drivers/md/dm-bufio.c

   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/jiffies.h>
  15 #include <linux/vmalloc.h>
  16 #include <linux/shrinker.h>
  17 #include <linux/module.h>
  18 #include <linux/rbtree.h>
  19 #include <linux/stacktrace.h>
  20
  21 #define DM_MSG_PREFIX "bufio"
  22
  23 /*
  24  * Memory management policy:
  25  *      Limit the number of buffers to DM_BUFIO_MEMORY_PERCENT of main memory
  26  *      or DM_BUFIO_VMALLOC_PERCENT of vmalloc memory (whichever is lower).
  27  *      Always allocate at least DM_BUFIO_MIN_BUFFERS buffers.
  28  *      Start background writeback when there are DM_BUFIO_WRITEBACK_PERCENT
  29  *      dirty buffers.
  30  */
  31 #define DM_BUFIO_MIN_BUFFERS            8
  32
  33 #define DM_BUFIO_MEMORY_PERCENT         2
  34 #define DM_BUFIO_VMALLOC_PERCENT        25
  35 #define DM_BUFIO_WRITEBACK_PERCENT      75
  36
  37 /*
  38  * Check buffer ages in this interval (seconds)
  39  */
  40 #define DM_BUFIO_WORK_TIMER_SECS        30
  41
  42 /*
  43  * Free buffers when they are older than this (seconds)
  44  */
  45 #define DM_BUFIO_DEFAULT_AGE_SECS       300
  46
  47 /*
  48  * The nr of bytes of cached data to keep around.
  49  */
  50 #define DM_BUFIO_DEFAULT_RETAIN_BYTES   (256 * 1024)
  51
  52 /*
  53  * The number of bvec entries that are embedded directly in the buffer.
  54  * If the chunk size is larger, dm-io is used to do the io.
  55  */
  56 #define DM_BUFIO_INLINE_VECS            16
  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  */
  87 struct 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         unsigned minimum_buffers;
 108
 109         struct rb_root buffer_tree;
 110         wait_queue_head_t free_buffer_wait;
 111
 112         int async_write_error;
 113
 114         struct list_head client_list;
 115         struct shrinker shrinker;
 116 };
 117
 118 /*
 119  * Buffer state bits.
 120  */
 121 #define B_READING       0
 122 #define B_WRITING       1
 123 #define B_DIRTY         2
 124
 125 /*
 126  * Describes how the block was allocated:
 127  * kmem_cache_alloc(), __get_free_pages() or vmalloc().
 128  * See the comment at alloc_buffer_data.
 129  */
 130 enum data_mode {
 131         DATA_MODE_SLAB = 0,
 132         DATA_MODE_GET_FREE_PAGES = 1,
 133         DATA_MODE_VMALLOC = 2,
 134         DATA_MODE_LIMIT = 3
 135 };
 136
 137 struct dm_buffer {
 138         struct rb_node node;
 139         struct list_head lru_list;
 140         sector_t block;
 141         void *data;
 142         enum data_mode data_mode;
 143         unsigned char list_mode;                /* LIST_* */
 144         unsigned hold_count;
 145         int read_error;
 146         int write_error;
 147         unsigned long state;
 148         unsigned long last_accessed;
 149         struct dm_bufio_client *c;
 150         struct list_head write_list;
 151         struct bio bio;
 152         struct bio_vec bio_vec[DM_BUFIO_INLINE_VECS];
 153 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
 154 #define MAX_STACK 10
 155         struct stack_trace stack_trace;
 156         unsigned long stack_entries[MAX_STACK];
 157 #endif
 158 };
 159
 160 /*----------------------------------------------------------------*/
 161
 162 static struct kmem_cache *dm_bufio_caches[PAGE_SHIFT - SECTOR_SHIFT];
 163 static char *dm_bufio_cache_names[PAGE_SHIFT - SECTOR_SHIFT];
 164
 165 static inline int dm_bufio_cache_index(struct dm_bufio_client *c)
 166 {
 167         unsigned ret = c->blocks_per_page_bits - 1;
 168
 169         BUG_ON(ret >= ARRAY_SIZE(dm_bufio_caches));
 170
 171         return ret;
 172 }
 173
 174 #define DM_BUFIO_CACHE(c)       (dm_bufio_caches[dm_bufio_cache_index(c)])
 175 #define DM_BUFIO_CACHE_NAME(c)  (dm_bufio_cache_names[dm_bufio_cache_index(c)])
 176
 177 #define dm_bufio_in_request()   (!!current->bio_list)
 178
 179 static void dm_bufio_lock(struct dm_bufio_client *c)
 180 {
 181         mutex_lock_nested(&c->lock, dm_bufio_in_request());
 182 }
 183
 184 static int dm_bufio_trylock(struct dm_bufio_client *c)
 185 {
 186         return mutex_trylock(&c->lock);
 187 }
 188
 189 static void dm_bufio_unlock(struct dm_bufio_client *c)
 190 {
 191         mutex_unlock(&c->lock);
 192 }
 193
 194 /*----------------------------------------------------------------*/
 195
 196 /*
 197  * Default cache size: available memory divided by the ratio.
 198  */
 199 static unsigned long dm_bufio_default_cache_size;
 200
 201 /*
 202  * Total cache size set by the user.
 203  */
 204 static unsigned long dm_bufio_cache_size;
 205
 206 /*
 207  * A copy of dm_bufio_cache_size because dm_bufio_cache_size can change
 208  * at any time.  If it disagrees, the user has changed cache size.
 209  */
 210 static unsigned long dm_bufio_cache_size_latch;
 211
 212 static DEFINE_SPINLOCK(param_spinlock);
 213
 214 /*
 215  * Buffers are freed after this timeout
 216  */
 217 static unsigned dm_bufio_max_age = DM_BUFIO_DEFAULT_AGE_SECS;
 218 static unsigned dm_bufio_retain_bytes = DM_BUFIO_DEFAULT_RETAIN_BYTES;
 219
 220 static unsigned long dm_bufio_peak_allocated;
 221 static unsigned long dm_bufio_allocated_kmem_cache;
 222 static unsigned long dm_bufio_allocated_get_free_pages;
 223 static unsigned long dm_bufio_allocated_vmalloc;
 224 static unsigned long dm_bufio_current_allocated;
 225
 226 /*----------------------------------------------------------------*/
 227
 228 /*
 229  * Per-client cache: dm_bufio_cache_size / dm_bufio_client_count
 230  */
 231 static unsigned long dm_bufio_cache_size_per_client;
 232
 233 /*
 234  * The current number of clients.
 235  */
 236 static int dm_bufio_client_count;
 237
 238 /*
 239  * The list of all clients.
 240  */
 241 static LIST_HEAD(dm_bufio_all_clients);
 242
 243 /*
 244  * This mutex protects dm_bufio_cache_size_latch,
 245  * dm_bufio_cache_size_per_client and dm_bufio_client_count
 246  */
 247 static DEFINE_MUTEX(dm_bufio_clients_lock);
 248
 249 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
 250 static void buffer_record_stack(struct dm_buffer *b)
 251 {
 252         b->stack_trace.nr_entries = 0;
 253         b->stack_trace.max_entries = MAX_STACK;
 254         b->stack_trace.entries = b->stack_entries;
 255         b->stack_trace.skip = 2;
 256         save_stack_trace(&b->stack_trace);
 257 }
 258 #endif
 259
 260 /*----------------------------------------------------------------
 261  * A red/black tree acts as an index for all the buffers.
 262  *--------------------------------------------------------------*/
 263 static struct dm_buffer *__find(struct dm_bufio_client *c, sector_t block)
 264 {
 265         struct rb_node *n = c->buffer_tree.rb_node;
 266         struct dm_buffer *b;
 267
 268         while (n) {
 269                 b = container_of(n, struct dm_buffer, node);
 270
 271                 if (b->block == block)
 272                         return b;
 273
 274                 n = (b->block < block) ? n->rb_left : n->rb_right;
 275         }
 276
 277         return NULL;
 278 }
 279
 280 static void __insert(struct dm_bufio_client *c, struct dm_buffer *b)
 281 {
 282         struct rb_node **new = &c->buffer_tree.rb_node, *parent = NULL;
 283         struct dm_buffer *found;
 284
 285         while (*new) {
 286                 found = container_of(*new, struct dm_buffer, node);
 287
 288                 if (found->block == b->block) {
 289                         BUG_ON(found != b);
 290                         return;
 291                 }
 292
 293                 parent = *new;
 294                 new = (found->block < b->block) ?
