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/sched/mm.h>
  15 #include <linux/jiffies.h>
  16 #include <linux/vmalloc.h>
  17 #include <linux/shrinker.h>
  18 #include <linux/module.h>
  19 #include <linux/rbtree.h>
  20 #include <linux/stacktrace.h>
  21
  22 #define DM_MSG_PREFIX "bufio"
  23
  24 /*
  25  * Memory management policy:
  26  *      Limit the number of buffers to DM_BUFIO_MEMORY_PERCENT of main memory
  27  *      or DM_BUFIO_VMALLOC_PERCENT of vmalloc memory (whichever is lower).
  28  *      Always allocate at least DM_BUFIO_MIN_BUFFERS buffers.
  29  *      Start background writeback when there are DM_BUFIO_WRITEBACK_PERCENT
  30  *      dirty buffers.
  31  */
  32 #define DM_BUFIO_MIN_BUFFERS            8
  33
  34 #define DM_BUFIO_MEMORY_PERCENT         2
  35 #define DM_BUFIO_VMALLOC_PERCENT        25
  36 #define DM_BUFIO_WRITEBACK_PERCENT      75
  37
  38 /*
  39  * Check buffer ages in this interval (seconds)
  40  */
  41 #define DM_BUFIO_WORK_TIMER_SECS        30
  42
  43 /*
  44  * Free buffers when they are older than this (seconds)
  45  */
  46 #define DM_BUFIO_DEFAULT_AGE_SECS       300
  47
  48 /*
  49  * The nr of bytes of cached data to keep around.
  50  */
  51 #define DM_BUFIO_DEFAULT_RETAIN_BYTES   (256 * 1024)
  52
  53 /*
  54  * The number of bvec entries that are embedded directly in the buffer.
  55  * If the chunk size is larger, dm-io is used to do the io.
  56  */
  57 #define DM_BUFIO_INLINE_VECS            16
  58
  59 /*
  60  * Don't try to use kmem_cache_alloc for blocks larger than this.
  61  * For explanation, see alloc_buffer_data below.
  62  */
  63 #define DM_BUFIO_BLOCK_SIZE_SLAB_LIMIT  (PAGE_SIZE >> 1)
  64 #define DM_BUFIO_BLOCK_SIZE_GFP_LIMIT   (PAGE_SIZE << (MAX_ORDER - 1))
  65
  66 /*
  67  * dm_buffer->list_mode
  68  */
  69 #define LIST_CLEAN      0
  70 #define LIST_DIRTY      1
  71 #define LIST_SIZE       2
  72
  73 /*
  74  * Linking of buffers:
  75  *      All buffers are linked to cache_hash with their hash_list field.
  76  *
  77  *      Clean buffers that are not being written (B_WRITING not set)
  78  *      are linked to lru[LIST_CLEAN] with their lru_list field.
  79  *
  80  *      Dirty and clean buffers that are being written are linked to
  81  *      lru[LIST_DIRTY] with their lru_list field. When the write
  82  *      finishes, the buffer cannot be relinked immediately (because we
  83  *      are in an interrupt context and relinking requires process
  84  *      context), so some clean-not-writing buffers can be held on
  85  *      dirty_lru too.  They are later added to lru in the process
  86  *      context.
  87  */
  88 struct dm_bufio_client {
  89         struct mutex lock;
  90
  91         struct list_head lru[LIST_SIZE];
  92         unsigned long n_buffers[LIST_SIZE];
  93
  94         struct block_device *bdev;
  95         unsigned block_size;
  96         unsigned char sectors_per_block_bits;
  97         unsigned char pages_per_block_bits;
  98         unsigned char blocks_per_page_bits;
  99         unsigned aux_size;
 100         void (*alloc_callback)(struct dm_buffer *);
 101         void (*write_callback)(struct dm_buffer *);
 102
 103         struct dm_io_client *dm_io;
 104
 105         struct list_head reserved_buffers;
 106         unsigned need_reserved_buffers;
 107
 108         unsigned minimum_buffers;
 109
 110         struct rb_root buffer_tree;
 111         wait_queue_head_t free_buffer_wait;
 112
 113         sector_t start;
 114
 115         int async_write_error;
 116
 117         struct list_head client_list;
 118         struct shrinker shrinker;
 119 };
 120
 121 /*
 122  * Buffer state bits.
 123  */
 124 #define B_READING       0
 125 #define B_WRITING       1
 126 #define B_DIRTY         2
 127
 128 /*
 129  * Describes how the block was allocated:
 130  * kmem_cache_alloc(), __get_free_pages() or vmalloc().
 131  * See the comment at alloc_buffer_data.
 132  */
 133 enum data_mode {
 134         DATA_MODE_SLAB = 0,
 135         DATA_MODE_GET_FREE_PAGES = 1,
 136         DATA_MODE_VMALLOC = 2,
 137         DATA_MODE_LIMIT = 3
 138 };
 139
 140 struct dm_buffer {
 141         struct rb_node node;
 142         struct list_head lru_list;
 143         sector_t block;
 144         void *data;
 145         enum data_mode data_mode;
 146         unsigned char list_mode;                /* LIST_* */
 147         unsigned hold_count;
 148         int read_error;
 149         int write_error;
 150         unsigned long state;
 151         unsigned long last_accessed;
 152         struct dm_bufio_client *c;
 153         struct list_head write_list;
 154         struct bio bio;
 155         struct bio_vec bio_vec[DM_BUFIO_INLINE_VECS];
 156 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
 157 #define MAX_STACK 10
 158         struct stack_trace stack_trace;
 159         unsigned long stack_entries[MAX_STACK];
 160 #endif
 161 };
 162
 163 /*----------------------------------------------------------------*/
 164
 165 static struct kmem_cache *dm_bufio_caches[PAGE_SHIFT - SECTOR_SHIFT];
 166 static char *dm_bufio_cache_names[PAGE_SHIFT - SECTOR_SHIFT];
 167
 168 static inline int dm_bufio_cache_index(struct dm_bufio_client *c)
 169 {
 170         unsigned ret = c->blocks_per_page_bits - 1;
 171
 172         BUG_ON(ret >= ARRAY_SIZE(dm_bufio_caches));
 173
 174         return ret;
 175 }
 176
 177 #define DM_BUFIO_CACHE(c)       (dm_bufio_caches[dm_bufio_cache_index(c)])
 178 #define DM_BUFIO_CACHE_NAME(c)  (dm_bufio_cache_names[dm_bufio_cache_index(c)])
 179
 180 #define dm_bufio_in_request()   (!!current->bio_list)
 181
 182 static void dm_bufio_lock(struct dm_bufio_client *c)
 183 {
 184         mutex_lock_nested(&c->lock, dm_bufio_in_request());
 185 }
 186
 187 static int dm_bufio_trylock(struct dm_bufio_client *c)
 188 {
 189         return mutex_trylock(&c->lock);
 190 }
 191
 192 static void dm_bufio_unlock(struct dm_bufio_client *c)
 193 {
 194         mutex_unlock(&c->lock);
 195 }
 196
 197 /*----------------------------------------------------------------*/
 198
 199 /*
 200  * Default cache size: available memory divided by the ratio.
 201  */
 202 static unsigned long dm_bufio_default_cache_size;
 203
 204 /*
 205  * Total cache size set by the user.
 206  */
 207 static unsigned long dm_bufio_cache_size;
 208
 209 /*
 210  * A copy of dm_bufio_cache_size because dm_bufio_cache_size can change
 211  * at any time.  If it disagrees, the user has changed cache size.
 212  */
 213 static unsigned long dm_bufio_cache_size_latch;
 214
 215 static DEFINE_SPINLOCK(param_spinlock);
 216
 217 /*
 218  * Buffers are freed after this timeout
 219  */
 220 static unsigned dm_bufio_max_age = DM_BUFIO_DEFAULT_AGE_SECS;
 221 static unsigned long dm_bufio_retain_bytes = DM_BUFIO_DEFAULT_RETAIN_BYTES;
 222
 223 static unsigned long dm_bufio_peak_allocated;
 224 static unsigned long dm_bufio_allocated_kmem_cache;
 225 static unsigned long dm_bufio_allocated_get_free_pages;
 226 static unsigned long dm_bufio_allocated_vmalloc;
 227 static unsigned long dm_bufio_current_allocated;
 228
 229 /*----------------------------------------------------------------*/
 230
 231 /*
 232  * Per-client cache: dm_bufio_cache_size / dm_bufio_client_count
 233  */
 234 static unsigned long dm_bufio_cache_size_per_client;
 235
 236 /*
 237  * The current number of clients.
