIXME Move to sched.h?
 188 */
 189#ifdef CONFIG_PREEMPT_VOLUNTARY
 190#  define dm_bufio_cond_resched()               \
 191do {                                            \
 192        if (unlikely(need_resched()))           \
 193                _cond_resched();                \
 194} while (0)
 195#else
 196#  define dm_bufio_cond_resched()                do { } while (0)
 197#endif
 198
 199/*----------------------------------------------------------------*/
 200
 201/*
 202 * Default cache size: available memory divided by the ratio.
 203 */
 204static unsigned long dm_bufio_default_cache_size;
 205
 206/*
 207 * Total cache size set by the user.
 208 */
 209static unsigned long dm_bufio_cache_size;
 210
 211/*
 212 * A copy of dm_bufio_cache_size because dm_bufio_cache_size can change
 213 * at any time.  If it disagrees, the user has changed cache size.
 214 */
 215static unsigned long dm_bufio_cache_size_latch;
 216
 217static DEFINE_SPINLOCK(param_spinlock);
 218
 219/*
 220 * Buffers are freed after this timeout
 221 */
 222static unsigned dm_bufio_max_age = DM_BUFIO_DEFAULT_AGE_SECS;
 223
 224static unsigned long dm_bufio_peak_allocated;
 225static unsigned long dm_bufio_allocated_kmem_cache;
 226static unsigned long dm_bufio_allocated_get_free_pages;
 227static unsigned long dm_bufio_allocated_vmalloc;
 228static unsigned long dm_bufio_current_allocated;
 229
 230/*----------------------------------------------------------------*/
 231
 232/*
 233 * Per-client cache: dm_bufio_cache_size / dm_bufio_client_count
 234 */
 235static unsigned long dm_bufio_cache_size_per_client;
 236
 237/*
 238 * The current number of clients.
 239 */
 240static int dm_bufio_client_count;
 241
 242/*
 243 * The list of all clients.
 244 */
 245static LIST_HEAD(dm_bufio_all_clients);
 246
 247/*
 248 * This mutex protects dm_bufio_cache_size_latch,
 249 * dm_bufio_cache_size_per_client and dm_bufio_client_count
 250 */
 251static DEFINE_MUTEX(dm_bufio_clients_lock);
 252
 253/*----------------------------------------------------------------*/
 254
 255static void adjust_total_allocated(enum data_mode data_mode, long diff)
 256{
 257        static unsigned long * const class_ptr[DATA_MODE_LIMIT] = {
 258                &dm_bufio_allocated_kmem_cache,
 259                &dm_bufio_allocated_get_free_pages,
 260                &dm_bufio_allocated_vmalloc,
 261        };
 262
 263        spin_lock(&param_spinlock);
 264
 265        *class_ptr[data_mode] += diff;
 266
 267        dm_bufio_current_allocated += diff;
 268
 269        if (dm_bufio_current_allocated > dm_bufio_peak_allocated)
 270                dm_bufio_peak_allocated = dm_bufio_current_allocated;
 271
 272        spin_unlock(&param_spinlock);
 273}
 274
 275/*
 276 * Change the number of clients and recalculate per-client limit.
 277 */
 278static void __cache_size_refresh(void)
 279{
 280        BUG_ON(!mutex_is_locked(&dm_bufio_clients_lock));
 281        BUG_ON(dm_bufio_client_count < 0);
 282
 283        dm_bufio_cache_size_latch = ACCESS_ONCE(dm_bufio_cache_size);
 284
 285        /*
 286         * Use default if set to 0 and report the actual cache size used.
 287         */
 288        if (!dm_bufio_cache_size_latch) {
 289                (void)cmpxchg(&dm_bufio_cache_size, 0,
 290                              dm_bufio_default_cache_size);
 291                dm_bufio_cache_size_latch = dm_bufio_default_cache_size;
 292        }
 293
 294        dm_bufio_cache_size_per_client = dm_bufio_cache_size_latch /
 295                                         (dm_bufio_client_count ? : 1);
 296}
 297
 298/*
 299 * Allocating buffer data.
 300 *
 301 * Small buffers are allocated with kmem_cache, to use space optimally.
 302 *
 303 * For large buffers, we choose between get_free_pages and vmalloc.
 304 * Each has advantages and disadvantages.
 305 *
 306 * __get_free_pages can randomly fail if the memory is fragmented.
 307 * __vmalloc won't randomly fail, but vmalloc space is limited (it may be
 308 * as low as 128M) so using it for caching is not appropriate.
 309 *
 310 * If the allocation may fail we use __get_free_pages. Memory fragmentation
 311 * won't have a fatal effect here, but it just causes flushes of some other
 312 * buffers and more I/O will be performed. Don't use __get_free_pages if it
 313 * always fails (i.e. order >= MAX_ORDER).
 314 *
 315 * If the allocation shouldn't fail we use __vmalloc. This is only for the
 316 * initial reserve allocation, so there's no risk of wasting all vmalloc
 317 * space.
 318 */
 319static void *alloc_buffer_data(struct dm_bufio_client *c, gfp_t gfp_mask,
 320                               enum data_mode *data_mode)
 321{
 322        unsigned noio_flag;
 323        void *ptr;
 324
 325        if (c->block_size <= DM_BUFIO_BLOCK_SIZE_SLAB_LIMIT) {
 326                *data_mode = DATA_MODE_SLAB;
 327                return kmem_cache_alloc(DM_BUFIO_CACHE(c), gfp_mask);
 328        }
 329
 330        if (c->block_size <= DM_BUFIO_BLOCK_SIZE_GFP_LIMIT &&
 331            gfp_mask & __GFP_NORETRY) {
 332                *data_mode = DATA_MODE_GET_FREE_PAGES;
 333                return (void *)__get_free_pages(gfp_mask,
 334                                                c->pages_per_block_bits);
 335        }
 336
 337        *data_mode = DATA_MODE_VMALLOC;
 338
 339        /*
 340         * __vmalloc allocates the data pages and auxiliary structures with
 341         * gfp_flags that were specified, but pagetables are always allocated
 342         * with GFP_KERNEL, no matter what was specified as gfp_mask.
 343         *
 344         * Consequently, we must set per-process flag PF_MEMALLOC_NOIO so that
 345         * all allocations done by this process (including pagetables) are done
 346         * as if GFP_NOIO was specified.
 347         */
 348
 349        if (gfp_mask & __GFP_NORETRY)
 350                noio_flag = memalloc_noio_save();
 351
 352        ptr = __vmalloc(c->block_size, gfp_mask, PAGE_KERNEL);
 353
 354        if (gfp_mask & __GFP_NORETRY)
 355                memalloc_noio_restore(noio_flag);
 356
 357        return ptr;
 358}
 359
 360/*
 361 * Free buffer's data.
 362 */
 363static void free_buffer_data(struct dm_bufio_client *c,
 364                             void *data, enum data_mode data_mode)
 365{
 366        switch (data_mode) {
 367        case DATA_MODE_SLAB:
 368                kmem_cache_free(DM_BUFIO_CACHE(c), data);
 369                break;
 370
 371        case DATA_MODE_GET_FREE_PAGES:
 372                free_pages((unsigned long)data, c->pages_per_block_bits);
 373                break;
 374
 375        case DATA_MODE_VMALLOC:
 376                vfree(data);
 377                break;
 378
 379        default:
 380                DMCRIT("dm_bufio_free_buffer_data: bad data mode: %d",
 381                       data_mode);
 382                BUG();
 383        }
 384}
 385
 386/*
 387 * Allocate buffer and its data.