 295                         &((*new)->rb_left) : &((*new)->rb_right);
 296         }
 297
 298         rb_link_node(&b->node, parent, new);
 299         rb_insert_color(&b->node, &c->buffer_tree);
 300 }
 301
 302 static void __remove(struct dm_bufio_client *c, struct dm_buffer *b)
 303 {
 304         rb_erase(&b->node, &c->buffer_tree);
 305 }
 306
 307 /*----------------------------------------------------------------*/
 308
 309 static void adjust_total_allocated(enum data_mode data_mode, long diff)
 310 {
 311         static unsigned long * const class_ptr[DATA_MODE_LIMIT] = {
 312                 &dm_bufio_allocated_kmem_cache,
 313                 &dm_bufio_allocated_get_free_pages,
 314                 &dm_bufio_allocated_vmalloc,
 315         };
 316
 317         spin_lock(&param_spinlock);
 318
 319         *class_ptr[data_mode] += diff;
 320
 321         dm_bufio_current_allocated += diff;
 322
 323         if (dm_bufio_current_allocated > dm_bufio_peak_allocated)
 324                 dm_bufio_peak_allocated = dm_bufio_current_allocated;
 325
 326         spin_unlock(&param_spinlock);
 327 }
 328
 329 /*
 330  * Change the number of clients and recalculate per-client limit.
 331  */
 332 static void __cache_size_refresh(void)
 333 {
 334         BUG_ON(!mutex_is_locked(&dm_bufio_clients_lock));
 335         BUG_ON(dm_bufio_client_count < 0);
 336
 337         dm_bufio_cache_size_latch = ACCESS_ONCE(dm_bufio_cache_size);
 338
 339         /*
 340          * Use default if set to 0 and report the actual cache size used.
 341          */
 342         if (!dm_bufio_cache_size_latch) {
 343                 (void)cmpxchg(&dm_bufio_cache_size, 0,
 344                               dm_bufio_default_cache_size);
 345                 dm_bufio_cache_size_latch = dm_bufio_default_cache_size;
 346         }
 347
 348         dm_bufio_cache_size_per_client = dm_bufio_cache_size_latch /
 349                                          (dm_bufio_client_count ? : 1);
 350 }
 351
 352 /*
 353  * Allocating buffer data.
 354  *
 355  * Small buffers are allocated with kmem_cache, to use space optimally.
 356  *
 357  * For large buffers, we choose between get_free_pages and vmalloc.
 358  * Each has advantages and disadvantages.
 359  *
 360  * __get_free_pages can randomly fail if the memory is fragmented.
 361  * __vmalloc won't randomly fail, but vmalloc space is limited (it may be
 362  * as low as 128M) so using it for caching is not appropriate.
 363  *
 364  * If the allocation may fail we use __get_free_pages. Memory fragmentation
 365  * won't have a fatal effect here, but it just causes flushes of some other
 366  * buffers and more I/O will be performed. Don't use __get_free_pages if it
 367  * always fails (i.e. order >= MAX_ORDER).
 368  *
 369  * If the allocation shouldn't fail we use __vmalloc. This is only for the
 370  * initial reserve allocation, so there's no risk of wasting all vmalloc
 371  * space.
 372  */
 373 static void *alloc_buffer_data(struct dm_bufio_client *c, gfp_t gfp_mask,
 374                                enum data_mode *data_mode)
 375 {
 376         unsigned noio_flag;
 377         void *ptr;
 378
 379         if (c->block_size <= DM_BUFIO_BLOCK_SIZE_SLAB_LIMIT) {
 380                 *data_mode = DATA_MODE_SLAB;
 381                 return kmem_cache_alloc(DM_BUFIO_CACHE(c), gfp_mask);
 382         }
 383
 384         if (c->block_size <= DM_BUFIO_BLOCK_SIZE_GFP_LIMIT &&
 385             gfp_mask & __GFP_NORETRY) {
 386                 *data_mode = DATA_MODE_GET_FREE_PAGES;
 387                 return (void *)__get_free_pages(gfp_mask,
 388                                                 c->pages_per_block_bits);
 389         }
 390
 391         *data_mode = DATA_MODE_VMALLOC;
 392
 393         /*
 394          * __vmalloc allocates the data pages and auxiliary structures with
 395          * gfp_flags that were specified, but pagetables are always allocated
 396          * with GFP_KERNEL, no matter what was specified as gfp_mask.
 397          *
 398          * Consequently, we must set per-process flag PF_MEMALLOC_NOIO so that
 399          * all allocations done by this process (including pagetables) are done
 400          * as if GFP_NOIO was specified.
 401          */
 402
 403         if (gfp_mask & __GFP_NORETRY)
 404                 noio_flag = memalloc_noio_save();
 405
 406         ptr = __vmalloc(c->block_size, gfp_mask | __GFP_HIGHMEM, PAGE_KERNEL);
 407
 408         if (gfp_mask & __GFP_NORETRY)
 409                 memalloc_noio_restore(noio_flag);
 410
 411         return ptr;
 412 }
 413
 414 /*
 415  * Free buffer's data.
 416  */
 417 static void free_buffer_data(struct dm_bufio_client *c,
 418                              void *data, enum data_mode data_mode)
 419 {
 420         switch (data_mode) {
 421         case DATA_MODE_SLAB:
 422                 kmem_cache_free(DM_BUFIO_CACHE(c), data);
 423                 break;
 424
 425         case DATA_MODE_GET_FREE_PAGES:
 426                 free_pages((unsigned long)data, c->pages_per_block_bits);
 427                 break;
 428
 429         case DATA_MODE_VMALLOC:
 430                 vfree(data);
 431                 break;
 432
 433         default:
 434                 DMCRIT("dm_bufio_free_buffer_data: bad data mode: %d",
 435                        data_mode);
 436                 BUG();
 437         }
 438 }
 439
 440 /*
 441  * Allocate buffer and its data.
 442  */
 443 static struct dm_buffer *alloc_buffer(struct dm_bufio_client *c, gfp_t gfp_mask)
 444 {
 445         struct dm_buffer *b = kmalloc(sizeof(struct dm_buffer) + c->aux_size,
 446                                       gfp_mask);
 447
 448         if (!b)
 449                 return NULL;
 450
 451         b->c = c;
 452
 453         b->data = alloc_buffer_data(c, gfp_mask, &b->data_mode);
 454         if (!b->data) {
 455                 kfree(b);
 456                 return NULL;
 457         }
 458
 459         adjust_total_allocated(b->data_mode, (long)c->block_size);
 460
 461 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
 462         memset(&b->stack_trace, 0, sizeof(b->stack_trace));
 463 #endif
 464         return b;
 465 }
 466
 467 /*
 468  * Free buffer and its data.
 469  */
 470 static void free_buffer(struct dm_buffer *b)
 471 {
 472         struct dm_bufio_client *c = b->c;
 473
 474         adjust_total_allocated(b->data_mode, -(long)c->block_size);
 475
 476         free_buffer_data(c, b->data, b->data_mode);
 477         kfree(b);
 478 }
 479
 480 /*
 481  * Link buffer to the hash list and clean or dirty queue.
 482  */
 483 static void __link_buffer(struct dm_buffer *b, sector_t block, int dirty)
 484 {
 485         struct dm_bufio_client *c = b->c;
 486
 487         c->n_buffers[dirty]++;
 488         b->block = block;
 489         b->list_mode = dirty;
 490         list_add(&b->lru_list, &c->lru[dirty]);
 491         __insert(b->c, b);
 492         b->last_accessed = jiffies;
 493 }
 494
 495 /*
 496  * Unlink buffer from the hash list and dirty or clean queue.
 497  */
 498 static void __unlink_buffer(struct dm_buffer *b)
 499 {
 500         struct dm_bufio_client *c = b->c;
 501
 502         BUG_ON(!c->n_buffers[b->list_mode]);
 503
 504         c->n_buffers[b->list_mode]--;
 505         __remove(b->c, b);
 506         list_del(&b->lru_list);
 507 }
 508
 509 /*
 510  * Place the buffer to the head of dirty or clean LRU queue.
 511  */
 512 static void __relink_lru(struct dm_buffer *b, int dirty)
 513 {
 514         struct dm_bufio_client *c = b->c;
 515
 516         BUG_ON(!c->n_buffers[b->list_mode]);
 517
 518         c->n_buffers[b->list_mode]--;
 519         c->n_buffers[dirty]++;
 520         b->list_mode = dirty;
 521         list_move(&b->lru_list, &c->lru[dirty]);
 522         b->last_accessed = jiffies;
 523 }
 524
 525 /*----------------------------------------------------------------
 526  * Submit I/O on the buffer.
 527  *
 528  * Bio interface is faster but it has some problems:
 529  *      the vector list is limited (increasing this limit increases
 530  *      memory-consumption per buffer, so it is not viable);
 531  *
 532  *      the memory must be direct-mapped, not vmalloced;
 533  *
 534  *      the I/O driver can reject requests spuriously if it thinks that
 535  *      the requests are too big for the device or if they cross a
 536  *      controller-defined memory boundary.