 238  */
 239 static int dm_bufio_client_count;
 240
 241 /*
 242  * The list of all clients.
 243  */
 244 static LIST_HEAD(dm_bufio_all_clients);
 245
 246 /*
 247  * This mutex protects dm_bufio_cache_size_latch,
 248  * dm_bufio_cache_size_per_client and dm_bufio_client_count
 249  */
 250 static DEFINE_MUTEX(dm_bufio_clients_lock);
 251
 252 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
 253 static void buffer_record_stack(struct dm_buffer *b)
 254 {
 255         b->stack_trace.nr_entries = 0;
 256         b->stack_trace.max_entries = MAX_STACK;
 257         b->stack_trace.entries = b->stack_entries;
 258         b->stack_trace.skip = 2;
 259         save_stack_trace(&b->stack_trace);
 260 }
 261 #endif
 262
 263 /*----------------------------------------------------------------
 264  * A red/black tree acts as an index for all the buffers.
 265  *--------------------------------------------------------------*/
 266 static struct dm_buffer *__find(struct dm_bufio_client *c, sector_t block)
 267 {
 268         struct rb_node *n = c->buffer_tree.rb_node;
 269         struct dm_buffer *b;
 270
 271         while (n) {
 272                 b = container_of(n, struct dm_buffer, node);
 273
 274                 if (b->block == block)
 275                         return b;
 276
 277                 n = (b->block < block) ? n->rb_left : n->rb_right;
 278         }
 279
 280         return NULL;
 281 }
 282
 283 static void __insert(struct dm_bufio_client *c, struct dm_buffer *b)
 284 {
 285         struct rb_node **new = &c->buffer_tree.rb_node, *parent = NULL;
 286         struct dm_buffer *found;
 287
 288         while (*new) {
 289                 found = container_of(*new, struct dm_buffer, node);
 290
 291                 if (found->block == b->block) {
 292                         BUG_ON(found != b);
 293                         return;
 294                 }
 295
 296                 parent = *new;
 297                 new = (found->block < b->block) ?
 298                         &((*new)->rb_left) : &((*new)->rb_right);
 299         }
 300
 301         rb_link_node(&b->node, parent, new);
 302         rb_insert_color(&b->node, &c->buffer_tree);
 303 }
 304
 305 static void __remove(struct dm_bufio_client *c, struct dm_buffer *b)
 306 {
 307         rb_erase(&b->node, &c->buffer_tree);
 308 }
 309
 310 /*----------------------------------------------------------------*/
 311
 312 static void adjust_total_allocated(enum data_mode data_mode, long diff)
 313 {
 314         static unsigned long * const class_ptr[DATA_MODE_LIMIT] = {
 315                 &dm_bufio_allocated_kmem_cache,
 316                 &dm_bufio_allocated_get_free_pages,
 317                 &dm_bufio_allocated_vmalloc,
 318         };
 319
 320         spin_lock(&param_spinlock);
 321
 322         *class_ptr[data_mode] += diff;
 323
 324         dm_bufio_current_allocated += diff;
 325
 326         if (dm_bufio_current_allocated > dm_bufio_peak_allocated)
 327                 dm_bufio_peak_allocated = dm_bufio_current_allocated;
 328
 329         spin_unlock(&param_spinlock);
 330 }
 331
 332 /*
 333  * Change the number of clients and recalculate per-client limit.
 334  */
 335 static void __cache_size_refresh(void)
 336 {
 337         BUG_ON(!mutex_is_locked(&dm_bufio_clients_lock));
 338         BUG_ON(dm_bufio_client_count < 0);
 339
 340         dm_bufio_cache_size_latch = ACCESS_ONCE(dm_bufio_cache_size);
 341
 342         /*
 343          * Use default if set to 0 and report the actual cache size used.
 344          */
 345         if (!dm_bufio_cache_size_latch) {
 346                 (void)cmpxchg(&dm_bufio_cache_size, 0,
 347                               dm_bufio_default_cache_size);
 348                 dm_bufio_cache_size_latch = dm_bufio_default_cache_size;
 349         }
 350
 351         dm_bufio_cache_size_per_client = dm_bufio_cache_size_latch /
 352                                          (dm_bufio_client_count ? : 1);
 353 }
 354
 355 /*
 356  * Allocating buffer data.
 357  *
 358  * Small buffers are allocated with kmem_cache, to use space optimally.
 359  *
 360  * For large buffers, we choose between get_free_pages and vmalloc.
 361  * Each has advantages and disadvantages.
 362  *
 363  * __get_free_pages can randomly fail if the memory is fragmented.
 364  * __vmalloc won't randomly fail, but vmalloc space is limited (it may be
 365  * as low as 128M) so using it for caching is not appropriate.
 366  *
 367  * If the allocation may fail we use __get_free_pages. Memory fragmentation
 368  * won't have a fatal effect here, but it just causes flushes of some other
 369  * buffers and more I/O will be performed. Don't use __get_free_pages if it
 370  * always fails (i.e. order >= MAX_ORDER).
 371  *
 372  * If the allocation shouldn't fail we use __vmalloc. This is only for the
 373  * initial reserve allocation, so there's no risk of wasting all vmalloc
 374  * space.
 375  */
 376 static void *alloc_buffer_data(struct dm_bufio_client *c, gfp_t gfp_mask,
 377                                enum data_mode *data_mode)
 378 {
 379         unsigned noio_flag;
 380         void *ptr;
 381
 382         if (c->block_size <= DM_BUFIO_BLOCK_SIZE_SLAB_LIMIT) {
 383                 *data_mode = DATA_MODE_SLAB;
 384                 return kmem_cache_alloc(DM_BUFIO_CACHE(c), gfp_mask);
 385         }
 386
 387         if (c->block_size <= DM_BUFIO_BLOCK_SIZE_GFP_LIMIT &&
 388             gfp_mask & __GFP_NORETRY) {
 389                 *data_mode = DATA_MODE_GET_FREE_PAGES;
 390                 return (void *)__get_free_pages(gfp_mask,
 391                                                 c->pages_per_block_bits);
 392         }
 393
 394         *data_mode = DATA_MODE_VMALLOC;
 395
 396         /*
 397          * __vmalloc allocates the data pages and auxiliary structures with
 398          * gfp_flags that were specified, but pagetables are always allocated
 399          * with GFP_KERNEL, no matter what was specified as gfp_mask.
 400          *
 401          * Consequently, we must set per-process flag PF_MEMALLOC_NOIO so that
 402          * all allocations done by this process (including pagetables) are done
 403          * as if GFP_NOIO was specified.
 404          */
 405
 406         if (gfp_mask & __GFP_NORETRY)
 407                 noio_flag = memalloc_noio_save();
 408
 409         ptr = __vmalloc(c->block_size, gfp_mask, PAGE_KERNEL);
 410
 411         if (gfp_mask & __GFP_NORETRY)
 412                 memalloc_noio_restore(noio_flag);
 413
 414         return ptr;
 415 }
 416
 417 /*
 418  * Free buffer's data.
 419  */
 420 static void free_buffer_data(struct dm_bufio_client *c,
 421                              void *data, enum data_mode data_mode)
 422 {
 423         switch (data_mode) {
 424         case DATA_MODE_SLAB:
 425                 kmem_cache_free(DM_BUFIO_CACHE(c), data);
 426                 break;
 427
 428         case DATA_MODE_GET_FREE_PAGES:
 429                 free_pages((unsigned long)data, c->pages_per_block_bits);
 430                 break;
 431
 432         case DATA_MODE_VMALLOC:
 433                 vfree(data);
 434                 break;
 435
 436         default:
 437                 DMCRIT("dm_bufio_free_buffer_data: bad data mode: %d",
 438                        data_mode);
 439                 BUG();
 440         }
 441 }
 442
 443 /*
 444  * Allocate buffer and its data.
 445  */
 446 static struct dm_buffer *alloc_buffer(struct dm_bufio_client *c, gfp_t gfp_mask)
 447 {
 448         struct dm_buffer *b = kmalloc(sizeof(struct dm_buffer) + c->aux_size,
 449                                       gfp_mask);
 450
 451         if (!b)
 452                 return NULL;
 453
 454         b->c = c;
 455
 456         b->data = alloc_buffer_data(c, gfp_mask, &b->data_mode);
 457         if (!b->data) {
 458                 kfree(b);
 459                 return NULL;
 460         }
 461
 462         adjust_total_allocated(b->data_mode, (long)c->block_size);
 463
 464 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
 465         memset(&b->stack_trace, 0, sizeof(b->stack_trace));
 466 #endif
 467         return b;
 468 }
 469
 470 /*
 471  * Free buffer and its data.