 388 */
 389static struct dm_buffer *alloc_buffer(struct dm_bufio_client *c, gfp_t gfp_mask)
 390{
 391        struct dm_buffer *b = kmalloc(sizeof(struct dm_buffer) + c->aux_size,
 392                                      gfp_mask);
 393
 394        if (!b)
 395                return NULL;
 396
 397        b->c = c;
 398
 399        b->data = alloc_buffer_data(c, gfp_mask, &b->data_mode);
 400        if (!b->data) {
 401                kfree(b);
 402                return NULL;
 403        }
 404
 405        adjust_total_allocated(b->data_mode, (long)c->block_size);
 406
 407        return b;
 408}
 409
 410/*
 411 * Free buffer and its data.
 412 */
 413static void free_buffer(struct dm_buffer *b)
 414{
 415        struct dm_bufio_client *c = b->c;
 416
 417        adjust_total_allocated(b->data_mode, -(long)c->block_size);
 418
 419        free_buffer_data(c, b->data, b->data_mode);
 420        kfree(b);
 421}
 422
 423/*
 424 * Link buffer to the hash list and clean or dirty queue.
 425 */
 426static void __link_buffer(struct dm_buffer *b, sector_t block, int dirty)
 427{
 428        struct dm_bufio_client *c = b->c;
 429
 430        c->n_buffers[dirty]++;
 431        b->block = block;
 432        b->list_mode = dirty;
 433        list_add(&b->lru_list, &c->lru[dirty]);
 434        hlist_add_head(&b->hash_list, &c->cache_hash[DM_BUFIO_HASH(block)]);
 435        b->last_accessed = jiffies;
 436}
 437
 438/*
 439 * Unlink buffer from the hash list and dirty or clean queue.
 440 */
 441static void __unlink_buffer(struct dm_buffer *b)
 442{
 443        struct dm_bufio_client *c = b->c;
 444
 445        BUG_ON(!c->n_buffers[b->list_mode]);
 446
 447        c->n_buffers[b->list_mode]--;
 448        hlist_del(&b->hash_list);
 449        list_del(&b->lru_list);
 450}
 451
 452/*
 453 * Place the buffer to the head of dirty or clean LRU queue.
 454 */
 455static void __relink_lru(struct dm_buffer *b, int dirty)
 456{
 457        struct dm_bufio_client *c = b->c;
 458
 459        BUG_ON(!c->n_buffers[b->list_mode]);
 460
 461        c->n_buffers[b->list_mode]--;
 462        c->n_buffers[dirty]++;
 463        b->list_mode = dirty;
 464        list_move(&b->lru_list, &c->lru[dirty]);
 465}
 466
 467/*----------------------------------------------------------------
 468 * Submit I/O on the buffer.
 469 *
 470 * Bio interface is faster but it has some problems:
 471 *      the vector list is limited (increasing this limit increases
 472 *      memory-consumption per buffer, so it is not viable);
 473 *
 474 *      the memory must be direct-mapped, not vmalloced;
 475 *
 476 *      the I/O driver can reject requests spuriously if it thinks that
 477 *      the requests are too big for the device or if they cross a
 478 *      controller-defined memory boundary.
 479 *
 480 * If the buffer is small enough (up to DM_BUFIO_INLINE_VECS pages) and
 481 * it is not vmalloced, try using the bio interface.
 482 *
 483 * If the buffer is big, if it is vmalloced or if the underlying device
 484 * rejects the bio because it is too large, use dm-io layer to do the I/O.
 485 * The dm-io layer splits the I/O into multiple requests, avoiding the above
 486 * shortcomings.
 487 *--------------------------------------------------------------*/
 488
 489/*
 490 * dm-io completion routine. It just calls b->bio.bi_end_io, pretending
 491 * that the request was handled directly with bio interface.
 492 */
 493static void dmio_complete(unsigned long error, void *context)
 494{
 495        struct dm_buffer *b = context;
 496
 497        b->bio.bi_end_io(&b->bio, error ? -EIO : 0);
 498}
 499
 500static void use_dmio(struct dm_buffer *b, int rw, sector_t block,
 501                     bio_end_io_t *end_io)
 502{
 503        int r;
 504        struct dm_io_request io_req = {
 505                .bi_rw = rw,
 506                .notify.fn = dmio_complete,
 507                .notify.context = b,
 508                .client = b->c->dm_io,
 509        };
 510        struct dm_io_region region = {
 511                .bdev = b->c->bdev,
 512                .sector = block << b->c->sectors_per_block_bits,
 513                .count = b->c->block_size >> SECTOR_SHIFT,
 514        };
 515
 516        if (b->data_mode != DATA_MODE_VMALLOC) {
 517                io_req.mem.type = DM_IO_KMEM;
 518                io_req.mem.ptr.addr = b->data;
 519        } else {
 520                io_req.mem.type = DM_IO_VMA;
 521                io_req.mem.ptr.vma = b->data;
 522        }
 523
 524        b->bio.bi_end_io = end_io;
 525
 526        r = dm_io(&io_req, 1, &region, NULL);
 527        if (r)
 528                end_io(&b->bio, r);
 529}
 530
 531static void use_inline_bio(struct dm_buffer *b, int rw, sector_t block,
 532                           bio_end_io_t *end_io)
 533{
 534        char *ptr;
 535        int len;
 536
 537        bio_init(&b->bio);
 538        b->bio.bi_io_vec = b->bio_vec;
 539        b->bio.bi_max_vecs = DM_BUFIO_INLINE_VECS;
 540        b->bio.bi_sector = block << b->c->sectors_per_block_bits;
 541        b->bio.bi_bdev = b->c->bdev;
 542        b->bio.bi_end_io = end_io;
 543
 544        /*
 545         * We assume that if len >= PAGE_SIZE ptr is page-aligned.
 546         * If len < PAGE_SIZE the buffer doesn't cross page boundary.
 547         */
 548        ptr = b->data;
 549        len = b->c->block_size;
 550
 551        if (len >= PAGE_SIZE)
 552                BUG_ON((unsigned long)ptr & (PAGE_SIZE - 1));
 553        else
 554                BUG_ON((unsigned long)ptr & (len - 1));
 555
 556        do {
 557                if (!bio_add_page(&b->bio, virt_to_page(ptr),
 558                                  len < PAGE_SIZE ? len : PAGE_SIZE,
 559                                  virt_to_phys(ptr) & (PAGE_SIZE - 1))) {
 560                        BUG_ON(b->c->block_size <= PAGE_SIZE);
 561                        use_dmio(b, rw, block, end_io);
 562                        return;
 563                }
 564
 565                len -= PAGE_SIZE;
 566                ptr += PAGE_SIZE;
 567        } while (len > 0);
 568
 569        submit_bio(rw, &b->bio);
 570}
 571
 572static void submit_io(struct dm_buffer *b, int rw, sector_t block,
 573                      bio_end_io_t *end_io)
 574{
 575        if (rw == WRITE && b->c->write_callback)
 576                b->c->write_callback(b);
 577
 578        if (b->c->block_size <= DM_BUFIO_INLINE_VECS * PAGE_SIZE &&
 579            b->data_mode != DATA_MODE_VMALLOC)
 580                use_inline_bio(b, rw, block, end_io);
 581        else
 582                use_dmio(b, rw, block, end_io);
 583}
 584
 585/*----------------------------------------------------------------
 586 * Writing dirty buffers
 587 *--------------------------------------------------------------*/
 588
 589/*
 590 * The endio routine for write.