 537  *
 538  * If the buffer is small enough (up to DM_BUFIO_INLINE_VECS pages) and
 539  * it is not vmalloced, try using the bio interface.
 540  *
 541  * If the buffer is big, if it is vmalloced or if the underlying device
 542  * rejects the bio because it is too large, use dm-io layer to do the I/O.
 543  * The dm-io layer splits the I/O into multiple requests, avoiding the above
 544  * shortcomings.
 545  *--------------------------------------------------------------*/
 546
 547 /*
 548  * dm-io completion routine. It just calls b->bio.bi_end_io, pretending
 549  * that the request was handled directly with bio interface.
 550  */
 551 static void dmio_complete(unsigned long error, void *context)
 552 {
 553         struct dm_buffer *b = context;
 554
 555         b->bio.bi_error = error ? -EIO : 0;
 556         b->bio.bi_end_io(&b->bio);
 557 }
 558
 559 static void use_dmio(struct dm_buffer *b, int rw, sector_t block,
 560                      bio_end_io_t *end_io)
 561 {
 562         int r;
 563         struct dm_io_request io_req = {
 564                 .bi_op = rw,
 565                 .bi_op_flags = 0,
 566                 .notify.fn = dmio_complete,
 567                 .notify.context = b,
 568                 .client = b->c->dm_io,
 569         };
 570         struct dm_io_region region = {
 571                 .bdev = b->c->bdev,
 572                 .sector = block << b->c->sectors_per_block_bits,
 573                 .count = b->c->block_size >> SECTOR_SHIFT,
 574         };
 575
 576         if (b->data_mode != DATA_MODE_VMALLOC) {
 577                 io_req.mem.type = DM_IO_KMEM;
 578                 io_req.mem.ptr.addr = b->data;
 579         } else {
 580                 io_req.mem.type = DM_IO_VMA;
 581                 io_req.mem.ptr.vma = b->data;
 582         }
 583
 584         b->bio.bi_end_io = end_io;
 585
 586         r = dm_io(&io_req, 1, &region, NULL);
 587         if (r) {
 588                 b->bio.bi_error = r;
 589                 end_io(&b->bio);
 590         }
 591 }
 592
 593 static void inline_endio(struct bio *bio)
 594 {
 595         bio_end_io_t *end_fn = bio->bi_private;
 596         int error = bio->bi_error;
 597
 598         /*
 599          * Reset the bio to free any attached resources
 600          * (e.g. bio integrity profiles).
 601          */
 602         bio_reset(bio);
 603
 604         bio->bi_error = error;
 605         end_fn(bio);
 606 }
 607
 608 static void use_inline_bio(struct dm_buffer *b, int rw, sector_t block,
 609                            bio_end_io_t *end_io)
 610 {
 611         char *ptr;
 612         int len;
 613
 614         bio_init(&b->bio);
 615         b->bio.bi_io_vec = b->bio_vec;
 616         b->bio.bi_max_vecs = DM_BUFIO_INLINE_VECS;
 617         b->bio.bi_iter.bi_sector = block << b->c->sectors_per_block_bits;
 618         b->bio.bi_bdev = b->c->bdev;
 619         b->bio.bi_end_io = inline_endio;
 620         /*
 621          * Use of .bi_private isn't a problem here because
 622          * the dm_buffer's inline bio is local to bufio.
 623          */
 624         b->bio.bi_private = end_io;
 625         bio_set_op_attrs(&b->bio, rw, 0);
 626
 627         /*
 628          * We assume that if len >= PAGE_SIZE ptr is page-aligned.
 629          * If len < PAGE_SIZE the buffer doesn't cross page boundary.
 630          */
 631         ptr = b->data;
 632         len = b->c->block_size;
 633
 634         if (len >= PAGE_SIZE)
 635                 BUG_ON((unsigned long)ptr & (PAGE_SIZE - 1));
 636         else
 637                 BUG_ON((unsigned long)ptr & (len - 1));
 638
 639         do {
 640                 if (!bio_add_page(&b->bio, virt_to_page(ptr),
 641                                   len < PAGE_SIZE ? len : PAGE_SIZE,
 642                                   offset_in_page(ptr))) {
 643                         BUG_ON(b->c->block_size <= PAGE_SIZE);
 644                         use_dmio(b, rw, block, end_io);
 645                         return;
 646                 }
 647
 648                 len -= PAGE_SIZE;
 649                 ptr += PAGE_SIZE;
 650         } while (len > 0);
 651
 652         submit_bio(&b->bio);
 653 }
 654
 655 static void submit_io(struct dm_buffer *b, int rw, sector_t block,
 656                       bio_end_io_t *end_io)
 657 {
 658         if (rw == WRITE && b->c->write_callback)
 659                 b->c->write_callback(b);
 660
 661         if (b->c->block_size <= DM_BUFIO_INLINE_VECS * PAGE_SIZE &&
 662             b->data_mode != DATA_MODE_VMALLOC)
 663                 use_inline_bio(b, rw, block, end_io);
 664         else
 665                 use_dmio(b, rw, block, end_io);
 666 }
 667
 668 /*----------------------------------------------------------------
 669  * Writing dirty buffers
 670  *--------------------------------------------------------------*/
 671
 672 /*
 673  * The endio routine for write.
 674  *
 675  * Set the error, clear B_WRITING bit and wake anyone who was waiting on
 676  * it.
 677  */
 678 static void write_endio(struct bio *bio)
 679 {
 680         struct dm_buffer *b = container_of(bio, struct dm_buffer, bio);
 681
 682         b->write_error = bio->bi_error;
 683         if (unlikely(bio->bi_error)) {
 684                 struct dm_bufio_client *c = b->c;
 685                 int error = bio->bi_error;
 686                 (void)cmpxchg(&c->async_write_error, 0, error);
 687         }
 688
 689         BUG_ON(!test_bit(B_WRITING, &b->state));
 690
 691         smp_mb__before_atomic();
 692         clear_bit(B_WRITING, &b->state);
 693         smp_mb__after_atomic();
 694
 695         wake_up_bit(&b->state, B_WRITING);
 696 }
 697
 698 /*
 699  * Initiate a write on a dirty buffer, but don't wait for it.
 700  *
 701  * - If the buffer is not dirty, exit.
 702  * - If there some previous write going on, wait for it to finish (we can't
 703  *   have two writes on the same buffer simultaneously).
 704  * - Submit our write and don't wait on it. We set B_WRITING indicating
 705  *   that there is a write in progress.
 706  */
 707 static void __write_dirty_buffer(struct dm_buffer *b,
 708                                  struct list_head *write_list)
 709 {
 710         if (!test_bit(B_DIRTY, &b->state))
 711                 return;
 712
 713         clear_bit(B_DIRTY, &b->state);
 714         wait_on_bit_lock_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
 715
 716         if (!write_list)
 717                 submit_io(b, WRITE, b->block, write_endio);
 718         else
 719                 list_add_tail(&b->write_list, write_list);
 720 }
 721
 722 static void __flush_write_list(struct list_head *write_list)
 723 {
 724         struct blk_plug plug;
 725         blk_start_plug(&plug);
 726         while (!list_empty(write_list)) {
 727                 struct dm_buffer *b =
 728                         list_entry(write_list->next, struct dm_buffer, write_list);
 729                 list_del(&b->write_list);
 730                 submit_io(b, WRITE, b->block, write_endio);
 731                 cond_resched();
 732         }
 733         blk_finish_plug(&plug);
 734 }
 735
 736 /*
 737  * Wait until any activity on the buffer finishes.  Possibly write the
 738  * buffer if it is dirty.  When this function finishes, there is no I/O
 739  * running on the buffer and the buffer is not dirty.
 740  */
 741 static void __make_buffer_clean(struct dm_buffer *b)
 742 {
 743         BUG_ON(b->hold_count);
 744
 745         if (!b->state)  /* fast case */
 746                 return;
 747
 748         wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE);
 749         __write_dirty_buffer(b, NULL);
 750         wait_on_bit_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
 751 }
 752
 753 /*
 754  * Find some buffer that is not held by anybody, clean it, unlink it and
 755  * return it.