 472  */
 473 static void free_buffer(struct dm_buffer *b)
 474 {
 475         struct dm_bufio_client *c = b->c;
 476
 477         adjust_total_allocated(b->data_mode, -(long)c->block_size);
 478
 479         free_buffer_data(c, b->data, b->data_mode);
 480         kfree(b);
 481 }
 482
 483 /*
 484  * Link buffer to the hash list and clean or dirty queue.
 485  */
 486 static void __link_buffer(struct dm_buffer *b, sector_t block, int dirty)
 487 {
 488         struct dm_bufio_client *c = b->c;
 489
 490         c->n_buffers[dirty]++;
 491         b->block = block;
 492         b->list_mode = dirty;
 493         list_add(&b->lru_list, &c->lru[dirty]);
 494         __insert(b->c, b);
 495         b->last_accessed = jiffies;
 496 }
 497
 498 /*
 499  * Unlink buffer from the hash list and dirty or clean queue.
 500  */
 501 static void __unlink_buffer(struct dm_buffer *b)
 502 {
 503         struct dm_bufio_client *c = b->c;
 504
 505         BUG_ON(!c->n_buffers[b->list_mode]);
 506
 507         c->n_buffers[b->list_mode]--;
 508         __remove(b->c, b);
 509         list_del(&b->lru_list);
 510 }
 511
 512 /*
 513  * Place the buffer to the head of dirty or clean LRU queue.
 514  */
 515 static void __relink_lru(struct dm_buffer *b, int dirty)
 516 {
 517         struct dm_bufio_client *c = b->c;
 518
 519         BUG_ON(!c->n_buffers[b->list_mode]);
 520
 521         c->n_buffers[b->list_mode]--;
 522         c->n_buffers[dirty]++;
 523         b->list_mode = dirty;
 524         list_move(&b->lru_list, &c->lru[dirty]);
 525         b->last_accessed = jiffies;
 526 }
 527
 528 /*----------------------------------------------------------------
 529  * Submit I/O on the buffer.
 530  *
 531  * Bio interface is faster but it has some problems:
 532  *      the vector list is limited (increasing this limit increases
 533  *      memory-consumption per buffer, so it is not viable);
 534  *
 535  *      the memory must be direct-mapped, not vmalloced;
 536  *
 537  *      the I/O driver can reject requests spuriously if it thinks that
 538  *      the requests are too big for the device or if they cross a
 539  *      controller-defined memory boundary.
 540  *
 541  * If the buffer is small enough (up to DM_BUFIO_INLINE_VECS pages) and
 542  * it is not vmalloced, try using the bio interface.
 543  *
 544  * If the buffer is big, if it is vmalloced or if the underlying device
 545  * rejects the bio because it is too large, use dm-io layer to do the I/O.
 546  * The dm-io layer splits the I/O into multiple requests, avoiding the above
 547  * shortcomings.
 548  *--------------------------------------------------------------*/
 549
 550 /*
 551  * dm-io completion routine. It just calls b->bio.bi_end_io, pretending
 552  * that the request was handled directly with bio interface.
 553  */
 554 static void dmio_complete(unsigned long error, void *context)
 555 {
 556         struct dm_buffer *b = context;
 557
 558         b->bio.bi_error = error ? -EIO : 0;
 559         b->bio.bi_end_io(&b->bio);
 560 }
 561
 562 static void use_dmio(struct dm_buffer *b, int rw, sector_t sector,
 563                      unsigned n_sectors, bio_end_io_t *end_io)
 564 {
 565         int r;
 566         struct dm_io_request io_req = {
 567                 .bi_op = rw,
 568                 .bi_op_flags = 0,
 569                 .notify.fn = dmio_complete,
 570                 .notify.context = b,
 571                 .client = b->c->dm_io,
 572         };
 573         struct dm_io_region region = {
 574                 .bdev = b->c->bdev,
 575                 .sector = sector,
 576                 .count = n_sectors,
 577         };
 578
 579         if (b->data_mode != DATA_MODE_VMALLOC) {
 580                 io_req.mem.type = DM_IO_KMEM;
 581                 io_req.mem.ptr.addr = b->data;
 582         } else {
 583                 io_req.mem.type = DM_IO_VMA;
 584                 io_req.mem.ptr.vma = b->data;
 585         }
 586
 587         b->bio.bi_end_io = end_io;
 588
 589         r = dm_io(&io_req, 1, &region, NULL);
 590         if (r) {
 591                 b->bio.bi_error = r;
 592                 end_io(&b->bio);
 593         }
 594 }
 595
 596 static void inline_endio(struct bio *bio)
 597 {
 598         bio_end_io_t *end_fn = bio->bi_private;
 599         int error = bio->bi_error;
 600
 601         /*
 602          * Reset the bio to free any attached resources
 603          * (e.g. bio integrity profiles).
 604          */
 605         bio_reset(bio);
 606
 607         bio->bi_error = error;
 608         end_fn(bio);
 609 }
 610
 611 static void use_inline_bio(struct dm_buffer *b, int rw, sector_t sector,
 612                            unsigned n_sectors, bio_end_io_t *end_io)
 613 {
 614         char *ptr;
 615         int len;
 616
 617         bio_init(&b->bio, b->bio_vec, DM_BUFIO_INLINE_VECS);
 618         b->bio.bi_iter.bi_sector = sector;
 619         b->bio.bi_bdev = b->c->bdev;
 620         b->bio.bi_end_io = inline_endio;
 621         /*
 622          * Use of .bi_private isn't a problem here because
 623          * the dm_buffer's inline bio is local to bufio.
 624          */
 625         b->bio.bi_private = end_io;
 626         bio_set_op_attrs(&b->bio, rw, 0);
 627
 628         /*
 629          * We assume that if len >= PAGE_SIZE ptr is page-aligned.
 630          * If len < PAGE_SIZE the buffer doesn't cross page boundary.
 631          */
 632         ptr = b->data;
 633         len = n_sectors << SECTOR_SHIFT;
 634
 635         if (len >= PAGE_SIZE)
 636                 BUG_ON((unsigned long)ptr & (PAGE_SIZE - 1));
 637         else
 638                 BUG_ON((unsigned long)ptr & (len - 1));
 639
 640         do {
 641                 if (!bio_add_page(&b->bio, virt_to_page(ptr),
 642                                   len < PAGE_SIZE ? len : PAGE_SIZE,
 643                                   offset_in_page(ptr))) {
 644                         BUG_ON(b->c->block_size <= PAGE_SIZE);
 645                         use_dmio(b, rw, sector, n_sectors, end_io);
 646                         return;
 647                 }
 648
 649                 len -= PAGE_SIZE;
 650                 ptr += PAGE_SIZE;
 651         } while (len > 0);
 652
 653         submit_bio(&b->bio);
 654 }
 655
 656 static void submit_io(struct dm_buffer *b, int rw, bio_end_io_t *end_io)
 657 {
 658         unsigned n_sectors;
 659         sector_t sector;
 660
 661         if (rw == WRITE && b->c->write_callback)
 662                 b->c->write_callback(b);
 663
 664         sector = (b->block << b->c->sectors_per_block_bits) + b->c->start;
 665         n_sectors = 1 << b->c->sectors_per_block_bits;
 666
 667         if (n_sectors <= ((DM_BUFIO_INLINE_VECS * PAGE_SIZE) >> SECTOR_SHIFT) &&
 668             b->data_mode != DATA_MODE_VMALLOC)
 669                 use_inline_bio(b, rw, sector, n_sectors, end_io);
 670         else
 671                 use_dmio(b, rw, sector, n_sectors, end_io);
 672 }
 673
 674 /*----------------------------------------------------------------
 675  * Writing dirty buffers
 676  *--------------------------------------------------------------*/
 677
 678 /*
 679  * The endio routine for write.
 680  *
 681  * Set the error, clear B_WRITING bit and wake anyone who was waiting on
 682  * it.
 683  */
 684 static void write_endio(struct bio *bio)
 685 {
 686         struct dm_buffer *b = container_of(bio, struct dm_buffer, bio);
 687
 688         b->write_error = bio->bi_error;
 689         if (unlikely(bio->bi_error)) {
 690                 struct dm_bufio_client *c = b->c;
 691                 int error = bio->bi_error;
 692                 (void)cmpxchg(&c->async_write_error, 0, error);
 693         }
 694
 695         BUG_ON(!test_bit(B_WRITING, &b->state));
 696
 697         smp_mb__before_atomic();
 698         clear_bit(B_WRITING, &b->state);
 699         smp_mb__after_atomic();
 700
 701         wake_up_bit(&b->state, B_WRITING);
 702 }
 703
 704 /*
 705  * Initiate a write on a dirty buffer, but don't wait for it.