 591 *
 592 * Set the error, clear B_WRITING bit and wake anyone who was waiting on
 593 * it.
 594 */
 595static void write_endio(struct bio *bio, int error)
 596{
 597        struct dm_buffer *b = container_of(bio, struct dm_buffer, bio);
 598
 599        b->write_error = error;
 600        if (unlikely(error)) {
 601                struct dm_bufio_client *c = b->c;
 602                (void)cmpxchg(&c->async_write_error, 0, error);
 603        }
 604
 605        BUG_ON(!test_bit(B_WRITING, &b->state));
 606
 607        smp_mb__before_clear_bit();
 608        clear_bit(B_WRITING, &b->state);
 609        smp_mb__after_clear_bit();
 610
 611        wake_up_bit(&b->state, B_WRITING);
 612}
 613
 614/*
 615 * This function is called when wait_on_bit is actually waiting.
 616 */
 617static int do_io_schedule(void *word)
 618{
 619        io_schedule();
 620
 621        return 0;
 622}
 623
 624/*
 625 * Initiate a write on a dirty buffer, but don't wait for it.
 626 *
 627 * - If the buffer is not dirty, exit.
 628 * - If there some previous write going on, wait for it to finish (we can't
 629 *   have two writes on the same buffer simultaneously).
 630 * - Submit our write and don't wait on it. We set B_WRITING indicating
 631 *   that there is a write in progress.
 632 */
 633static void __write_dirty_buffer(struct dm_buffer *b)
 634{
 635        if (!test_bit(B_DIRTY, &b->state))
 636                return;
 637
 638        clear_bit(B_DIRTY, &b->state);
 639        wait_on_bit_lock(&b->state, B_WRITING,
 640                         do_io_schedule, TASK_UNINTERRUPTIBLE);
 641
 642        submit_io(b, WRITE, b->block, write_endio);
 643}
 644
 645/*
 646 * Wait until any activity on the buffer finishes.  Possibly write the
 647 * buffer if it is dirty.  When this function finishes, there is no I/O
 648 * running on the buffer and the buffer is not dirty.
 649 */
 650static void __make_buffer_clean(struct dm_buffer *b)
 651{
 652        BUG_ON(b->hold_count);
 653
 654        if (!b->state)  /* fast case */
 655                return;
 656
 657        wait_on_bit(&b->state, B_READING, do_io_schedule, TASK_UNINTERRUPTIBLE);
 658        __write_dirty_buffer(b);
 659        wait_on_bit(&b->state, B_WRITING, do_io_schedule, TASK_UNINTERRUPTIBLE);
 660}
 661
 662/*
 663 * Find some buffer that is not held by anybody, clean it, unlink it and
 664 * return it.
 665 */
 666static struct dm_buffer *__get_unclaimed_buffer(struct dm_bufio_client *c)
 667{
 668        struct dm_buffer *b;
 669
 670        list_for_each_entry_reverse(b, &c->lru[LIST_CLEAN], lru_list) {
 671                BUG_ON(test_bit(B_WRITING, &b->state));
 672                BUG_ON(test_bit(B_DIRTY, &b->state));
 673
 674                if (!b->hold_count) {
 675                        __make_buffer_clean(b);
 676                        __unlink_buffer(b);
 677                        return b;
 678                }
 679                dm_bufio_cond_resched();
 680        }
 681
 682        list_for_each_entry_reverse(b, &c->lru[LIST_DIRTY], lru_list) {
 683                BUG_ON(test_bit(B_READING, &b->state));
 684
 685                if (!b->hold_count) {
 686                        __make_buffer_clean(b);
 687                        __unlink_buffer(b);
 688                        return b;
 689                }
 690                dm_bufio_cond_resched();
 691        }
 692
 693        return NULL;
 694}
 695
 696/*
 697 * Wait until some other threads free some buffer or release hold count on
 698 * some buffer.
 699 *
 700 * This function is entered with c->lock held, drops it and regains it
 701 * before exiting.
 702 */
 703static void __wait_for_free_buffer(struct dm_bufio_client *c)
 704{
 705        DECLARE_WAITQUEUE(wait, current);
 706
 707        add_wait_queue(&c->free_buffer_wait, &wait);
 708        set_task_state(current, TASK_UNINTERRUPTIBLE);
 709        dm_bufio_unlock(c);
 710
 711        io_schedule();
 712
 713        set_task_state(current, TASK_RUNNING);
 714        remove_wait_queue(&c->free_buffer_wait, &wait);
 715
 716        dm_bufio_lock(c);
 717}
 718
 719enum new_flag {
 720        NF_FRESH = 0,
 721        NF_READ = 1,
 722        NF_GET = 2,
 723        NF_PREFETCH = 3
 724};
 725
 726/*
 727 * Allocate a new buffer. If the allocation is not possible, wait until
 728 * some other thread frees a buffer.
 729 *
 730 * May drop the lock and regain it.
 731 */
 732static struct dm_buffer *__alloc_buffer_wait_no_callback(struct dm_bufio_client *c, enum new_flag nf)
 733{
 734        struct dm_buffer *b;
 735
 736        /*
 737         * dm-bufio is resistant to allocation failures (it just keeps
 738         * one buffer reserved in cases all the allocations fail).
 739         * So set flags to not try too hard:
 740         *      GFP_NOIO: don't recurse into the I/O layer
 741         *      __GFP_NORETRY: don't retry and rather return failure
 742         *      __GFP_NOMEMALLOC: don't use emergency reserves
 743         *      __GFP_NOWARN: don't print a warning in case of failure
 744         *
 745         * For debugging, if we set the cache size to 1, no new buffers will
 746         * be allocated.
 747         */
 748        while (1) {
 749                if (dm_bufio_cache_size_latch != 1) {
 750                        b = alloc_buffer(c, GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
 751                        if (b)
 752                                return b;
 753                }
 754
 755                if (nf == NF_PREFETCH)
 756                        return NULL;
 757
 758                if (!list_empty(&c->reserved_buffers)) {
 759                        b = list_entry(c->reserved_buffers.next,
 760                                       struct dm_buffer, lru_list);
 761                        list_del(&b->lru_list);
 762                        c->need_reserved_buffers++;
 763
 764                        return b;
 765                }
 766
 767                b = __get_unclaimed_buffer(c);
 768                if (b)
 769                        return b;
 770
 771                __wait_for_free_buffer(c);
 772        }
 773}
 774
 775static struct dm_buffer *__alloc_buffer_wait(struct dm_bufio_client *c, enum new_flag nf)
 776{
 777        struct dm_buffer *b = __alloc_buffer_wait_no_callback(c, nf);
 778
 779        if (!b)
 780                return NULL;
 781
 782        if (c->alloc_callback)
 783                c->alloc_callback(b);
 784
 785        return b;
 786}
 787
 788/*
 789 * Free a buffer and wake other threads waiting for free buffers.