 756  */
 757 static struct dm_buffer *__get_unclaimed_buffer(struct dm_bufio_client *c)
 758 {
 759         struct dm_buffer *b;
 760
 761         list_for_each_entry_reverse(b, &c->lru[LIST_CLEAN], lru_list) {
 762                 BUG_ON(test_bit(B_WRITING, &b->state));
 763                 BUG_ON(test_bit(B_DIRTY, &b->state));
 764
 765                 if (!b->hold_count) {
 766                         __make_buffer_clean(b);
 767                         __unlink_buffer(b);
 768                         return b;
 769                 }
 770                 cond_resched();
 771         }
 772
 773         list_for_each_entry_reverse(b, &c->lru[LIST_DIRTY], lru_list) {
 774                 BUG_ON(test_bit(B_READING, &b->state));
 775
 776                 if (!b->hold_count) {
 777                         __make_buffer_clean(b);
 778                         __unlink_buffer(b);
 779                         return b;
 780                 }
 781                 cond_resched();
 782         }
 783
 784         return NULL;
 785 }
 786
 787 /*
 788  * Wait until some other threads free some buffer or release hold count on
 789  * some buffer.
 790  *
 791  * This function is entered with c->lock held, drops it and regains it
 792  * before exiting.
 793  */
 794 static void __wait_for_free_buffer(struct dm_bufio_client *c)
 795 {
 796         DECLARE_WAITQUEUE(wait, current);
 797
 798         add_wait_queue(&c->free_buffer_wait, &wait);
 799         set_task_state(current, TASK_UNINTERRUPTIBLE);
 800         dm_bufio_unlock(c);
 801
 802         io_schedule();
 803
 804         remove_wait_queue(&c->free_buffer_wait, &wait);
 805
 806         dm_bufio_lock(c);
 807 }
 808
 809 enum new_flag {
 810         NF_FRESH = 0,
 811         NF_READ = 1,
 812         NF_GET = 2,
 813         NF_PREFETCH = 3
 814 };
 815
 816 /*
 817  * Allocate a new buffer. If the allocation is not possible, wait until
 818  * some other thread frees a buffer.
 819  *
 820  * May drop the lock and regain it.
 821  */
 822 static struct dm_buffer *__alloc_buffer_wait_no_callback(struct dm_bufio_client *c, enum new_flag nf)
 823 {
 824         struct dm_buffer *b;
 825
 826         /*
 827          * dm-bufio is resistant to allocation failures (it just keeps
 828          * one buffer reserved in cases all the allocations fail).
 829          * So set flags to not try too hard:
 830          *      GFP_NOIO: don't recurse into the I/O layer
 831          *      __GFP_NORETRY: don't retry and rather return failure
 832          *      __GFP_NOMEMALLOC: don't use emergency reserves
 833          *      __GFP_NOWARN: don't print a warning in case of failure
 834          *
 835          * For debugging, if we set the cache size to 1, no new buffers will
 836          * be allocated.
 837          */
 838         while (1) {
 839                 if (dm_bufio_cache_size_latch != 1) {
 840                         b = alloc_buffer(c, GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
 841                         if (b)
 842                                 return b;
 843                 }
 844
 845                 if (nf == NF_PREFETCH)
 846                         return NULL;
 847
 848                 if (!list_empty(&c->reserved_buffers)) {
 849                         b = list_entry(c->reserved_buffers.next,
 850                                        struct dm_buffer, lru_list);
 851                         list_del(&b->lru_list);
 852                         c->need_reserved_buffers++;
 853
 854                         return b;
 855                 }
 856
 857                 b = __get_unclaimed_buffer(c);
 858                 if (b)
 859                         return b;
 860
 861                 __wait_for_free_buffer(c);
 862         }
 863 }
 864
 865 static struct dm_buffer *__alloc_buffer_wait(struct dm_bufio_client *c, enum new_flag nf)
 866 {
 867         struct dm_buffer *b = __alloc_buffer_wait_no_callback(c, nf);
 868
 869         if (!b)
 870                 return NULL;
 871
 872         if (c->alloc_callback)
 873                 c->alloc_callback(b);
 874
 875         return b;
 876 }
 877
 878 /*
 879  * Free a buffer and wake other threads waiting for free buffers.
 880  */
 881 static void __free_buffer_wake(struct dm_buffer *b)
 882 {
 883         struct dm_bufio_client *c = b->c;
 884
 885         if (!c->need_reserved_buffers)
 886                 free_buffer(b);
 887         else {
 888                 list_add(&b->lru_list, &c->reserved_buffers);
 889                 c->need_reserved_buffers--;
 890         }
 891
 892         wake_up(&c->free_buffer_wait);
 893 }
 894
 895 static void __write_dirty_buffers_async(struct dm_bufio_client *c, int no_wait,
 896                                         struct list_head *write_list)
 897 {
 898         struct dm_buffer *b, *tmp;
 899
 900         list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
 901                 BUG_ON(test_bit(B_READING, &b->state));
 902
 903                 if (!test_bit(B_DIRTY, &b->state) &&
 904                     !test_bit(B_WRITING, &b->state)) {
 905                         __relink_lru(b, LIST_CLEAN);
 906                         continue;
 907                 }
 908
 909                 if (no_wait && test_bit(B_WRITING, &b->state))
 910                         return;
 911
 912                 __write_dirty_buffer(b, write_list);
 913                 cond_resched();
 914         }
 915 }
 916
 917 /*
 918  * Get writeback threshold and buffer limit for a given client.
 919  */
 920 static void __get_memory_limit(struct dm_bufio_client *c,
 921                                unsigned long *threshold_buffers,
 922                                unsigned long *limit_buffers)
 923 {
 924         unsigned long buffers;
 925
 926         if (ACCESS_ONCE(dm_bufio_cache_size) != dm_bufio_cache_size_latch) {
 927                 mutex_lock(&dm_bufio_clients_lock);
 928                 __cache_size_refresh();
 929                 mutex_unlock(&dm_bufio_clients_lock);
 930         }
 931
 932         buffers = dm_bufio_cache_size_per_client >>
 933                   (c->sectors_per_block_bits + SECTOR_SHIFT);
 934
 935         if (buffers < c->minimum_buffers)
 936                 buffers = c->minimum_buffers;
 937
 938         *limit_buffers = buffers;
 939         *threshold_buffers = buffers * DM_BUFIO_WRITEBACK_PERCENT / 100;
 940 }
 941
 942 /*
 943  * Check if we're over watermark.
 944  * If we are over threshold_buffers, start freeing buffers.
 945  * If we're over "limit_buffers", block until we get under the limit.
 946  */
 947 static void __check_watermark(struct dm_bufio_client *c,
 948                               struct list_head *write_list)
 949 {
 950         unsigned long threshold_buffers, limit_buffers;
 951
 952         __get_memory_limit(c, &threshold_buffers, &limit_buffers);
 953
 954         while (c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY] >
 955                limit_buffers) {
 956
 957                 struct dm_buffer *b = __get_unclaimed_buffer(c);
 958
 959                 if (!b)
 960                         return;
 961
 962                 __free_buffer_wake(b);
 963                 cond_resched();
 964         }
 965
 966         if (c->n_buffers[LIST_DIRTY] > threshold_buffers)
 967                 __write_dirty_buffers_async(c, 1, write_list);
 968 }
 969
 970 /*----------------------------------------------------------------
 971  * Getting a buffer
 972  *--------------------------------------------------------------*/
 973
 974 static struct dm_buffer *__bufio_new(struct dm_bufio_client *c, sector_t block,
 975                                      enum new_flag nf, int *need_submit,
 976                                      struct list_head *write_list)
 977 {
 978         struct dm_buffer *b, *new_b = NULL;
 979
 980         *need_submit = 0;
 981
 982         b = __find(c, block);
 983         if (b)
 984                 goto found_buffer;
 985
 986         if (nf == NF_GET)
 987                 return NULL;
 988
 989         new_b = __alloc_buffer_wait(c, nf);
 990         if (!new_b)
 991                 return NULL;
 992
 993         /*
 994          * We've had a period where the mutex was unlocked, so need to
 995          * recheck the hash table.
 996          */
 997         b = __find(c, block);
 998         if (b) {
 999                 __free_buffer_wake(new_b);
1000                 goto found_buffer;
1001         }
1002
1003         __check_watermark(c, write_list);
1004
1005         b = new_b;
1006         b->hold_count = 1;
1007         b->read_error = 0;
1008         b->write_error = 0;
1009         __link_buffer(b, block, LIST_CLEAN);
1010
1011         if (nf == NF_FRESH) {
1012                 b->state = 0;
1013                 return b;
1014         }
1015
1016         b->state = 1 << B_READING;
1017         *need_submit = 1;
1018
1019         return b;
1020
1021 found_buffer:
1022         if (nf == NF_PREFETCH)
1023                 return NULL;
1024         /*
1025          * Note: it is essential that we don't wait for the buffer to be
1026          * read if dm_bufio_get function is used. Both dm_bufio_get and
1027          * dm_bufio_prefetch can be used in the driver request routine.
1028          * If the user called both dm_bufio_prefetch and dm_bufio_get on
1029          * the same buffer, it would deadlock if we waited.