 706  *
 707  * - If the buffer is not dirty, exit.
 708  * - If there some previous write going on, wait for it to finish (we can't
 709  *   have two writes on the same buffer simultaneously).
 710  * - Submit our write and don't wait on it. We set B_WRITING indicating
 711  *   that there is a write in progress.
 712  */
 713 static void __write_dirty_buffer(struct dm_buffer *b,
 714                                  struct list_head *write_list)
 715 {
 716         if (!test_bit(B_DIRTY, &b->state))
 717                 return;
 718
 719         clear_bit(B_DIRTY, &b->state);
 720         wait_on_bit_lock_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
 721
 722         if (!write_list)
 723                 submit_io(b, WRITE, write_endio);
 724         else
 725                 list_add_tail(&b->write_list, write_list);
 726 }
 727
 728 static void __flush_write_list(struct list_head *write_list)
 729 {
 730         struct blk_plug plug;
 731         blk_start_plug(&plug);
 732         while (!list_empty(write_list)) {
 733                 struct dm_buffer *b =
 734                         list_entry(write_list->next, struct dm_buffer, write_list);
 735                 list_del(&b->write_list);
 736                 submit_io(b, WRITE, write_endio);
 737                 cond_resched();
 738         }
 739         blk_finish_plug(&plug);
 740 }
 741
 742 /*
 743  * Wait until any activity on the buffer finishes.  Possibly write the
 744  * buffer if it is dirty.  When this function finishes, there is no I/O
 745  * running on the buffer and the buffer is not dirty.
 746  */
 747 static void __make_buffer_clean(struct dm_buffer *b)
 748 {
 749         BUG_ON(b->hold_count);
 750
 751         if (!b->state)  /* fast case */
 752                 return;
 753
 754         wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE);
 755         __write_dirty_buffer(b, NULL);
 756         wait_on_bit_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
 757 }
 758
 759 /*
 760  * Find some buffer that is not held by anybody, clean it, unlink it and
 761  * return it.
 762  */
 763 static struct dm_buffer *__get_unclaimed_buffer(struct dm_bufio_client *c)
 764 {
 765         struct dm_buffer *b;
 766
 767         list_for_each_entry_reverse(b, &c->lru[LIST_CLEAN], lru_list) {
 768                 BUG_ON(test_bit(B_WRITING, &b->state));
 769                 BUG_ON(test_bit(B_DIRTY, &b->state));
 770
 771                 if (!b->hold_count) {
 772                         __make_buffer_clean(b);
 773                         __unlink_buffer(b);
 774                         return b;
 775                 }
 776                 cond_resched();
 777         }
 778
 779         list_for_each_entry_reverse(b, &c->lru[LIST_DIRTY], lru_list) {
 780                 BUG_ON(test_bit(B_READING, &b->state));
 781
 782                 if (!b->hold_count) {
 783                         __make_buffer_clean(b);
 784                         __unlink_buffer(b);
 785                         return b;
 786                 }
 787                 cond_resched();
 788         }
 789
 790         return NULL;
 791 }
 792
 793 /*
 794  * Wait until some other threads free some buffer or release hold count on
 795  * some buffer.
 796  *
 797  * This function is entered with c->lock held, drops it and regains it
 798  * before exiting.
 799  */
 800 static void __wait_for_free_buffer(struct dm_bufio_client *c)
 801 {
 802         DECLARE_WAITQUEUE(wait, current);
 803
 804         add_wait_queue(&c->free_buffer_wait, &wait);
 805         set_current_state(TASK_UNINTERRUPTIBLE);
 806         dm_bufio_unlock(c);
 807
 808         io_schedule();
 809
 810         remove_wait_queue(&c->free_buffer_wait, &wait);
 811
 812         dm_bufio_lock(c);
 813 }
 814
 815 enum new_flag {
 816         NF_FRESH = 0,
 817         NF_READ = 1,
 818         NF_GET = 2,
 819         NF_PREFETCH = 3
 820 };
 821
 822 /*
 823  * Allocate a new buffer. If the allocation is not possible, wait until
 824  * some other thread frees a buffer.
 825  *
 826  * May drop the lock and regain it.
 827  */
 828 static struct dm_buffer *__alloc_buffer_wait_no_callback(struct dm_bufio_client *c, enum new_flag nf)
 829 {
 830         struct dm_buffer *b;
 831         bool tried_noio_alloc = false;
 832
 833         /*
 834          * dm-bufio is resistant to allocation failures (it just keeps
 835          * one buffer reserved in cases all the allocations fail).
 836          * So set flags to not try too hard:
 837          *      GFP_NOWAIT: don't wait; if we need to sleep we'll release our
 838          *                  mutex and wait ourselves.
 839          *      __GFP_NORETRY: don't retry and rather return failure
 840          *      __GFP_NOMEMALLOC: don't use emergency reserves
 841          *      __GFP_NOWARN: don't print a warning in case of failure
 842          *
 843          * For debugging, if we set the cache size to 1, no new buffers will
 844          * be allocated.
 845          */
 846         while (1) {
 847                 if (dm_bufio_cache_size_latch != 1) {
 848                         b = alloc_buffer(c, GFP_NOWAIT | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
 849                         if (b)
 850                                 return b;
 851                 }
 852
 853                 if (nf == NF_PREFETCH)
 854                         return NULL;
 855
 856                 if (dm_bufio_cache_size_latch != 1 && !tried_noio_alloc) {
 857                         dm_bufio_unlock(c);
 858                         b = alloc_buffer(c, GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
 859                         dm_bufio_lock(c);
 860                         if (b)
 861                                 return b;
 862                         tried_noio_alloc = true;
 863                 }
 864
 865                 if (!list_empty(&c->reserved_buffers)) {
 866                         b = list_entry(c->reserved_buffers.next,
 867                                        struct dm_buffer, lru_list);
 868                         list_del(&b->lru_list);
 869                         c->need_reserved_buffers++;
 870
 871                         return b;
 872                 }
 873
 874                 b = __get_unclaimed_buffer(c);
 875                 if (b)
 876                         return b;
 877
 878                 __wait_for_free_buffer(c);
 879         }
 880 }
 881
 882 static struct dm_buffer *__alloc_buffer_wait(struct dm_bufio_client *c, enum new_flag nf)
 883 {
 884         struct dm_buffer *b = __alloc_buffer_wait_no_callback(c, nf);
 885
 886         if (!b)
 887                 return NULL;
 888
 889         if (c->alloc_callback)
 890                 c->alloc_callback(b);
 891
 892         return b;
 893 }
 894
 895 /*
 896  * Free a buffer and wake other threads waiting for free buffers.
 897  */
 898 static void __free_buffer_wake(struct dm_buffer *b)
 899 {
 900         struct dm_bufio_client *c = b->c;
 901
 902         if (!c->need_reserved_buffers)
 903                 free_buffer(b);
 904         else {
 905                 list_add(&b->lru_list, &c->reserved_buffers);
 906                 c->need_reserved_buffers--;
 907         }
 908
 909         wake_up(&c->free_buffer_wait);
 910 }
 911
 912 static void __write_dirty_buffers_async(struct dm_bufio_client *c, int no_wait,
 913                                         struct list_head *write_list)
 914 {
 915         struct dm_buffer *b, *tmp;
 916
 917         list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
 918                 BUG_ON(test_bit(B_READING, &b->state));
 919
 920                 if (!test_bit(B_DIRTY, &b->state) &&
 921                     !test_bit(B_WRITING, &b->state)) {
 922                         __relink_lru(b, LIST_CLEAN);
 923                         continue;
 924                 }
 925
 926                 if (no_wait && test_bit(B_WRITING, &b->state))
 927                         return;
 928
 929                 __write_dirty_buffer(b, write_list);
 930                 cond_resched();
 931         }
 932 }
 933
 934 /*
 935  * Get writeback threshold and buffer limit for a given client.