 790 */
 791static void __free_buffer_wake(struct dm_buffer *b)
 792{
 793        struct dm_bufio_client *c = b->c;
 794
 795        if (!c->need_reserved_buffers)
 796                free_buffer(b);
 797        else {
 798                list_add(&b->lru_list, &c->reserved_buffers);
 799                c->need_reserved_buffers--;
 800        }
 801
 802        wake_up(&c->free_buffer_wait);
 803}
 804
 805static void __write_dirty_buffers_async(struct dm_bufio_client *c, int no_wait)
 806{
 807        struct dm_buffer *b, *tmp;
 808
 809        list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
 810                BUG_ON(test_bit(B_READING, &b->state));
 811
 812                if (!test_bit(B_DIRTY, &b->state) &&
 813                    !test_bit(B_WRITING, &b->state)) {
 814                        __relink_lru(b, LIST_CLEAN);
 815                        continue;
 816                }
 817
 818                if (no_wait && test_bit(B_WRITING, &b->state))
 819                        return;
 820
 821                __write_dirty_buffer(b);
 822                dm_bufio_cond_resched();
 823        }
 824}
 825
 826/*
 827 * Get writeback threshold and buffer limit for a given client.
 828 */
 829static void __get_memory_limit(struct dm_bufio_client *c,
 830                               unsigned long *threshold_buffers,
 831                               unsigned long *limit_buffers)
 832{
 833        unsigned long buffers;
 834
 835        if (ACCESS_ONCE(dm_bufio_cache_size) != dm_bufio_cache_size_latch) {
 836                mutex_lock(&dm_bufio_clients_lock);
 837                __cache_size_refresh();
 838                mutex_unlock(&dm_bufio_clients_lock);
 839        }
 840
 841        buffers = dm_bufio_cache_size_per_client >>
 842                  (c->sectors_per_block_bits + SECTOR_SHIFT);
 843
 844        if (buffers < DM_BUFIO_MIN_BUFFERS)
 845                buffers = DM_BUFIO_MIN_BUFFERS;
 846
 847        *limit_buffers = buffers;
 848        *threshold_buffers = buffers * DM_BUFIO_WRITEBACK_PERCENT / 100;
 849}
 850
 851/*
 852 * Check if we're over watermark.
 853 * If we are over threshold_buffers, start freeing buffers.
 854 * If we're over "limit_buffers", block until we get under the limit.
 855 */
 856static void __check_watermark(struct dm_bufio_client *c)
 857{
 858        unsigned long threshold_buffers, limit_buffers;
 859
 860        __get_memory_limit(c, &threshold_buffers, &limit_buffers);
 861
 862        while (c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY] >
 863               limit_buffers) {
 864
 865                struct dm_buffer *b = __get_unclaimed_buffer(c);
 866
 867                if (!b)
 868                        return;
 869
 870                __free_buffer_wake(b);
 871                dm_bufio_cond_resched();
 872        }
 873
 874        if (c->n_buffers[LIST_DIRTY] > threshold_buffers)
 875                __write_dirty_buffers_async(c, 1);
 876}
 877
 878/*
 879 * Find a buffer in the hash.
 880 */
 881static struct dm_buffer *__find(struct dm_bufio_client *c, sector_t block)
 882{
 883        struct dm_buffer *b;
 884
 885        hlist_for_each_entry(b, &c->cache_hash[DM_BUFIO_HASH(block)],
 886                             hash_list) {
 887                dm_bufio_cond_resched();
 888                if (b->block == block)
 889                        return b;
 890        }
 891
 892        return NULL;
 893}
 894
 895/*----------------------------------------------------------------
 896 * Getting a buffer
 897 *--------------------------------------------------------------*/
 898
 899static struct dm_buffer *__bufio_new(struct dm_bufio_client *c, sector_t block,
 900                                     enum new_flag nf, int *need_submit)
 901{
 902        struct dm_buffer *b, *new_b = NULL;
 903
 904        *need_submit = 0;
 905
 906        b = __find(c, block);
 907        if (b)
 908                goto found_buffer;
 909
 910        if (nf == NF_GET)
 911                return NULL;
 912
 913        new_b = __alloc_buffer_wait(c, nf);
 914        if (!new_b)
 915                return NULL;
 916
 917        /*
 918         * We've had a period where the mutex was unlocked, so need to
 919         * recheck the hash table.
 920         */
 921        b = __find(c, block);
 922        if (b) {
 923                __free_buffer_wake(new_b);
 924                goto found_buffer;
 925        }
 926
 927        __check_watermark(c);
 928
 929        b = new_b;
 930        b->hold_count = 1;
 931        b->read_error = 0;
 932        b->write_error = 0;
 933        __link_buffer(b, block, LIST_CLEAN);
 934
 935        if (nf == NF_FRESH) {
 936                b->state = 0;
 937                return b;
 938        }
 939
 940        b->state = 1 << B_READING;
 941        *need_submit = 1;
 942
 943        return b;
 944
 945found_buffer:
 946        if (nf == NF_PREFETCH)
 947                return NULL;
 948        /*
 949         * Note: it is essential that we don't wait for the buffer to be
 950         * read if dm_bufio_get function is used. Both dm_bufio_get and
 951         * dm_bufio_prefetch can be used in the driver request routine.
 952         * If the user called both dm_bufio_prefetch and dm_bufio_get on
 953         * the same buffer, it would deadlock if we waited.
 954         */
 955        if (nf == NF_GET && unlikely(test_bit(B_READING, &b->state)))
 956                return NULL;
 957
 958        b->hold_count++;
 959        __relink_lru(b, test_bit(B_DIRTY, &b->state) ||
 960                     test_bit(B_WRITING, &b->state));
 961        return b;
 962}
 963
 964/*
 965 * The endio routine for reading: set the error, clear the bit and wake up
 966 * anyone waiting on the buffer.
 967 */
 968static void read_endio(struct bio *bio, int error)
 969{
 970        struct dm_buffer *b = container_of(bio, struct dm_buffer, bio);
 971
 972        b->read_error = error;
 973
 974        BUG_ON(!test_bit(B_READING, &b->state));
 975
 976        smp_mb__before_clear_bit();
 977        clear_bit(B_READING, &b->state);
 978        smp_mb__after_clear_bit();
 979
 980        wake_up_bit(&b->state, B_READING);
 981}
 982
 983/*
 984 * A common routine for dm_bufio_new and dm_bufio_read.  Operation of these
 985 * functions is similar except that dm_bufio_new doesn't read the
 986 * buffer from the disk (assuming that the caller overwrites all the data
 987 * and uses dm_bufio_mark_buffer_dirty to write new data back).