1030          */
1031         if (nf == NF_GET && unlikely(test_bit(B_READING, &b->state)))
1032                 return NULL;
1033
1034         b->hold_count++;
1035         __relink_lru(b, test_bit(B_DIRTY, &b->state) ||
1036                      test_bit(B_WRITING, &b->state));
1037         return b;
1038 }
1039
1040 /*
1041  * The endio routine for reading: set the error, clear the bit and wake up
1042  * anyone waiting on the buffer.
1043  */
1044 static void read_endio(struct bio *bio)
1045 {
1046         struct dm_buffer *b = container_of(bio, struct dm_buffer, bio);
1047
1048         b->read_error = bio->bi_error;
1049
1050         BUG_ON(!test_bit(B_READING, &b->state));
1051
1052         smp_mb__before_atomic();
1053         clear_bit(B_READING, &b->state);
1054         smp_mb__after_atomic();
1055
1056         wake_up_bit(&b->state, B_READING);
1057 }
1058
1059 /*
1060  * A common routine for dm_bufio_new and dm_bufio_read.  Operation of these
1061  * functions is similar except that dm_bufio_new doesn't read the
1062  * buffer from the disk (assuming that the caller overwrites all the data
1063  * and uses dm_bufio_mark_buffer_dirty to write new data back).
1064  */
1065 static void *new_read(struct dm_bufio_client *c, sector_t block,
1066                       enum new_flag nf, struct dm_buffer **bp)
1067 {
1068         int need_submit;
1069         struct dm_buffer *b;
1070
1071         LIST_HEAD(write_list);
1072
1073         dm_bufio_lock(c);
1074         b = __bufio_new(c, block, nf, &need_submit, &write_list);
1075 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1076         if (b && b->hold_count == 1)
1077                 buffer_record_stack(b);
1078 #endif
1079         dm_bufio_unlock(c);
1080
1081         __flush_write_list(&write_list);
1082
1083         if (!b)
1084                 return NULL;
1085
1086         if (need_submit)
1087                 submit_io(b, READ, b->block, read_endio);
1088
1089         wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE);
1090
1091         if (b->read_error) {
1092                 int error = b->read_error;
1093
1094                 dm_bufio_release(b);
1095
1096                 return ERR_PTR(error);
1097         }
1098
1099         *bp = b;
1100
1101         return b->data;
1102 }
1103
1104 void *dm_bufio_get(struct dm_bufio_client *c, sector_t block,
1105                    struct dm_buffer **bp)
1106 {
1107         return new_read(c, block, NF_GET, bp);
1108 }
1109 EXPORT_SYMBOL_GPL(dm_bufio_get);
1110
1111 void *dm_bufio_read(struct dm_bufio_client *c, sector_t block,
1112                     struct dm_buffer **bp)
1113 {
1114         BUG_ON(dm_bufio_in_request());
1115
1116         return new_read(c, block, NF_READ, bp);
1117 }
1118 EXPORT_SYMBOL_GPL(dm_bufio_read);
1119
1120 void *dm_bufio_new(struct dm_bufio_client *c, sector_t block,
1121                    struct dm_buffer **bp)
1122 {
1123         BUG_ON(dm_bufio_in_request());
1124
1125         return new_read(c, block, NF_FRESH, bp);
1126 }
1127 EXPORT_SYMBOL_GPL(dm_bufio_new);
1128
1129 void dm_bufio_prefetch(struct dm_bufio_client *c,
1130                        sector_t block, unsigned n_blocks)
1131 {
1132         struct blk_plug plug;
1133
1134         LIST_HEAD(write_list);
1135
1136         BUG_ON(dm_bufio_in_request());
1137
1138         blk_start_plug(&plug);
1139         dm_bufio_lock(c);
1140
1141         for (; n_blocks--; block++) {
1142                 int need_submit;
1143                 struct dm_buffer *b;
1144                 b = __bufio_new(c, block, NF_PREFETCH, &need_submit,
1145                                 &write_list);
1146                 if (unlikely(!list_empty(&write_list))) {
1147                         dm_bufio_unlock(c);
1148                         blk_finish_plug(&plug);
1149                         __flush_write_list(&write_list);
1150                         blk_start_plug(&plug);
1151                         dm_bufio_lock(c);
1152                 }
1153                 if (unlikely(b != NULL)) {
1154                         dm_bufio_unlock(c);
1155
1156                         if (need_submit)
1157                                 submit_io(b, READ, b->block, read_endio);
1158                         dm_bufio_release(b);
1159
1160                         cond_resched();
1161
1162                         if (!n_blocks)
1163                                 goto flush_plug;
1164                         dm_bufio_lock(c);
1165                 }
1166         }
1167
1168         dm_bufio_unlock(c);
1169
1170 flush_plug:
1171         blk_finish_plug(&plug);
1172 }
1173 EXPORT_SYMBOL_GPL(dm_bufio_prefetch);
1174
1175 void dm_bufio_release(struct dm_buffer *b)
1176 {
1177         struct dm_bufio_client *c = b->c;
1178
1179         dm_bufio_lock(c);
1180
1181         BUG_ON(!b->hold_count);
1182
1183         b->hold_count--;
1184         if (!b->hold_count) {
1185                 wake_up(&c->free_buffer_wait);
1186
1187                 /*
1188                  * If there were errors on the buffer, and the buffer is not
1189                  * to be written, free the buffer. There is no point in caching
1190                  * invalid buffer.
1191                  */
1192                 if ((b->read_error || b->write_error) &&
1193                     !test_bit(B_READING, &b->state) &&
1194                     !test_bit(B_WRITING, &b->state) &&
1195                     !test_bit(B_DIRTY, &b->state)) {
1196                         __unlink_buffer(b);
1197                         __free_buffer_wake(b);
1198                 }
1199         }
1200
1201         dm_bufio_unlock(c);
1202 }
1203 EXPORT_SYMBOL_GPL(dm_bufio_release);
1204
1205 void dm_bufio_mark_buffer_dirty(struct dm_buffer *b)
1206 {
1207         struct dm_bufio_client *c = b->c;
1208
1209         dm_bufio_lock(c);
1210
1211         BUG_ON(test_bit(B_READING, &b->state));
1212
1213         if (!test_and_set_bit(B_DIRTY, &b->state))
1214                 __relink_lru(b, LIST_DIRTY);
1215
1216         dm_bufio_unlock(c);
1217 }
1218 EXPORT_SYMBOL_GPL(dm_bufio_mark_buffer_dirty);
1219
1220 void dm_bufio_write_dirty_buffers_async(struct dm_bufio_client *c)
1221 {
1222         LIST_HEAD(write_list);
1223
1224         BUG_ON(dm_bufio_in_request());
1225
1226         dm_bufio_lock(c);
1227         __write_dirty_buffers_async(c, 0, &write_list);
1228         dm_bufio_unlock(c);
1229         __flush_write_list(&write_list);
1230 }
1231 EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers_async);
1232
1233 /*
1234  * For performance, it is essential that the buffers are written asynchronously
1235  * and simultaneously (so that the block layer can merge the writes) and then
1236  * waited upon.
1237  *
1238  * Finally, we flush hardware disk cache.
1239  */
1240 int dm_bufio_write_dirty_buffers(struct dm_bufio_client *c)
1241 {
1242         int a, f;
1243         unsigned long buffers_processed = 0;
1244         struct dm_buffer *b, *tmp;
1245
1246         LIST_HEAD(write_list);
1247
1248         dm_bufio_lock(c);
1249         __write_dirty_buffers_async(c, 0, &write_list);
1250         dm_bufio_unlock(c);
1251         __flush_write_list(&write_list);
1252         dm_bufio_lock(c);
1253
1254 again:
1255         list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
1256                 int dropped_lock = 0;
1257
1258                 if (buffers_processed < c->n_buffers[LIST_DIRTY])
1259                         buffers_processed++;
1260
1261                 BUG_ON(test_bit(B_READING, &b->state));
1262
1263                 if (test_bit(B_WRITING, &b->state)) {
1264                         if (buffers_processed < c->n_buffers[LIST_DIRTY]) {
1265                                 dropped_lock = 1;
1266                                 b->hold_count++;
1267                                 dm_bufio_unlock(c);
1268                                 wait_on_bit_io(&b->state, B_WRITING,
1269                                                TASK_UNINTERRUPTIBLE);
1270                                 dm_bufio_lock(c);
1271                                 b->hold_count--;
1272                         } else
1273                                 wait_on_bit_io(&b->state, B_WRITING,
1274                                                TASK_UNINTERRUPTIBLE);
1275                 }
1276
1277                 if (!test_bit(B_DIRTY, &b->state) &&
1278                     !test_bit(B_WRITING, &b->state))
1279                         __relink_lru(b, LIST_CLEAN);
1280
1281                 cond_resched();
1282
1283                 /*
1284                  * If we dropped the lock, the list is no longer consistent,
1285                  * so we must restart the search.