 936  */
 937 static void __get_memory_limit(struct dm_bufio_client *c,
 938                                unsigned long *threshold_buffers,
 939                                unsigned long *limit_buffers)
 940 {
 941         unsigned long buffers;
 942
 943         if (unlikely(ACCESS_ONCE(dm_bufio_cache_size) != dm_bufio_cache_size_latch)) {
 944                 if (mutex_trylock(&dm_bufio_clients_lock)) {
 945                         __cache_size_refresh();
 946                         mutex_unlock(&dm_bufio_clients_lock);
 947                 }
 948         }
 949
 950         buffers = dm_bufio_cache_size_per_client >>
 951                   (c->sectors_per_block_bits + SECTOR_SHIFT);
 952
 953         if (buffers < c->minimum_buffers)
 954                 buffers = c->minimum_buffers;
 955
 956         *limit_buffers = buffers;
 957         *threshold_buffers = buffers * DM_BUFIO_WRITEBACK_PERCENT / 100;
 958 }
 959
 960 /*
 961  * Check if we're over watermark.
 962  * If we are over threshold_buffers, start freeing buffers.
 963  * If we're over "limit_buffers", block until we get under the limit.
 964  */
 965 static void __check_watermark(struct dm_bufio_client *c,
 966                               struct list_head *write_list)
 967 {
 968         unsigned long threshold_buffers, limit_buffers;
 969
 970         __get_memory_limit(c, &threshold_buffers, &limit_buffers);
 971
 972         while (c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY] >
 973                limit_buffers) {
 974
 975                 struct dm_buffer *b = __get_unclaimed_buffer(c);
 976
 977                 if (!b)
 978                         return;
 979
 980                 __free_buffer_wake(b);
 981                 cond_resched();
 982         }
 983
 984         if (c->n_buffers[LIST_DIRTY] > threshold_buffers)
 985                 __write_dirty_buffers_async(c, 1, write_list);
 986 }
 987
 988 /*----------------------------------------------------------------
 989  * Getting a buffer
 990  *--------------------------------------------------------------*/
 991
 992 static struct dm_buffer *__bufio_new(struct dm_bufio_client *c, sector_t block,
 993                                      enum new_flag nf, int *need_submit,
 994                                      struct list_head *write_list)
 995 {
 996         struct dm_buffer *b, *new_b = NULL;
 997
 998         *need_submit = 0;
 999
1000         b = __find(c, block);
1001         if (b)
1002                 goto found_buffer;
1003
1004         if (nf == NF_GET)
1005                 return NULL;
1006
1007         new_b = __alloc_buffer_wait(c, nf);
1008         if (!new_b)
1009                 return NULL;
1010
1011         /*
1012          * We've had a period where the mutex was unlocked, so need to
1013          * recheck the hash table.
1014          */
1015         b = __find(c, block);
1016         if (b) {
1017                 __free_buffer_wake(new_b);
1018                 goto found_buffer;
1019         }
1020
1021         __check_watermark(c, write_list);
1022
1023         b = new_b;
1024         b->hold_count = 1;
1025         b->read_error = 0;
1026         b->write_error = 0;
1027         __link_buffer(b, block, LIST_CLEAN);
1028
1029         if (nf == NF_FRESH) {
1030                 b->state = 0;
1031                 return b;
1032         }
1033
1034         b->state = 1 << B_READING;
1035         *need_submit = 1;
1036
1037         return b;
1038
1039 found_buffer:
1040         if (nf == NF_PREFETCH)
1041                 return NULL;
1042         /*
1043          * Note: it is essential that we don't wait for the buffer to be
1044          * read if dm_bufio_get function is used. Both dm_bufio_get and
1045          * dm_bufio_prefetch can be used in the driver request routine.
1046          * If the user called both dm_bufio_prefetch and dm_bufio_get on
1047          * the same buffer, it would deadlock if we waited.
1048          */
1049         if (nf == NF_GET && unlikely(test_bit(B_READING, &b->state)))
1050                 return NULL;
1051
1052         b->hold_count++;
1053         __relink_lru(b, test_bit(B_DIRTY, &b->state) ||
1054                      test_bit(B_WRITING, &b->state));
1055         return b;
1056 }
1057
1058 /*
1059  * The endio routine for reading: set the error, clear the bit and wake up
1060  * anyone waiting on the buffer.
1061  */
1062 static void read_endio(struct bio *bio)
1063 {
1064         struct dm_buffer *b = container_of(bio, struct dm_buffer, bio);
1065
1066         b->read_error = bio->bi_error;
1067
1068         BUG_ON(!test_bit(B_READING, &b->state));
1069
1070         smp_mb__before_atomic();
1071         clear_bit(B_READING, &b->state);
1072         smp_mb__after_atomic();
1073
1074         wake_up_bit(&b->state, B_READING);
1075 }
1076
1077 /*
1078  * A common routine for dm_bufio_new and dm_bufio_read.  Operation of these
1079  * functions is similar except that dm_bufio_new doesn't read the
1080  * buffer from the disk (assuming that the caller overwrites all the data
1081  * and uses dm_bufio_mark_buffer_dirty to write new data back).
1082  */
1083 static void *new_read(struct dm_bufio_client *c, sector_t block,
1084                       enum new_flag nf, struct dm_buffer **bp)
1085 {
1086         int need_submit;
1087         struct dm_buffer *b;
1088
1089         LIST_HEAD(write_list);
1090
1091         dm_bufio_lock(c);
1092         b = __bufio_new(c, block, nf, &need_submit, &write_list);
1093 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1094         if (b && b->hold_count == 1)
1095                 buffer_record_stack(b);
1096 #endif
1097         dm_bufio_unlock(c);
1098
1099         __flush_write_list(&write_list);
1100
1101         if (!b)
1102                 return NULL;
1103
1104         if (need_submit)
1105                 submit_io(b, READ, read_endio);
1106
1107         wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE);
1108
1109         if (b->read_error) {
1110                 int error = b->read_error;
1111
1112                 dm_bufio_release(b);
1113
1114                 return ERR_PTR(error);
1115         }
1116
1117         *bp = b;
1118
1119         return b->data;
1120 }
1121
1122 void *dm_bufio_get(struct dm_bufio_client *c, sector_t block,
1123                    struct dm_buffer **bp)
1124 {
1125         return new_read(c, block, NF_GET, bp);
1126 }
1127 EXPORT_SYMBOL_GPL(dm_bufio_get);
1128
1129 void *dm_bufio_read(struct dm_bufio_client *c, sector_t block,
1130                     struct dm_buffer **bp)
1131 {
1132         BUG_ON(dm_bufio_in_request());
1133
1134         return new_read(c, block, NF_READ, bp);
1135 }
1136 EXPORT_SYMBOL_GPL(dm_bufio_read);
1137
1138 void *dm_bufio_new(struct dm_bufio_client *c, sector_t block,
1139                    struct dm_buffer **bp)
1140 {
1141         BUG_ON(dm_bufio_in_request());
1142
1143         return new_read(c, block, NF_FRESH, bp);
1144 }
1145 EXPORT_SYMBOL_GPL(dm_bufio_new);
1146
1147 void dm_bufio_prefetch(struct dm_bufio_client *c,
1148                        sector_t block, unsigned n_blocks)
1149 {
1150         struct blk_plug plug;
1151
1152         LIST_HEAD(write_list);
1153
1154         BUG_ON(dm_bufio_in_request());
1155
1156         blk_start_plug(&plug);
1157         dm_bufio_lock(c);
1158
1159         for (; n_blocks--; block++) {
1160                 int need_submit;
1161                 struct dm_buffer *b;
1162                 b = __bufio_new(c, block, NF_PREFETCH, &need_submit,
1163                                 &write_list);
1164                 if (unlikely(!list_empty(&write_list))) {
1165                         dm_bufio_unlock(c);
1166                         blk_finish_plug(&plug);
1167                         __flush_write_list(&write_list);
1168                         blk_start_plug(&plug);
1169                         dm_bufio_lock(c);
1170                 }
1171                 if (unlikely(b != NULL)) {
1172                         dm_bufio_unlock(c);
1173
1174                         if (need_submit)
1175                                 submit_io(b, READ, read_endio);
1176                         dm_bufio_release(b);
1177
1178                         cond_resched();
1179
1180                         if (!n_blocks)
1181                                 goto flush_plug;
1182                         dm_bufio_lock(c);
1183                 }
1184         }
1185
1186         dm_bufio_unlock(c);
1187
1188 flush_plug:
1189         blk_finish_plug(&plug);
1190 }
1191 EXPORT_SYMBOL_GPL(dm_bufio_prefetch);
1192
1193 void dm_bufio_release(struct dm_buffer *b)
1194 {
1195         struct dm_bufio_client *c = b->c;
1196
1197         dm_bufio_lock(c);
1198
1199         BUG_ON(!b->hold_count);
1200
1201         b->hold_count--;
1202         if (!b->hold_count) {
1203                 wake_up(&c->free_buffer_wait);
1204
1205                 /*
1206                  * If there were errors on the buffer, and the buffer is not
1207                  * to be written, free the buffer. There is no point in caching
1208                  * invalid buffer.