 988 */
 989static void *new_read(struct dm_bufio_client *c, sector_t block,
 990                      enum new_flag nf, struct dm_buffer **bp)
 991{
 992        int need_submit;
 993        struct dm_buffer *b;
 994
 995        dm_bufio_lock(c);
 996        b = __bufio_new(c, block, nf, &need_submit);
 997        dm_bufio_unlock(c);
 998
 999        if (!b)
1000                return b;
1001
1002        if (need_submit)
1003                submit_io(b, READ, b->block, read_endio);
1004
1005        wait_on_bit(&b->state, B_READING, do_io_schedule, TASK_UNINTERRUPTIBLE);
1006
1007        if (b->read_error) {
1008                int error = b->read_error;
1009
1010                dm_bufio_release(b);
1011
1012                return ERR_PTR(error);
1013        }
1014
1015        *bp = b;
1016
1017        return b->data;
1018}
1019
1020void *dm_bufio_get(struct dm_bufio_client *c, sector_t block,
1021                   struct dm_buffer **bp)
1022{
1023        return new_read(c, block, NF_GET, bp);
1024}
1025EXPORT_SYMBOL_GPL(dm_bufio_get);
1026
1027void *dm_bufio_read(struct dm_bufio_client *c, sector_t block,
1028                    struct dm_buffer **bp)
1029{
1030        BUG_ON(dm_bufio_in_request());
1031
1032        return new_read(c, block, NF_READ, bp);
1033}
1034EXPORT_SYMBOL_GPL(dm_bufio_read);
1035
1036void *dm_bufio_new(struct dm_bufio_client *c, sector_t block,
1037                   struct dm_buffer **bp)
1038{
1039        BUG_ON(dm_bufio_in_request());
1040
1041        return new_read(c, block, NF_FRESH, bp);
1042}
1043EXPORT_SYMBOL_GPL(dm_bufio_new);
1044
1045void dm_bufio_prefetch(struct dm_bufio_client *c,
1046                       sector_t block, unsigned n_blocks)
1047{
1048        struct blk_plug plug;
1049
1050        BUG_ON(dm_bufio_in_request());
1051
1052        blk_start_plug(&plug);
1053        dm_bufio_lock(c);
1054
1055        for (; n_blocks--; block++) {
1056                int need_submit;
1057                struct dm_buffer *b;
1058                b = __bufio_new(c, block, NF_PREFETCH, &need_submit);
1059                if (unlikely(b != NULL)) {
1060                        dm_bufio_unlock(c);
1061
1062                        if (need_submit)
1063                                submit_io(b, READ, b->block, read_endio);
1064                        dm_bufio_release(b);
1065
1066                        dm_bufio_cond_resched();
1067
1068                        if (!n_blocks)
1069                                goto flush_plug;
1070                        dm_bufio_lock(c);
1071                }
1072
1073        }
1074
1075        dm_bufio_unlock(c);
1076
1077flush_plug:
1078        blk_finish_plug(&plug);
1079}
1080EXPORT_SYMBOL_GPL(dm_bufio_prefetch);
1081
1082void dm_bufio_release(struct dm_buffer *b)
1083{
1084        struct dm_bufio_client *c = b->c;
1085
1086        dm_bufio_lock(c);
1087
1088        BUG_ON(!b->hold_count);
1089
1090        b->hold_count--;
1091        if (!b->hold_count) {
1092                wake_up(&c->free_buffer_wait);
1093
1094                /*
1095                 * If there were errors on the buffer, and the buffer is not
1096                 * to be written, free the buffer. There is no point in caching
1097                 * invalid buffer.
1098                 */
1099                if ((b->read_error || b->write_error) &&
1100                    !test_bit(B_READING, &b->state) &&
1101                    !test_bit(B_WRITING, &b->state) &&
1102                    !test_bit(B_DIRTY, &b->state)) {
1103                        __unlink_buffer(b);
1104                        __free_buffer_wake(b);
1105                }
1106        }
1107
1108        dm_bufio_unlock(c);
1109}
1110EXPORT_SYMBOL_GPL(dm_bufio_release);
1111
1112void dm_bufio_mark_buffer_dirty(struct dm_buffer *b)
1113{
1114        struct dm_bufio_client *c = b->c;
1115
1116        dm_bufio_lock(c);
1117
1118        BUG_ON(test_bit(B_READING, &b->state));
1119
1120        if (!test_and_set_bit(B_DIRTY, &b->state))
1121                __relink_lru(b, LIST_DIRTY);
1122
1123        dm_bufio_unlock(c);
1124}
1125EXPORT_SYMBOL_GPL(dm_bufio_mark_buffer_dirty);
1126
1127void dm_bufio_write_dirty_buffers_async(struct dm_bufio_client *c)
1128{
1129        BUG_ON(dm_bufio_in_request());
1130
1131        dm_bufio_lock(c);
1132        __write_dirty_buffers_async(c, 0);
1133        dm_bufio_unlock(c);
1134}
1135EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers_async);
1136
1137/*
1138 * For performance, it is essential that the buffers are written asynchronously
1139 * and simultaneously (so that the block layer can merge the writes) and then
1140 * waited upon.
1141 *
1142 * Finally, we flush hardware disk cache.
1143 */
1144int dm_bufio_write_dirty_buffers(struct dm_bufio_client *c)
1145{
1146        int a, f;
1147        unsigned long buffers_processed = 0;
1148        struct dm_buffer *b, *tmp;
1149
1150        dm_bufio_lock(c);
1151        __write_dirty_buffers_async(c, 0);
1152
1153again:
1154        list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
1155                int dropped_lock = 0;
1156
1157                if (buffers_processed < c->n_buffers[LIST_DIRTY])
1158                        buffers_processed++;
1159
1160                BUG_ON(test_bit(B_READING, &b->state));
1161
1162                if (test_bit(B_WRITING, &b->state)) {
1163                        if (buffers_processed < c->n_buffers[LIST_DIRTY]) {
1164                                dropped_lock = 1;
1165                                b->hold_count++;
1166                                dm_bufio_unlock(c);
1167                                wait_on_bit(&b->state, B_WRITING,
1168                                            do_io_schedule,
1169                                            TASK_UNINTERRUPTIBLE);
1170                                dm_bufio_lock(c);
1171                                b->hold_count--;
1172                        } else
1173                                wait_on_bit(&b->state, B_WRITING,
1174                                            do_io_schedule,
1175                                            TASK_UNINTERRUPTIBLE);
1176                }
1177
1178                if (!test_bit(B_DIRTY, &b->state) &&
1179                    !test_bit(B_WRITING, &b->state))
1180                        __relink_lru(b, LIST_CLEAN);
1181
1182                dm_bufio_cond_resched();
1183
1184                /*
1185                 * If we dropped the lock, the list is no longer consistent,
1186                 * so we must restart the search.
1187                 *
1188                 * In the most common case, the buffer just processed is
1189                 * relinked to the clean list, so we won't loop scanning the
1190                 * same buffer again and again.
1191                 *
1192                 * This may livelock if there is another thread simultaneously
1193                 * dirtying buffers, so we count the number of buffers walked
1194                 * and if it exceeds the total number of buffers, it means that
1195                 * someone is doing some writes simultaneously with us.  In
1196                 * this case, stop, dropping the lock.
1197                 */
1198                if (dropped_lock)
1199                        goto again;
1200        }
1201        wake_up(&c->free_buffer_wait);
1202        dm_bufio_unlock(c);
1203
1204        a = xchg(&c->async_write_error, 0);
1205        f = dm_bufio_issue_flush(c);
1206        if (a)
1207                return a;
1208
1209        return f;
1210}
1211EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers);
1212
1213/*
1214 * Use dm-io to send and empty barrier flush the device.