1286                  *
1287                  * In the most common case, the buffer just processed is
1288                  * relinked to the clean list, so we won't loop scanning the
1289                  * same buffer again and again.
1290                  *
1291                  * This may livelock if there is another thread simultaneously
1292                  * dirtying buffers, so we count the number of buffers walked
1293                  * and if it exceeds the total number of buffers, it means that
1294                  * someone is doing some writes simultaneously with us.  In
1295                  * this case, stop, dropping the lock.
1296                  */
1297                 if (dropped_lock)
1298                         goto again;
1299         }
1300         wake_up(&c->free_buffer_wait);
1301         dm_bufio_unlock(c);
1302
1303         a = xchg(&c->async_write_error, 0);
1304         f = dm_bufio_issue_flush(c);
1305         if (a)
1306                 return a;
1307
1308         return f;
1309 }
1310 EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers);
1311
1312 /*
1313  * Use dm-io to send and empty barrier flush the device.
1314  */
1315 int dm_bufio_issue_flush(struct dm_bufio_client *c)
1316 {
1317         struct dm_io_request io_req = {
1318                 .bi_op = REQ_OP_WRITE,
1319                 .bi_op_flags = WRITE_FLUSH,
1320                 .mem.type = DM_IO_KMEM,
1321                 .mem.ptr.addr = NULL,
1322                 .client = c->dm_io,
1323         };
1324         struct dm_io_region io_reg = {
1325                 .bdev = c->bdev,
1326                 .sector = 0,
1327                 .count = 0,
1328         };
1329
1330         BUG_ON(dm_bufio_in_request());
1331
1332         return dm_io(&io_req, 1, &io_reg, NULL);
1333 }
1334 EXPORT_SYMBOL_GPL(dm_bufio_issue_flush);
1335
1336 /*
1337  * We first delete any other buffer that may be at that new location.
1338  *
1339  * Then, we write the buffer to the original location if it was dirty.
1340  *
1341  * Then, if we are the only one who is holding the buffer, relink the buffer
1342  * in the hash queue for the new location.
1343  *
1344  * If there was someone else holding the buffer, we write it to the new
1345  * location but not relink it, because that other user needs to have the buffer
1346  * at the same place.
1347  */
1348 void dm_bufio_release_move(struct dm_buffer *b, sector_t new_block)
1349 {
1350         struct dm_bufio_client *c = b->c;
1351         struct dm_buffer *new;
1352
1353         BUG_ON(dm_bufio_in_request());
1354
1355         dm_bufio_lock(c);
1356
1357 retry:
1358         new = __find(c, new_block);
1359         if (new) {
1360                 if (new->hold_count) {
1361                         __wait_for_free_buffer(c);
1362                         goto retry;
1363                 }
1364
1365                 /*
1366                  * FIXME: Is there any point waiting for a write that's going
1367                  * to be overwritten in a bit?
1368                  */
1369                 __make_buffer_clean(new);
1370                 __unlink_buffer(new);
1371                 __free_buffer_wake(new);
1372         }
1373
1374         BUG_ON(!b->hold_count);
1375         BUG_ON(test_bit(B_READING, &b->state));
1376
1377         __write_dirty_buffer(b, NULL);
1378         if (b->hold_count == 1) {
1379                 wait_on_bit_io(&b->state, B_WRITING,
1380                                TASK_UNINTERRUPTIBLE);
1381                 set_bit(B_DIRTY, &b->state);
1382                 __unlink_buffer(b);
1383                 __link_buffer(b, new_block, LIST_DIRTY);
1384         } else {
1385                 sector_t old_block;
1386                 wait_on_bit_lock_io(&b->state, B_WRITING,
1387                                     TASK_UNINTERRUPTIBLE);
1388                 /*
1389                  * Relink buffer to "new_block" so that write_callback
1390                  * sees "new_block" as a block number.
1391                  * After the write, link the buffer back to old_block.
1392                  * All this must be done in bufio lock, so that block number
1393                  * change isn't visible to other threads.
1394                  */
1395                 old_block = b->block;
1396                 __unlink_buffer(b);
1397                 __link_buffer(b, new_block, b->list_mode);
1398                 submit_io(b, WRITE, new_block, write_endio);
1399                 wait_on_bit_io(&b->state, B_WRITING,
1400                                TASK_UNINTERRUPTIBLE);
1401                 __unlink_buffer(b);
1402                 __link_buffer(b, old_block, b->list_mode);
1403         }
1404
1405         dm_bufio_unlock(c);
1406         dm_bufio_release(b);
1407 }
1408 EXPORT_SYMBOL_GPL(dm_bufio_release_move);
1409
1410 /*
1411  * Free the given buffer.
1412  *
1413  * This is just a hint, if the buffer is in use or dirty, this function
1414  * does nothing.
1415  */
1416 void dm_bufio_forget(struct dm_bufio_client *c, sector_t block)
1417 {
1418         struct dm_buffer *b;
1419
1420         dm_bufio_lock(c);
1421
1422         b = __find(c, block);
1423         if (b && likely(!b->hold_count) && likely(!b->state)) {
1424                 __unlink_buffer(b);
1425                 __free_buffer_wake(b);
1426         }
1427
1428         dm_bufio_unlock(c);
1429 }
1430 EXPORT_SYMBOL(dm_bufio_forget);
1431
1432 void dm_bufio_set_minimum_buffers(struct dm_bufio_client *c, unsigned n)
1433 {
1434         c->minimum_buffers = n;
1435 }
1436 EXPORT_SYMBOL(dm_bufio_set_minimum_buffers);
1437
1438 unsigned dm_bufio_get_block_size(struct dm_bufio_client *c)
1439 {
1440         return c->block_size;
1441 }
1442 EXPORT_SYMBOL_GPL(dm_bufio_get_block_size);
1443
1444 sector_t dm_bufio_get_device_size(struct dm_bufio_client *c)
1445 {
1446         return i_size_read(c->bdev->bd_inode) >>
1447                            (SECTOR_SHIFT + c->sectors_per_block_bits);
1448 }
1449 EXPORT_SYMBOL_GPL(dm_bufio_get_device_size);
1450
1451 sector_t dm_bufio_get_block_number(struct dm_buffer *b)
1452 {
1453         return b->block;
1454 }
1455 EXPORT_SYMBOL_GPL(dm_bufio_get_block_number);
1456
1457 void *dm_bufio_get_block_data(struct dm_buffer *b)
1458 {
1459         return b->data;
1460 }
1461 EXPORT_SYMBOL_GPL(dm_bufio_get_block_data);
1462
1463 void *dm_bufio_get_aux_data(struct dm_buffer *b)
1464 {
1465         return b + 1;
1466 }
1467 EXPORT_SYMBOL_GPL(dm_bufio_get_aux_data);
1468
1469 struct dm_bufio_client *dm_bufio_get_client(struct dm_buffer *b)
1470 {
1471         return b->c;
1472 }
1473 EXPORT_SYMBOL_GPL(dm_bufio_get_client);
1474
1475 static void drop_buffers(struct dm_bufio_client *c)
1476 {
1477         struct dm_buffer *b;
1478         int i;
1479         bool warned = false;
1480
1481         BUG_ON(dm_bufio_in_request());
1482
1483         /*
1484          * An optimization so that the buffers are not written one-by-one.
1485          */
1486         dm_bufio_write_dirty_buffers_async(c);
1487
1488         dm_bufio_lock(c);
1489
1490         while ((b = __get_unclaimed_buffer(c)))
1491                 __free_buffer_wake(b);
1492
1493         for (i = 0; i < LIST_SIZE; i++)
1494                 list_for_each_entry(b, &c->lru[i], lru_list) {
1495                         WARN_ON(!warned);
1496                         warned = true;
1497                         DMERR("leaked buffer %llx, hold count %u, list %d",
1498                               (unsigned long long)b->block, b->hold_count, i);
1499 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1500                         print_stack_trace(&b->stack_trace, 1);
1501                         b->hold_count = 0; /* mark unclaimed to avoid BUG_ON below */
1502 #endif
1503                 }
1504
1505 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1506         while ((b = __get_unclaimed_buffer(c)))
1507                 __free_buffer_wake(b);
1508 #endif
1509
1510         for (i = 0; i < LIST_SIZE; i++)
1511                 BUG_ON(!list_empty(&c->lru[i]));
1512
1513         dm_bufio_unlock(c);
1514 }
1515
1516 /*
1517  * We may not be able to evict this buffer if IO pending or the client
1518  * is still using it.  Caller is expected to know buffer is too old.
1519  *
1520  * And if GFP_NOFS is used, we must not do any I/O because we hold
1521  * dm_bufio_clients_lock and we would risk deadlock if the I/O gets
1522  * rerouted to different bufio client.