1209                  */
1210                 if ((b->read_error || b->write_error) &&
1211                     !test_bit(B_READING, &b->state) &&
1212                     !test_bit(B_WRITING, &b->state) &&
1213                     !test_bit(B_DIRTY, &b->state)) {
1214                         __unlink_buffer(b);
1215                         __free_buffer_wake(b);
1216                 }
1217         }
1218
1219         dm_bufio_unlock(c);
1220 }
1221 EXPORT_SYMBOL_GPL(dm_bufio_release);
1222
1223 void dm_bufio_mark_buffer_dirty(struct dm_buffer *b)
1224 {
1225         struct dm_bufio_client *c = b->c;
1226
1227         dm_bufio_lock(c);
1228
1229         BUG_ON(test_bit(B_READING, &b->state));
1230
1231         if (!test_and_set_bit(B_DIRTY, &b->state))
1232                 __relink_lru(b, LIST_DIRTY);
1233
1234         dm_bufio_unlock(c);
1235 }
1236 EXPORT_SYMBOL_GPL(dm_bufio_mark_buffer_dirty);
1237
1238 void dm_bufio_write_dirty_buffers_async(struct dm_bufio_client *c)
1239 {
1240         LIST_HEAD(write_list);
1241
1242         BUG_ON(dm_bufio_in_request());
1243
1244         dm_bufio_lock(c);
1245         __write_dirty_buffers_async(c, 0, &write_list);
1246         dm_bufio_unlock(c);
1247         __flush_write_list(&write_list);
1248 }
1249 EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers_async);
1250
1251 /*
1252  * For performance, it is essential that the buffers are written asynchronously
1253  * and simultaneously (so that the block layer can merge the writes) and then
1254  * waited upon.
1255  *
1256  * Finally, we flush hardware disk cache.
1257  */
1258 int dm_bufio_write_dirty_buffers(struct dm_bufio_client *c)
1259 {
1260         int a, f;
1261         unsigned long buffers_processed = 0;
1262         struct dm_buffer *b, *tmp;
1263
1264         LIST_HEAD(write_list);
1265
1266         dm_bufio_lock(c);
1267         __write_dirty_buffers_async(c, 0, &write_list);
1268         dm_bufio_unlock(c);
1269         __flush_write_list(&write_list);
1270         dm_bufio_lock(c);
1271
1272 again:
1273         list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
1274                 int dropped_lock = 0;
1275
1276                 if (buffers_processed < c->n_buffers[LIST_DIRTY])
1277                         buffers_processed++;
1278
1279                 BUG_ON(test_bit(B_READING, &b->state));
1280
1281                 if (test_bit(B_WRITING, &b->state)) {
1282                         if (buffers_processed < c->n_buffers[LIST_DIRTY]) {
1283                                 dropped_lock = 1;
1284                                 b->hold_count++;
1285                                 dm_bufio_unlock(c);
1286                                 wait_on_bit_io(&b->state, B_WRITING,
1287                                                TASK_UNINTERRUPTIBLE);
1288                                 dm_bufio_lock(c);
1289                                 b->hold_count--;
1290                         } else
1291                                 wait_on_bit_io(&b->state, B_WRITING,
1292                                                TASK_UNINTERRUPTIBLE);
1293                 }
1294
1295                 if (!test_bit(B_DIRTY, &b->state) &&
1296                     !test_bit(B_WRITING, &b->state))
1297                         __relink_lru(b, LIST_CLEAN);
1298
1299                 cond_resched();
1300
1301                 /*
1302                  * If we dropped the lock, the list is no longer consistent,
1303                  * so we must restart the search.
1304                  *
1305                  * In the most common case, the buffer just processed is
1306                  * relinked to the clean list, so we won't loop scanning the
1307                  * same buffer again and again.
1308                  *
1309                  * This may livelock if there is another thread simultaneously
1310                  * dirtying buffers, so we count the number of buffers walked
1311                  * and if it exceeds the total number of buffers, it means that
1312                  * someone is doing some writes simultaneously with us.  In
1313                  * this case, stop, dropping the lock.
1314                  */
1315                 if (dropped_lock)
1316                         goto again;
1317         }
1318         wake_up(&c->free_buffer_wait);
1319         dm_bufio_unlock(c);
1320
1321         a = xchg(&c->async_write_error, 0);
1322         f = dm_bufio_issue_flush(c);
1323         if (a)
1324                 return a;
1325
1326         return f;
1327 }
1328 EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers);
1329
1330 /*
1331  * Use dm-io to send and empty barrier flush the device.
1332  */
1333 int dm_bufio_issue_flush(struct dm_bufio_client *c)
1334 {
1335         struct dm_io_request io_req = {
1336                 .bi_op = REQ_OP_WRITE,
1337                 .bi_op_flags = REQ_PREFLUSH | REQ_SYNC,
1338                 .mem.type = DM_IO_KMEM,
1339                 .mem.ptr.addr = NULL,
1340                 .client = c->dm_io,
1341         };
1342         struct dm_io_region io_reg = {
1343                 .bdev = c->bdev,
1344                 .sector = 0,
1345                 .count = 0,
1346         };
1347
1348         BUG_ON(dm_bufio_in_request());
1349
1350         return dm_io(&io_req, 1, &io_reg, NULL);
1351 }
1352 EXPORT_SYMBOL_GPL(dm_bufio_issue_flush);
1353
1354 /*
1355  * We first delete any other buffer that may be at that new location.
1356  *
1357  * Then, we write the buffer to the original location if it was dirty.
1358  *
1359  * Then, if we are the only one who is holding the buffer, relink the buffer
1360  * in the hash queue for the new location.
1361  *
1362  * If there was someone else holding the buffer, we write it to the new
1363  * location but not relink it, because that other user needs to have the buffer
1364  * at the same place.
1365  */
1366 void dm_bufio_release_move(struct dm_buffer *b, sector_t new_block)
1367 {
1368         struct dm_bufio_client *c = b->c;
1369         struct dm_buffer *new;
1370
1371         BUG_ON(dm_bufio_in_request());
1372
1373         dm_bufio_lock(c);
1374
1375 retry:
1376         new = __find(c, new_block);
1377         if (new) {
1378                 if (new->hold_count) {
1379                         __wait_for_free_buffer(c);
1380                         goto retry;
1381                 }
1382
1383                 /*
1384                  * FIXME: Is there any point waiting for a write that's going
1385                  * to be overwritten in a bit?
1386                  */
1387                 __make_buffer_clean(new);
1388                 __unlink_buffer(new);
1389                 __free_buffer_wake(new);
1390         }
1391
1392         BUG_ON(!b->hold_count);
1393         BUG_ON(test_bit(B_READING, &b->state));
1394
1395         __write_dirty_buffer(b, NULL);
1396         if (b->hold_count == 1) {
1397                 wait_on_bit_io(&b->state, B_WRITING,
1398                                TASK_UNINTERRUPTIBLE);
1399                 set_bit(B_DIRTY, &b->state);
1400                 __unlink_buffer(b);
1401                 __link_buffer(b, new_block, LIST_DIRTY);
1402         } else {
1403                 sector_t old_block;
1404                 wait_on_bit_lock_io(&b->state, B_WRITING,
1405                                     TASK_UNINTERRUPTIBLE);
1406                 /*
1407                  * Relink buffer to "new_block" so that write_callback
1408                  * sees "new_block" as a block number.
1409                  * After the write, link the buffer back to old_block.
1410                  * All this must be done in bufio lock, so that block number
1411                  * change isn't visible to other threads.
1412                  */
1413                 old_block = b->block;
1414                 __unlink_buffer(b);
1415                 __link_buffer(b, new_block, b->list_mode);
1416                 submit_io(b, WRITE, write_endio);
1417                 wait_on_bit_io(&b->state, B_WRITING,
1418                                TASK_UNINTERRUPTIBLE);
1419                 __unlink_buffer(b);
1420                 __link_buffer(b, old_block, b->list_mode);
1421         }
1422
1423         dm_bufio_unlock(c);
1424         dm_bufio_release(b);
1425 }
1426 EXPORT_SYMBOL_GPL(dm_bufio_release_move);
1427
1428 /*
1429  * Free the given buffer.
1430  *
1431  * This is just a hint, if the buffer is in use or dirty, this function
1432  * does nothing.