1215 */
1216int dm_bufio_issue_flush(struct dm_bufio_client *c)
1217{
1218        struct dm_io_request io_req = {
1219                .bi_rw = WRITE_FLUSH,
1220                .mem.type = DM_IO_KMEM,
1221                .mem.ptr.addr = NULL,
1222                .client = c->dm_io,
1223        };
1224        struct dm_io_region io_reg = {
1225                .bdev = c->bdev,
1226                .sector = 0,
1227                .count = 0,
1228        };
1229
1230        BUG_ON(dm_bufio_in_request());
1231
1232        return dm_io(&io_req, 1, &io_reg, NULL);
1233}
1234EXPORT_SYMBOL_GPL(dm_bufio_issue_flush);
1235
1236/*
1237 * We first delete any other buffer that may be at that new location.
1238 *
1239 * Then, we write the buffer to the original location if it was dirty.
1240 *
1241 * Then, if we are the only one who is holding the buffer, relink the buffer
1242 * in the hash queue for the new location.
1243 *
1244 * If there was someone else holding the buffer, we write it to the new
1245 * location but not relink it, because that other user needs to have the buffer
1246 * at the same place.
1247 */
1248void dm_bufio_release_move(struct dm_buffer *b, sector_t new_block)
1249{
1250        struct dm_bufio_client *c = b->c;
1251        struct dm_buffer *new;
1252
1253        BUG_ON(dm_bufio_in_request());
1254
1255        dm_bufio_lock(c);
1256
1257retry:
1258        new = __find(c, new_block);
1259        if (new) {
1260                if (new->hold_count) {
1261                        __wait_for_free_buffer(c);
1262                        goto retry;
1263                }
1264
1265                /*
1266                 * FIXME: Is there any point waiting for a write that's going
1267                 * to be overwritten in a bit?
1268                 */
1269                __make_buffer_clean(new);
1270                __unlink_buffer(new);
1271                __free_buffer_wake(new);
1272        }
1273
1274        BUG_ON(!b->hold_count);
1275        BUG_ON(test_bit(B_READING, &b->state));
1276
1277        __write_dirty_buffer(b);
1278        if (b->hold_count == 1) {
1279                wait_on_bit(&b->state, B_WRITING,
1280                            do_io_schedule, TASK_UNINTERRUPTIBLE);
1281                set_bit(B_DIRTY, &b->state);
1282                __unlink_buffer(b);
1283                __link_buffer(b, new_block, LIST_DIRTY);
1284        } else {
1285                sector_t old_block;
1286                wait_on_bit_lock(&b->state, B_WRITING,
1287                                 do_io_schedule, TASK_UNINTERRUPTIBLE);
1288                /*
1289                 * Relink buffer to "new_block" so that write_callback
1290                 * sees "new_block" as a block number.
1291                 * After the write, link the buffer back to old_block.
1292                 * All this must be done in bufio lock, so that block number
1293                 * change isn't visible to other threads.
1294                 */
1295                old_block = b->block;
1296                __unlink_buffer(b);
1297                __link_buffer(b, new_block, b->list_mode);
1298                submit_io(b, WRITE, new_block, write_endio);
1299                wait_on_bit(&b->state, B_WRITING,
1300                            do_io_schedule, TASK_UNINTERRUPTIBLE);
1301                __unlink_buffer(b);
1302                __link_buffer(b, old_block, b->list_mode);
1303        }
1304
1305        dm_bufio_unlock(c);
1306        dm_bufio_release(b);
1307}
1308EXPORT_SYMBOL_GPL(dm_bufio_release_move);
1309
1310unsigned dm_bufio_get_block_size(struct dm_bufio_client *c)
1311{
1312        return c->block_size;
1313}
1314EXPORT_SYMBOL_GPL(dm_bufio_get_block_size);
1315
1316sector_t dm_bufio_get_device_size(struct dm_bufio_client *c)
1317{
1318        return i_size_read(c->bdev->bd_inode) >>
1319                           (SECTOR_SHIFT + c->sectors_per_block_bits);
1320}
1321EXPORT_SYMBOL_GPL(dm_bufio_get_device_size);
1322
1323sector_t dm_bufio_get_block_number(struct dm_buffer *b)
1324{
1325        return b->block;
1326}
1327EXPORT_SYMBOL_GPL(dm_bufio_get_block_number);
1328
1329void *dm_bufio_get_block_data(struct dm_buffer *b)
1330{
1331        return b->data;
1332}
1333EXPORT_SYMBOL_GPL(dm_bufio_get_block_data);
1334
1335void *dm_bufio_get_aux_data(struct dm_buffer *b)
1336{
1337        return b + 1;
1338}
1339EXPORT_SYMBOL_GPL(dm_bufio_get_aux_data);
1340
1341struct dm_bufio_client *dm_bufio_get_client(struct dm_buffer *b)
1342{
1343        return b->c;
1344}
1345EXPORT_SYMBOL_GPL(dm_bufio_get_client);
1346
1347static void drop_buffers(struct dm_bufio_client *c)
1348{
1349        struct dm_buffer *b;
1350        int i;
1351
1352        BUG_ON(dm_bufio_in_request());
1353
1354        /*
1355         * An optimization so that the buffers are not written one-by-one.
1356         */
1357        dm_bufio_write_dirty_buffers_async(c);
1358
1359        dm_bufio_lock(c);
1360
1361        while ((b = __get_unclaimed_buffer(c)))
1362                __free_buffer_wake(b);
1363
1364        for (i = 0; i < LIST_SIZE; i++)
1365                list_for_each_entry(b, &c->lru[i], lru_list)
1366                        DMERR("leaked buffer %llx, hold count %u, list %d",
1367                              (unsigned long long)b->block, b->hold_count, i);
1368
1369        for (i = 0; i < LIST_SIZE; i++)
1370                BUG_ON(!list_empty(&c->lru[i]));
1371
1372        dm_bufio_unlock(c);
1373}
1374
1375/*
1376 * Test if the buffer is unused and too old, and commit it.
1377 * At if noio is set, we must not do any I/O because we hold
1378 * dm_bufio_clients_lock and we would risk deadlock if the I/O gets rerouted to
1379 * different bufio client.