1523  */
1524 static bool __try_evict_buffer(struct dm_buffer *b, gfp_t gfp)
1525 {
1526         if (!(gfp & __GFP_FS)) {
1527                 if (test_bit(B_READING, &b->state) ||
1528                     test_bit(B_WRITING, &b->state) ||
1529                     test_bit(B_DIRTY, &b->state))
1530                         return false;
1531         }
1532
1533         if (b->hold_count)
1534                 return false;
1535
1536         __make_buffer_clean(b);
1537         __unlink_buffer(b);
1538         __free_buffer_wake(b);
1539
1540         return true;
1541 }
1542
1543 static unsigned get_retain_buffers(struct dm_bufio_client *c)
1544 {
1545         unsigned retain_bytes = ACCESS_ONCE(dm_bufio_retain_bytes);
1546         return retain_bytes / c->block_size;
1547 }
1548
1549 static unsigned long __scan(struct dm_bufio_client *c, unsigned long nr_to_scan,
1550                             gfp_t gfp_mask)
1551 {
1552         int l;
1553         struct dm_buffer *b, *tmp;
1554         unsigned long freed = 0;
1555         unsigned long count = nr_to_scan;
1556         unsigned retain_target = get_retain_buffers(c);
1557
1558         for (l = 0; l < LIST_SIZE; l++) {
1559                 list_for_each_entry_safe_reverse(b, tmp, &c->lru[l], lru_list) {
1560                         if (__try_evict_buffer(b, gfp_mask))
1561                                 freed++;
1562                         if (!--nr_to_scan || ((count - freed) <= retain_target))
1563                                 return freed;
1564                         cond_resched();
1565                 }
1566         }
1567         return freed;
1568 }
1569
1570 static unsigned long
1571 dm_bufio_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
1572 {
1573         struct dm_bufio_client *c;
1574         unsigned long freed;
1575
1576         c = container_of(shrink, struct dm_bufio_client, shrinker);
1577         if (sc->gfp_mask & __GFP_FS)
1578                 dm_bufio_lock(c);
1579         else if (!dm_bufio_trylock(c))
1580                 return SHRINK_STOP;
1581
1582         freed  = __scan(c, sc->nr_to_scan, sc->gfp_mask);
1583         dm_bufio_unlock(c);
1584         return freed;
1585 }
1586
1587 static unsigned long
1588 dm_bufio_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
1589 {
1590         struct dm_bufio_client *c;
1591         unsigned long count;
1592
1593         c = container_of(shrink, struct dm_bufio_client, shrinker);
1594         if (sc->gfp_mask & __GFP_FS)
1595                 dm_bufio_lock(c);
1596         else if (!dm_bufio_trylock(c))
1597                 return 0;
1598
1599         count = c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY];
1600         dm_bufio_unlock(c);
1601         return count;
1602 }
1603
1604 /*
1605  * Create the buffering interface
1606  */
1607 struct dm_bufio_client *dm_bufio_client_create(struct block_device *bdev, unsigned block_size,
1608                                                unsigned reserved_buffers, unsigned aux_size,
1609                                                void (*alloc_callback)(struct dm_buffer *),
1610                                                void (*write_callback)(struct dm_buffer *))
1611 {
1612         int r;
1613         struct dm_bufio_client *c;
1614         unsigned i;
1615
1616         BUG_ON(block_size < 1 << SECTOR_SHIFT ||
1617                (block_size & (block_size - 1)));
1618
1619         c = kzalloc(sizeof(*c), GFP_KERNEL);
1620         if (!c) {
1621                 r = -ENOMEM;
1622                 goto bad_client;
1623         }
1624         c->buffer_tree = RB_ROOT;
1625
1626         c->bdev = bdev;
1627         c->block_size = block_size;
1628         c->sectors_per_block_bits = __ffs(block_size) - SECTOR_SHIFT;
1629         c->pages_per_block_bits = (__ffs(block_size) >= PAGE_SHIFT) ?
1630                                   __ffs(block_size) - PAGE_SHIFT : 0;
1631         c->blocks_per_page_bits = (__ffs(block_size) < PAGE_SHIFT ?
1632                                   PAGE_SHIFT - __ffs(block_size) : 0);
1633
1634         c->aux_size = aux_size;
1635         c->alloc_callback = alloc_callback;
1636         c->write_callback = write_callback;
1637
1638         for (i = 0; i < LIST_SIZE; i++) {
1639                 INIT_LIST_HEAD(&c->lru[i]);
1640                 c->n_buffers[i] = 0;
1641         }
1642
1643         mutex_init(&c->lock);
1644         INIT_LIST_HEAD(&c->reserved_buffers);
1645         c->need_reserved_buffers = reserved_buffers;
1646
1647         c->minimum_buffers = DM_BUFIO_MIN_BUFFERS;
1648
1649         init_waitqueue_head(&c->free_buffer_wait);
1650         c->async_write_error = 0;
1651
1652         c->dm_io = dm_io_client_create();
1653         if (IS_ERR(c->dm_io)) {
1654                 r = PTR_ERR(c->dm_io);
1655                 goto bad_dm_io;
1656         }
1657
1658         mutex_lock(&dm_bufio_clients_lock);
1659         if (c->blocks_per_page_bits) {
1660                 if (!DM_BUFIO_CACHE_NAME(c)) {
1661                         DM_BUFIO_CACHE_NAME(c) = kasprintf(GFP_KERNEL, "dm_bufio_cache-%u", c->block_size);
1662                         if (!DM_BUFIO_CACHE_NAME(c)) {
1663                                 r = -ENOMEM;
1664                                 mutex_unlock(&dm_bufio_clients_lock);
1665                                 goto bad_cache;
1666                         }
1667                 }
1668
1669                 if (!DM_BUFIO_CACHE(c)) {
1670                         DM_BUFIO_CACHE(c) = kmem_cache_create(DM_BUFIO_CACHE_NAME(c),
1671                                                               c->block_size,
1672                                                               c->block_size, 0, NULL);
1673                         if (!DM_BUFIO_CACHE(c)) {
1674                                 r = -ENOMEM;
1675                                 mutex_unlock(&dm_bufio_clients_lock);
1676                                 goto bad_cache;
1677                         }
1678                 }
1679         }
1680         mutex_unlock(&dm_bufio_clients_lock);
1681
1682         while (c->need_reserved_buffers) {
1683                 struct dm_buffer *b = alloc_buffer(c, GFP_KERNEL);
1684
1685                 if (!b) {
1686                         r = -ENOMEM;
1687                         goto bad_buffer;
1688                 }
1689                 __free_buffer_wake(b);
1690         }
1691
1692         mutex_lock(&dm_bufio_clients_lock);
1693         dm_bufio_client_count++;
1694         list_add(&c->client_list, &dm_bufio_all_clients);
1695         __cache_size_refresh();
1696         mutex_unlock(&dm_bufio_clients_lock);
1697
1698         c->shrinker.count_objects = dm_bufio_shrink_count;
1699         c->shrinker.scan_objects = dm_bufio_shrink_scan;
1700         c->shrinker.seeks = 1;
1701         c->shrinker.batch = 0;
1702         register_shrinker(&c->shrinker);
1703
1704         return c;
1705
1706 bad_buffer:
1707 bad_cache:
1708         while (!list_empty(&c->reserved_buffers)) {
1709                 struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1710                                                  struct dm_buffer, lru_list);
1711                 list_del(&b->lru_list);
1712                 free_buffer(b);
1713         }
1714         dm_io_client_destroy(c->dm_io);
1715 bad_dm_io:
1716         kfree(c);
1717 bad_client:
1718         return ERR_PTR(r);
1719 }
1720 EXPORT_SYMBOL_GPL(dm_bufio_client_create);
1721
1722 /*
1723  * Free the buffering interface.
1724  * It is required that there are no references on any buffers.