1433  */
1434 void dm_bufio_forget(struct dm_bufio_client *c, sector_t block)
1435 {
1436         struct dm_buffer *b;
1437
1438         dm_bufio_lock(c);
1439
1440         b = __find(c, block);
1441         if (b && likely(!b->hold_count) && likely(!b->state)) {
1442                 __unlink_buffer(b);
1443                 __free_buffer_wake(b);
1444         }
1445
1446         dm_bufio_unlock(c);
1447 }
1448 EXPORT_SYMBOL(dm_bufio_forget);
1449
1450 void dm_bufio_set_minimum_buffers(struct dm_bufio_client *c, unsigned n)
1451 {
1452         c->minimum_buffers = n;
1453 }
1454 EXPORT_SYMBOL(dm_bufio_set_minimum_buffers);
1455
1456 unsigned dm_bufio_get_block_size(struct dm_bufio_client *c)
1457 {
1458         return c->block_size;
1459 }
1460 EXPORT_SYMBOL_GPL(dm_bufio_get_block_size);
1461
1462 sector_t dm_bufio_get_device_size(struct dm_bufio_client *c)
1463 {
1464         return i_size_read(c->bdev->bd_inode) >>
1465                            (SECTOR_SHIFT + c->sectors_per_block_bits);
1466 }
1467 EXPORT_SYMBOL_GPL(dm_bufio_get_device_size);
1468
1469 sector_t dm_bufio_get_block_number(struct dm_buffer *b)
1470 {
1471         return b->block;
1472 }
1473 EXPORT_SYMBOL_GPL(dm_bufio_get_block_number);
1474
1475 void *dm_bufio_get_block_data(struct dm_buffer *b)
1476 {
1477         return b->data;
1478 }
1479 EXPORT_SYMBOL_GPL(dm_bufio_get_block_data);
1480
1481 void *dm_bufio_get_aux_data(struct dm_buffer *b)
1482 {
1483         return b + 1;
1484 }
1485 EXPORT_SYMBOL_GPL(dm_bufio_get_aux_data);
1486
1487 struct dm_bufio_client *dm_bufio_get_client(struct dm_buffer *b)
1488 {
1489         return b->c;
1490 }
1491 EXPORT_SYMBOL_GPL(dm_bufio_get_client);
1492
1493 static void drop_buffers(struct dm_bufio_client *c)
1494 {
1495         struct dm_buffer *b;
1496         int i;
1497         bool warned = false;
1498
1499         BUG_ON(dm_bufio_in_request());
1500
1501         /*
1502          * An optimization so that the buffers are not written one-by-one.
1503          */
1504         dm_bufio_write_dirty_buffers_async(c);
1505
1506         dm_bufio_lock(c);
1507
1508         while ((b = __get_unclaimed_buffer(c)))
1509                 __free_buffer_wake(b);
1510
1511         for (i = 0; i < LIST_SIZE; i++)
1512                 list_for_each_entry(b, &c->lru[i], lru_list) {
1513                         WARN_ON(!warned);
1514                         warned = true;
1515                         DMERR("leaked buffer %llx, hold count %u, list %d",
1516                               (unsigned long long)b->block, b->hold_count, i);
1517 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1518                         print_stack_trace(&b->stack_trace, 1);
1519                         b->hold_count = 0; /* mark unclaimed to avoid BUG_ON below */
1520 #endif
1521                 }
1522
1523 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1524         while ((b = __get_unclaimed_buffer(c)))
1525                 __free_buffer_wake(b);
1526 #endif
1527
1528         for (i = 0; i < LIST_SIZE; i++)
1529                 BUG_ON(!list_empty(&c->lru[i]));
1530
1531         dm_bufio_unlock(c);
1532 }
1533
1534 /*
1535  * We may not be able to evict this buffer if IO pending or the client
1536  * is still using it.  Caller is expected to know buffer is too old.
1537  *
1538  * And if GFP_NOFS is used, we must not do any I/O because we hold
1539  * dm_bufio_clients_lock and we would risk deadlock if the I/O gets
1540  * rerouted to different bufio client.
1541  */
1542 static bool __try_evict_buffer(struct dm_buffer *b, gfp_t gfp)
1543 {
1544         if (!(gfp & __GFP_FS)) {
1545                 if (test_bit(B_READING, &b->state) ||
1546                     test_bit(B_WRITING, &b->state) ||
1547                     test_bit(B_DIRTY, &b->state))
1548                         return false;
1549         }
1550
1551         if (b->hold_count)
1552                 return false;
1553
1554         __make_buffer_clean(b);
1555         __unlink_buffer(b);
1556         __free_buffer_wake(b);
1557
1558         return true;
1559 }
1560
1561 static unsigned long get_retain_buffers(struct dm_bufio_client *c)
1562 {
1563         unsigned long retain_bytes = ACCESS_ONCE(dm_bufio_retain_bytes);
1564         return retain_bytes >> (c->sectors_per_block_bits + SECTOR_SHIFT);
1565 }
1566
1567 static unsigned long __scan(struct dm_bufio_client *c, unsigned long nr_to_scan,
1568                             gfp_t gfp_mask)
1569 {
1570         int l;
1571         struct dm_buffer *b, *tmp;
1572         unsigned long freed = 0;
1573         unsigned long count = nr_to_scan;
1574         unsigned long retain_target = get_retain_buffers(c);
1575
1576         for (l = 0; l < LIST_SIZE; l++) {
1577                 list_for_each_entry_safe_reverse(b, tmp, &c->lru[l], lru_list) {
1578                         if (__try_evict_buffer(b, gfp_mask))
1579                                 freed++;
1580                         if (!--nr_to_scan || ((count - freed) <= retain_target))
1581                                 return freed;
1582                         cond_resched();
1583                 }
1584         }
1585         return freed;
1586 }
1587
1588 static unsigned long
1589 dm_bufio_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
1590 {
1591         struct dm_bufio_client *c;
1592         unsigned long freed;
1593
1594         c = container_of(shrink, struct dm_bufio_client, shrinker);
1595         if (sc->gfp_mask & __GFP_FS)
1596                 dm_bufio_lock(c);
1597         else if (!dm_bufio_trylock(c))
1598                 return SHRINK_STOP;
1599
1600         freed  = __scan(c, sc->nr_to_scan, sc->gfp_mask);
1601         dm_bufio_unlock(c);
1602         return freed;
1603 }
1604
1605 static unsigned long
1606 dm_bufio_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
1607 {
1608         struct dm_bufio_client *c = container_of(shrink, struct dm_bufio_client, shrinker);
1609
1610         return ACCESS_ONCE(c->n_buffers[LIST_CLEAN]) + ACCESS_ONCE(c->n_buffers[LIST_DIRTY]);
1611 }
1612
1613 /*
1614  * Create the buffering interface
1615  */
1616 struct dm_bufio_client *dm_bufio_client_create(struct block_device *bdev, unsigned block_size,
1617                                                unsigned reserved_buffers, unsigned aux_size,
1618                                                void (*alloc_callback)(struct dm_buffer *),
1619                                                void (*write_callback)(struct dm_buffer *))
1620 {
1621         int r;
1622         struct dm_bufio_client *c;
1623         unsigned i;
1624
1625         BUG_ON(block_size < 1 << SECTOR_SHIFT ||
1626                (block_size & (block_size - 1)));
1627
1628         c = kzalloc(sizeof(*c), GFP_KERNEL);
1629         if (!c) {
1630                 r = -ENOMEM;
1631                 goto bad_client;
1632         }
1633         c->buffer_tree = RB_ROOT;
1634
1635         c->bdev = bdev;
1636         c->block_size = block_size;
1637         c->sectors_per_block_bits = __ffs(block_size) - SECTOR_SHIFT;
1638         c->pages_per_block_bits = (__ffs(block_size) >= PAGE_SHIFT) ?
1639                                   __ffs(block_size) - PAGE_SHIFT : 0;
1640         c->blocks_per_page_bits = (__ffs(block_size) < PAGE_SHIFT ?