1380 */
1381static int __cleanup_old_buffer(struct dm_buffer *b, gfp_t gfp,
1382                                unsigned long max_jiffies)
1383{
1384        if (jiffies - b->last_accessed < max_jiffies)
1385                return 1;
1386
1387        if (!(gfp & __GFP_IO)) {
1388                if (test_bit(B_READING, &b->state) ||
1389                    test_bit(B_WRITING, &b->state) ||
1390                    test_bit(B_DIRTY, &b->state))
1391                        return 1;
1392        }
1393
1394        if (b->hold_count)
1395                return 1;
1396
1397        __make_buffer_clean(b);
1398        __unlink_buffer(b);
1399        __free_buffer_wake(b);
1400
1401        return 0;
1402}
1403
1404static void __scan(struct dm_bufio_client *c, unsigned long nr_to_scan,
1405                   struct shrink_control *sc)
1406{
1407        int l;
1408        struct dm_buffer *b, *tmp;
1409
1410        for (l = 0; l < LIST_SIZE; l++) {
1411                list_for_each_entry_safe_reverse(b, tmp, &c->lru[l], lru_list)
1412                        if (!__cleanup_old_buffer(b, sc->gfp_mask, 0) &&
1413                            !--nr_to_scan)
1414                                return;
1415                dm_bufio_cond_resched();
1416        }
1417}
1418
1419static int shrink(struct shrinker *shrinker, struct shrink_control *sc)
1420{
1421        struct dm_bufio_client *c =
1422            container_of(shrinker, struct dm_bufio_client, shrinker);
1423        unsigned long r;
1424        unsigned long nr_to_scan = sc->nr_to_scan;
1425
1426        if (sc->gfp_mask & __GFP_IO)
1427                dm_bufio_lock(c);
1428        else if (!dm_bufio_trylock(c))
1429                return !nr_to_scan ? 0 : -1;
1430
1431        if (nr_to_scan)
1432                __scan(c, nr_to_scan, sc);
1433
1434        r = c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY];
1435        if (r > INT_MAX)
1436                r = INT_MAX;
1437
1438        dm_bufio_unlock(c);
1439
1440        return r;
1441}
1442
1443/*
1444 * Create the buffering interface
1445 */
1446struct dm_bufio_client *dm_bufio_client_create(struct block_device *bdev, unsigned block_size,
1447                                               unsigned reserved_buffers, unsigned aux_size,
1448                                               void (*alloc_callback)(struct dm_buffer *),
1449                                               void (*write_callback)(struct dm_buffer *))
1450{
1451        int r;
1452        struct dm_bufio_client *c;
1453        unsigned i;
1454
1455        BUG_ON(block_size < 1 << SECTOR_SHIFT ||
1456               (block_size & (block_size - 1)));
1457
1458        c = kmalloc(sizeof(*c), GFP_KERNEL);
1459        if (!c) {
1460                r = -ENOMEM;
1461                goto bad_client;
1462        }
1463        c->cache_hash = vmalloc(sizeof(struct hlist_head) << DM_BUFIO_HASH_BITS);
1464        if (!c->cache_hash) {
1465                r = -ENOMEM;
1466                goto bad_hash;
1467        }
1468
1469        c->bdev = bdev;
1470        c->block_size = block_size;
1471        c->sectors_per_block_bits = ffs(block_size) - 1 - SECTOR_SHIFT;
1472        c->pages_per_block_bits = (ffs(block_size) - 1 >= PAGE_SHIFT) ?
1473                                  ffs(block_size) - 1 - PAGE_SHIFT : 0;
1474        c->blocks_per_page_bits = (ffs(block_size) - 1 < PAGE_SHIFT ?
1475                                  PAGE_SHIFT - (ffs(block_size) - 1) : 0);
1476
1477        c->aux_size = aux_size;
1478        c->alloc_callback = alloc_callback;
1479        c->write_callback = write_callback;
1480
1481        for (i = 0; i < LIST_SIZE; i++) {
1482                INIT_LIST_HEAD(&c->lru[i]);
1483                c->n_buffers[i] = 0;
1484        }
1485
1486        for (i = 0; i < 1 << DM_BUFIO_HASH_BITS; i++)
1487                INIT_HLIST_HEAD(&c->cache_hash[i]);
1488
1489        mutex_init(&c->lock);
1490        INIT_LIST_HEAD(&c->reserved_buffers);
1491        c->need_reserved_buffers = reserved_buffers;
1492
1493        init_waitqueue_head(&c->free_buffer_wait);
1494        c->async_write_error = 0;
1495
1496        c->dm_io = dm_io_client_create();
1497        if (IS_ERR(c->dm_io)) {
1498                r = PTR_ERR(c->dm_io);
1499                goto bad_dm_io;
1500        }
1501
1502        mutex_lock(&dm_bufio_clients_lock);
1503        if (c->blocks_per_page_bits) {
1504                if (!DM_BUFIO_CACHE_NAME(c)) {
1505                        DM_BUFIO_CACHE_NAME(c) = kasprintf(GFP_KERNEL, "dm_bufio_cache-%u", c->block_size);
1506                        if (!DM_BUFIO_CACHE_NAME(c)) {
1507                                r = -ENOMEM;
1508                                mutex_unlock(&dm_bufio_clients_lock);
1509                                goto bad_cache;
1510                        }
1511                }
1512
1513                if (!DM_BUFIO_CACHE(c)) {
1514                        DM_BUFIO_CACHE(c) = kmem_cache_create(DM_BUFIO_CACHE_NAME(c),
1515                                                              c->block_size,
1516                                                              c->block_size, 0, NULL);
1517                        if (!DM_BUFIO_CACHE(c)) {
1518                                r = -ENOMEM;
1519                                mutex_unlock(&dm_bufio_clients_lock);
1520                                goto bad_cache;
1521                        }
1522                }
1523        }
1524        mutex_unlock(&dm_bufio_clients_lock);
1525
1526        while (c->need_reserved_buffers) {
1527                struct dm_buffer *b = alloc_buffer(c, GFP_KERNEL);
1528
1529                if (!b) {
1530                        r = -ENOMEM;
1531                        goto bad_buffer;
1532                }
1533                __free_buffer_wake(b);
1534        }
1535
1536        mutex_lock(&dm_bufio_clients_lock);
1537        dm_bufio_client_count++;
1538        list_add(&c->client_list, &dm_bufio_all_clients);
1539        __cache_size_refresh();
1540        mutex_unlock(&dm_bufio_clients_lock);
1541
1542        c->shrinker.shrink = shrink;
1543        c->shrinker.seeks = 1;
1544        c->shrinker.batch = 0;
1545        register_shrinker(&c->shrinker);
1546
1547        return c;
1548
1549bad_buffer:
1550bad_cache:
1551        while (!list_empty(&c->reserved_buffers)) {
1552                struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1553                                                 struct dm_buffer, lru_list);
1554                list_del(&b->lru_list);
1555                free_buffer(b);
1556        }
1557        dm_io_client_destroy(c->dm_io);
1558bad_dm_io:
1559        vfree(c->cache_hash);
1560bad_hash:
1561        kfree(c);
1562bad_client:
1563        return ERR_PTR(r);
1564}
1565EXPORT_SYMBOL_GPL(dm_bufio_client_create);
1566
1567/*
1568 * Free the buffering interface.
1569 * It is required that there are no references on any buffers.