1725  */
1726 void dm_bufio_client_destroy(struct dm_bufio_client *c)
1727 {
1728         unsigned i;
1729
1730         drop_buffers(c);
1731
1732         unregister_shrinker(&c->shrinker);
1733
1734         mutex_lock(&dm_bufio_clients_lock);
1735
1736         list_del(&c->client_list);
1737         dm_bufio_client_count--;
1738         __cache_size_refresh();
1739
1740         mutex_unlock(&dm_bufio_clients_lock);
1741
1742         BUG_ON(!RB_EMPTY_ROOT(&c->buffer_tree));
1743         BUG_ON(c->need_reserved_buffers);
1744
1745         while (!list_empty(&c->reserved_buffers)) {
1746                 struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1747                                                  struct dm_buffer, lru_list);
1748                 list_del(&b->lru_list);
1749                 free_buffer(b);
1750         }
1751
1752         for (i = 0; i < LIST_SIZE; i++)
1753                 if (c->n_buffers[i])
1754                         DMERR("leaked buffer count %d: %ld", i, c->n_buffers[i]);
1755
1756         for (i = 0; i < LIST_SIZE; i++)
1757                 BUG_ON(c->n_buffers[i]);
1758
1759         dm_io_client_destroy(c->dm_io);
1760         kfree(c);
1761 }
1762 EXPORT_SYMBOL_GPL(dm_bufio_client_destroy);
1763
1764 static unsigned get_max_age_hz(void)
1765 {
1766         unsigned max_age = ACCESS_ONCE(dm_bufio_max_age);
1767
1768         if (max_age > UINT_MAX / HZ)
1769                 max_age = UINT_MAX / HZ;
1770
1771         return max_age * HZ;
1772 }
1773
1774 static bool older_than(struct dm_buffer *b, unsigned long age_hz)
1775 {
1776         return time_after_eq(jiffies, b->last_accessed + age_hz);
1777 }
1778
1779 static void __evict_old_buffers(struct dm_bufio_client *c, unsigned long age_hz)
1780 {
1781         struct dm_buffer *b, *tmp;
1782         unsigned retain_target = get_retain_buffers(c);
1783         unsigned count;
1784
1785         dm_bufio_lock(c);
1786
1787         count = c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY];
1788         list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_CLEAN], lru_list) {
1789                 if (count <= retain_target)
1790                         break;
1791
1792                 if (!older_than(b, age_hz))
1793                         break;
1794
1795                 if (__try_evict_buffer(b, 0))
1796                         count--;
1797
1798                 cond_resched();
1799         }
1800
1801         dm_bufio_unlock(c);
1802 }
1803
1804 static void cleanup_old_buffers(void)
1805 {
1806         unsigned long max_age_hz = get_max_age_hz();
1807         struct dm_bufio_client *c;
1808
1809         mutex_lock(&dm_bufio_clients_lock);
1810
1811         list_for_each_entry(c, &dm_bufio_all_clients, client_list)
1812                 __evict_old_buffers(c, max_age_hz);
1813
1814         mutex_unlock(&dm_bufio_clients_lock);
1815 }
1816
1817 static struct workqueue_struct *dm_bufio_wq;
1818 static struct delayed_work dm_bufio_work;
1819
1820 static void work_fn(struct work_struct *w)
1821 {
1822         cleanup_old_buffers();
1823
1824         queue_delayed_work(dm_bufio_wq, &dm_bufio_work,
1825                            DM_BUFIO_WORK_TIMER_SECS * HZ);
1826 }
1827
1828 /*----------------------------------------------------------------
1829  * Module setup
1830  *--------------------------------------------------------------*/
1831
1832 /*
1833  * This is called only once for the whole dm_bufio module.
1834  * It initializes memory limit.
1835  */
1836 static int __init dm_bufio_init(void)
1837 {
1838         __u64 mem;
1839
1840         dm_bufio_allocated_kmem_cache = 0;
1841         dm_bufio_allocated_get_free_pages = 0;
1842         dm_bufio_allocated_vmalloc = 0;
1843         dm_bufio_current_allocated = 0;
1844
1845         memset(&dm_bufio_caches, 0, sizeof dm_bufio_caches);
1846         memset(&dm_bufio_cache_names, 0, sizeof dm_bufio_cache_names);
1847
1848         mem = (__u64)((totalram_pages - totalhigh_pages) *
1849                       DM_BUFIO_MEMORY_PERCENT / 100) << PAGE_SHIFT;
1850
1851         if (mem > ULONG_MAX)
1852                 mem = ULONG_MAX;
1853
1854 #ifdef CONFIG_MMU
1855         /*
1856          * Get the size of vmalloc space the same way as VMALLOC_TOTAL
1857          * in fs/proc/internal.h
1858          */
1859         if (mem > (VMALLOC_END - VMALLOC_START) * DM_BUFIO_VMALLOC_PERCENT / 100)
1860                 mem = (VMALLOC_END - VMALLOC_START) * DM_BUFIO_VMALLOC_PERCENT / 100;
1861 #endif
1862
1863         dm_bufio_default_cache_size = mem;
1864
1865         mutex_lock(&dm_bufio_clients_lock);
1866         __cache_size_refresh();
1867         mutex_unlock(&dm_bufio_clients_lock);
1868
1869         dm_bufio_wq = alloc_workqueue("dm_bufio_cache", WQ_MEM_RECLAIM, 0);
1870         if (!dm_bufio_wq)
1871                 return -ENOMEM;
1872
1873         INIT_DELAYED_WORK(&dm_bufio_work, work_fn);
1874         queue_delayed_work(dm_bufio_wq, &dm_bufio_work,
1875                            DM_BUFIO_WORK_TIMER_SECS * HZ);
1876
1877         return 0;
1878 }
1879
1880 /*
1881  * This is called once when unloading the dm_bufio module.
1882  */
1883 static void __exit dm_bufio_exit(void)
1884 {
1885         int bug = 0;
1886         int i;
1887
1888         cancel_delayed_work_sync(&dm_bufio_work);
1889         destroy_workqueue(dm_bufio_wq);
1890
1891         for (i = 0; i < ARRAY_SIZE(dm_bufio_caches); i++)
1892                 kmem_cache_destroy(dm_bufio_caches[i]);
1893
1894         for (i = 0; i < ARRAY_SIZE(dm_bufio_cache_names); i++)
1895                 kfree(dm_bufio_cache_names[i]);
1896
1897         if (dm_bufio_client_count) {
1898                 DMCRIT("%s: dm_bufio_client_count leaked: %d",
1899                         __func__, dm_bufio_client_count);
1900                 bug = 1;
1901         }
1902
1903         if (dm_bufio_current_allocated) {
1904                 DMCRIT("%s: dm_bufio_current_allocated leaked: %lu",
1905                         __func__, dm_bufio_current_allocated);
1906                 bug = 1;
1907         }
1908
1909         if (dm_bufio_allocated_get_free_pages) {
1910                 DMCRIT("%s: dm_bufio_allocated_get_free_pages leaked: %lu",
1911                        __func__, dm_bufio_allocated_get_free_pages);
1912                 bug = 1;
1913         }
1914
1915         if (dm_bufio_allocated_vmalloc) {
1916                 DMCRIT("%s: dm_bufio_vmalloc leaked: %lu",
1917                        __func__, dm_bufio_allocated_vmalloc);
1918                 bug = 1;
1919         }
1920
1921         BUG_ON(bug);
1922 }
1923
1924 module_init(dm_bufio_init)
1925 module_exit(dm_bufio_exit)
1926
1927 module_param_named(max_cache_size_bytes, dm_bufio_cache_size, ulong, S_IRUGO | S_IWUSR);
1928 MODULE_PARM_DESC(max_cache_size_bytes, "Size of metadata cache");
1929
1930 module_param_named(max_age_seconds, dm_bufio_max_age, uint, S_IRUGO | S_IWUSR);
1931 MODULE_PARM_DESC(max_age_seconds, "Max age of a buffer in seconds");
1932
1933 module_param_named(retain_bytes, dm_bufio_retain_bytes, uint, S_IRUGO | S_IWUSR);
1934 MODULE_PARM_DESC(retain_bytes, "Try to keep at least this many bytes cached in memory");
1935
1936 module_param_named(peak_allocated_bytes, dm_bufio_peak_allocated, ulong, S_IRUGO | S_IWUSR);
1937 MODULE_PARM_DESC(peak_allocated_bytes, "Tracks the maximum allocated memory");
1938
1939 module_param_named(allocated_kmem_cache_bytes, dm_bufio_allocated_kmem_cache, ulong, S_IRUGO);
1940 MODULE_PARM_DESC(allocated_kmem_cache_bytes, "Memory allocated with kmem_cache_alloc");
1941
1942 module_param_named(allocated_get_free_pages_bytes, dm_bufio_allocated_get_free_pages, ulong, S_IRUGO);
1943 MODULE_PARM_DESC(allocated_get_free_pages_bytes, "Memory allocated with get_free_pages");
1944
1945 module_param_named(allocated_vmalloc_bytes, dm_bufio_allocated_vmalloc, ulong, S_IRUGO);
1946 MODULE_PARM_DESC(allocated_vmalloc_bytes, "Memory allocated with vmalloc");
1947
1948 module_param_named(current_allocated_bytes, dm_bufio_current_allocated, ulong, S_IRUGO);
1949 MODULE_PARM_DESC(current_allocated_bytes, "Memory currently used by the cache");
1950
1951 MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
1952 MODULE_DESCRIPTION(DM_NAME " buffered I/O library");
1953 MODULE_LICENSE("GPL");