1570 */
1571void dm_bufio_client_destroy(struct dm_bufio_client *c)
1572{
1573        unsigned i;
1574
1575        drop_buffers(c);
1576
1577        unregister_shrinker(&c->shrinker);
1578
1579        mutex_lock(&dm_bufio_clients_lock);
1580
1581        list_del(&c->client_list);
1582        dm_bufio_client_count--;
1583        __cache_size_refresh();
1584
1585        mutex_unlock(&dm_bufio_clients_lock);
1586
1587        for (i = 0; i < 1 << DM_BUFIO_HASH_BITS; i++)
1588                BUG_ON(!hlist_empty(&c->cache_hash[i]));
1589
1590        BUG_ON(c->need_reserved_buffers);
1591
1592        while (!list_empty(&c->reserved_buffers)) {
1593                struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1594                                                 struct dm_buffer, lru_list);
1595                list_del(&b->lru_list);
1596                free_buffer(b);
1597        }
1598
1599        for (i = 0; i < LIST_SIZE; i++)
1600                if (c->n_buffers[i])
1601                        DMERR("leaked buffer count %d: %ld", i, c->n_buffers[i]);
1602
1603        for (i = 0; i < LIST_SIZE; i++)
1604                BUG_ON(c->n_buffers[i]);
1605
1606        dm_io_client_destroy(c->dm_io);
1607        vfree(c->cache_hash);
1608        kfree(c);
1609}
1610EXPORT_SYMBOL_GPL(dm_bufio_client_destroy);
1611
1612static void cleanup_old_buffers(void)
1613{
1614        unsigned long max_age = ACCESS_ONCE(dm_bufio_max_age);
1615        struct dm_bufio_client *c;
1616
1617        if (max_age > ULONG_MAX / HZ)
1618                max_age = ULONG_MAX / HZ;
1619
1620        mutex_lock(&dm_bufio_clients_lock);
1621        list_for_each_entry(c, &dm_bufio_all_clients, client_list) {
1622                if (!dm_bufio_trylock(c))
1623                        continue;
1624
1625                while (!list_empty(&c->lru[LIST_CLEAN])) {
1626                        struct dm_buffer *b;
1627                        b = list_entry(c->lru[LIST_CLEAN].prev,
1628                                       struct dm_buffer, lru_list);
1629                        if (__cleanup_old_buffer(b, 0, max_age * HZ))
1630                                break;
1631                        dm_bufio_cond_resched();
1632                }
1633
1634                dm_bufio_unlock(c);
1635                dm_bufio_cond_resched();
1636        }
1637        mutex_unlock(&dm_bufio_clients_lock);
1638}
1639
1640static struct workqueue_struct *dm_bufio_wq;
1641static struct delayed_work dm_bufio_work;
1642
1643static void work_fn(struct work_struct *w)
1644{
1645        cleanup_old_buffers();
1646
1647        queue_delayed_work(dm_bufio_wq, &dm_bufio_work,
1648                           DM_BUFIO_WORK_TIMER_SECS * HZ);
1649}
1650
1651/*----------------------------------------------------------------
1652 * Module setup
1653 *--------------------------------------------------------------*/
1654
1655/*
1656 * This is called only once for the whole dm_bufio module.
1657 * It initializes memory limit.
1658 */
1659static int __init dm_bufio_init(void)
1660{
1661        __u64 mem;
1662
1663        memset(&dm_bufio_caches, 0, sizeof dm_bufio_caches);
1664        memset(&dm_bufio_cache_names, 0, sizeof dm_bufio_cache_names);
1665
1666        mem = (__u64)((totalram_pages - totalhigh_pages) *
1667                      DM_BUFIO_MEMORY_PERCENT / 100) << PAGE_SHIFT;
1668
1669        if (mem > ULONG_MAX)
1670                mem = ULONG_MAX;
1671
1672#ifdef CONFIG_MMU
1673        /*
1674         * Get the size of vmalloc space the same way as VMALLOC_TOTAL
1675         * in fs/proc/internal.h
1676         */
1677        if (mem > (VMALLOC_END - VMALLOC_START) * DM_BUFIO_VMALLOC_PERCENT / 100)
1678                mem = (VMALLOC_END - VMALLOC_START) * DM_BUFIO_VMALLOC_PERCENT / 100;
1679#endif
1680
1681        dm_bufio_default_cache_size = mem;
1682
1683        mutex_lock(&dm_bufio_clients_lock);
1684        __cache_size_refresh();
1685        mutex_unlock(&dm_bufio_clients_lock);
1686
1687        dm_bufio_wq = create_singlethread_workqueue("dm_bufio_cache");
1688        if (!dm_bufio_wq)
1689                return -ENOMEM;
1690
1691        INIT_DELAYED_WORK(&dm_bufio_work, work_fn);
1692        queue_delayed_work(dm_bufio_wq, &dm_bufio_work,
1693                           DM_BUFIO_WORK_TIMER_SECS * HZ);
1694
1695        return 0;
1696}
1697
1698/*
1699 * This is called once when unloading the dm_bufio module.
1700 */
1701static void __exit dm_bufio_exit(void)
1702{
1703        int bug = 0;
1704        int i;
1705
1706        cancel_delayed_work_sync(&dm_bufio_work);
1707        destroy_workqueue(dm_bufio_wq);
1708
1709        for (i = 0; i < ARRAY_SIZE(dm_bufio_caches); i++) {
1710                struct kmem_cache *kc = dm_bufio_caches[i];
1711
1712                if (kc)
1713                        kmem_cache_destroy(kc);
1714        }
1715
1716        for (i = 0; i < ARRAY_SIZE(dm_bufio_cache_names); i++)
1717                kfree(dm_bufio_cache_names[i]);
1718
1719        if (dm_bufio_client_count) {
1720                DMCRIT("%s: dm_bufio_client_count leaked: %d",
1721                        __func__, dm_bufio_client_count);
1722                bug = 1;
1723        }
1724
1725        if (dm_bufio_current_allocated) {
1726                DMCRIT("%s: dm_bufio_current_allocated leaked: %lu",
1727                        __func__, dm_bufio_current_allocated);
1728                bug = 1;
1729        }
1730
1731        if (dm_bufio_allocated_get_free_pages) {
1732                DMCRIT("%s: dm_bufio_allocated_get_free_pages leaked: %lu",
1733                       __func__, dm_bufio_allocated_get_free_pages);
1734                bug = 1;
1735        }
1736
1737        if (dm_bufio_allocated_vmalloc) {
1738                DMCRIT("%s: dm_bufio_vmalloc leaked: %lu",
1739                       __func__, dm_bufio_allocated_vmalloc);
1740                bug = 1;
1741        }
1742
1743        if (bug)
1744                BUG();
1745}
1746
1747module_init(dm_bufio_init)
1748module_exit(dm_bufio_exit)
1749
1750module_param_named(max_cache_size_bytes, dm_bufio_cache_size, ulong, S_IRUGO | S_IWUSR);
1751MODULE_PARM_DESC(max_cache_size_bytes, "Size of metadata cache");
1752
1753module_param_named(max_age_seconds, dm_bufio_max_age, uint, S_IRUGO | S_IWUSR);
1754MODULE_PARM_DESC(max_age_seconds, "Max age of a buffer in seconds");
1755
1756module_param_named(peak_allocated_bytes, dm_bufio_peak_allocated, ulong, S_IRUGO | S_IWUSR);
1757MODULE_PARM_DESC(peak_allocated_bytes, "Tracks the maximum allocated memory");
1758
1759module_param_named(allocated_kmem_cache_bytes, dm_bufio_allocated_kmem_cache, ulong, S_IRUGO);
1760MODULE_PARM_DESC(allocated_kmem_cache_bytes, "Memory allocated with kmem_cache_alloc");
1761
1762module_param_named(allocated_get_free_pages_bytes, dm_bufio_allocated_get_free_pages, ulong, S_IRUGO);
1763MODULE_PARM_DESC(allocated_get_free_pages_bytes, "Memory allocated with get_free_pages");
1764
1765module_param_named(allocated_vmalloc_bytes, dm_bufio_allocated_vmalloc, ulong, S_IRUGO);
1766MODULE_PARM_DESC(allocated_vmalloc_bytes, "Memory allocated with vmalloc");
1767
1768module_param_named(current_allocated_bytes, dm_bufio_current_allocated, ulong, S_IRUGO);
1769MODULE_PARM_DESC(current_allocated_bytes, "Memory currently used by the cache");
1770
1771MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
1772MODULE_DESCRIPTION(DM_NAME " buffered I/O library");
1773MODULE_LICENSE("GPL");
1774
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.