1641                                   PAGE_SHIFT - __ffs(block_size) : 0);
1642
1643         c->aux_size = aux_size;
1644         c->alloc_callback = alloc_callback;
1645         c->write_callback = write_callback;
1646
1647         for (i = 0; i < LIST_SIZE; i++) {
1648                 INIT_LIST_HEAD(&c->lru[i]);
1649                 c->n_buffers[i] = 0;
1650         }
1651
1652         mutex_init(&c->lock);
1653         INIT_LIST_HEAD(&c->reserved_buffers);
1654         c->need_reserved_buffers = reserved_buffers;
1655
1656         c->minimum_buffers = DM_BUFIO_MIN_BUFFERS;
1657
1658         init_waitqueue_head(&c->free_buffer_wait);
1659         c->async_write_error = 0;
1660
1661         c->dm_io = dm_io_client_create();
1662         if (IS_ERR(c->dm_io)) {
1663                 r = PTR_ERR(c->dm_io);
1664                 goto bad_dm_io;
1665         }
1666
1667         mutex_lock(&dm_bufio_clients_lock);
1668         if (c->blocks_per_page_bits) {
1669                 if (!DM_BUFIO_CACHE_NAME(c)) {
1670                         DM_BUFIO_CACHE_NAME(c) = kasprintf(GFP_KERNEL, "dm_bufio_cache-%u", c->block_size);
1671                         if (!DM_BUFIO_CACHE_NAME(c)) {
1672                                 r = -ENOMEM;
1673                                 mutex_unlock(&dm_bufio_clients_lock);
1674                                 goto bad_cache;
1675                         }
1676                 }
1677
1678                 if (!DM_BUFIO_CACHE(c)) {
1679                         DM_BUFIO_CACHE(c) = kmem_cache_create(DM_BUFIO_CACHE_NAME(c),
1680                                                               c->block_size,
1681                                                               c->block_size, 0, NULL);
1682                         if (!DM_BUFIO_CACHE(c)) {
1683                                 r = -ENOMEM;
1684                                 mutex_unlock(&dm_bufio_clients_lock);
1685                                 goto bad_cache;
1686                         }
1687                 }
1688         }
1689         mutex_unlock(&dm_bufio_clients_lock);
1690
1691         while (c->need_reserved_buffers) {
1692                 struct dm_buffer *b = alloc_buffer(c, GFP_KERNEL);
1693
1694                 if (!b) {
1695                         r = -ENOMEM;
1696                         goto bad_buffer;
1697                 }
1698                 __free_buffer_wake(b);
1699         }
1700
1701         mutex_lock(&dm_bufio_clients_lock);
1702         dm_bufio_client_count++;
1703         list_add(&c->client_list, &dm_bufio_all_clients);
1704         __cache_size_refresh();
1705         mutex_unlock(&dm_bufio_clients_lock);
1706
1707         c->shrinker.count_objects = dm_bufio_shrink_count;
1708         c->shrinker.scan_objects = dm_bufio_shrink_scan;
1709         c->shrinker.seeks = 1;
1710         c->shrinker.batch = 0;
1711         register_shrinker(&c->shrinker);
1712
1713         return c;
1714
1715 bad_buffer:
1716 bad_cache:
1717         while (!list_empty(&c->reserved_buffers)) {
1718                 struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1719                                                  struct dm_buffer, lru_list);
1720                 list_del(&b->lru_list);
1721                 free_buffer(b);
1722         }
1723         dm_io_client_destroy(c->dm_io);
1724 bad_dm_io:
1725         kfree(c);
1726 bad_client:
1727         return ERR_PTR(r);
1728 }
1729 EXPORT_SYMBOL_GPL(dm_bufio_client_create);
1730
1731 /*
1732  * Free the buffering interface.
1733  * It is required that there are no references on any buffers.
1734  */
1735 void dm_bufio_client_destroy(struct dm_bufio_client *c)
1736 {
1737         unsigned i;
1738
1739         drop_buffers(c);
1740
1741         unregister_shrinker(&c->shrinker);
1742
1743         mutex_lock(&dm_bufio_clients_lock);
1744
1745         list_del(&c->client_list);
1746         dm_bufio_client_count--;
1747         __cache_size_refresh();
1748
1749         mutex_unlock(&dm_bufio_clients_lock);
1750
1751         BUG_ON(!RB_EMPTY_ROOT(&c->buffer_tree));
1752         BUG_ON(c->need_reserved_buffers);
1753
1754         while (!list_empty(&c->reserved_buffers)) {
1755                 struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1756                                                  struct dm_buffer, lru_list);
1757                 list_del(&b->lru_list);
1758                 free_buffer(b);
1759         }
1760
1761         for (i = 0; i < LIST_SIZE; i++)
1762                 if (c->n_buffers[i])
1763                         DMERR("leaked buffer count %d: %ld", i, c->n_buffers[i]);
1764
1765         for (i = 0; i < LIST_SIZE; i++)
1766                 BUG_ON(c->n_buffers[i]);
1767
1768         dm_io_client_destroy(c->dm_io);
1769         kfree(c);
1770 }
1771 EXPORT_SYMBOL_GPL(dm_bufio_client_destroy);
1772
1773 void dm_bufio_set_sector_offset(struct dm_bufio_client *c, sector_t start)
1774 {
1775         c->start = start;
1776 }
1777 EXPORT_SYMBOL_GPL(dm_bufio_set_sector_offset);
1778
1779 static unsigned get_max_age_hz(void)
1780 {
1781         unsigned max_age = ACCESS_ONCE(dm_bufio_max_age);
1782
1783         if (max_age > UINT_MAX / HZ)
1784                 max_age = UINT_MAX / HZ;
1785
1786         return max_age * HZ;
1787 }
1788
1789 static bool older_than(struct dm_buffer *b, unsigned long age_hz)
1790 {
1791         return time_after_eq(jiffies, b->last_accessed + age_hz);
1792 }
1793
1794 static void __evict_old_buffers(struct dm_bufio_client *c, unsigned long age_hz)
1795 {
1796         struct dm_buffer *b, *tmp;
1797         unsigned long retain_target = get_retain_buffers(c);
1798         unsigned long count;
1799         LIST_HEAD(write_list);
1800
1801         dm_bufio_lock(c);
1802
1803         __check_watermark(c, &write_list);
1804         if (unlikely(!list_empty(&write_list))) {
1805                 dm_bufio_unlock(c);
1806                 __flush_write_list(&write_list);
1807                 dm_bufio_lock(c);
1808         }
1809
1810         count = c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY];
1811         list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_CLEAN], lru_list) {
1812                 if (count <= retain_target)
1813                         break;
1814
1815                 if (!older_than(b, age_hz))
1816                         break;
1817
1818                 if (__try_evict_buffer(b, 0))
1819                         count--;
1820
1821                 cond_resched();
1822         }
1823
1824         dm_bufio_unlock(c);
1825 }
1826
1827 static void cleanup_old_buffers(void)
1828 {
1829         unsigned long max_age_hz = get_max_age_hz();
1830         struct dm_bufio_client *c;
1831
1832         mutex_lock(&dm_bufio_clients_lock);
1833
1834         __cache_size_refresh();
1835
1836         list_for_each_entry(c, &dm_bufio_all_clients, client_list)
1837                 __evict_old_buffers(c, max_age_hz);
1838
1839         mutex_unlock(&dm_bufio_clients_lock);
1840 }
1841
1842 static struct workqueue_struct *dm_bufio_wq;
1843 static struct delayed_work dm_bufio_work;
1844
1845 static void work_fn(struct work_struct *w)
1846 {
1847         cleanup_old_buffers();
1848
1849         queue_delayed_work(dm_bufio_wq, &dm_bufio_work,
1850                            DM_BUFIO_WORK_TIMER_SECS * HZ);
1851 }
1852
1853 /*----------------------------------------------------------------
1854  * Module setup
1855  *--------------------------------------------------------------*/
1856
1857 /*
1858  * This is called only once for the whole dm_bufio module.
1859  * It initializes memory limit.
1860  */
1861 static int __init dm_bufio_init(void)
1862 {
1863         __u64 mem;
1864
1865         dm_bufio_allocated_kmem_cache = 0;
1866         dm_bufio_allocated_get_free_pages = 0;
1867         dm_bufio_allocated_vmalloc = 0;
1868         dm_bufio_current_allocated = 0;
1869
1870         memset(&dm_bufio_caches, 0, sizeof dm_bufio_caches);
1871         memset(&dm_bufio_cache_names, 0, sizeof dm_bufio_cache_names);
1872
1873         mem = (__u64)((totalram_pages - totalhigh_pages) *
1874                       DM_BUFIO_MEMORY_PERCENT / 100) << PAGE_SHIFT;
1875
1876         if (mem > ULONG_MAX)
1877                 mem = ULONG_MAX;
1878
1879 #ifdef CONFIG_MMU
1880         /*
1881          * Get the size of vmalloc space the same way as VMALLOC_TOTAL
1882          * in fs/proc/internal.h
1883          */
1884         if (mem > (VMALLOC_END - VMALLOC_START) * DM_BUFIO_VMALLOC_PERCENT / 100)
1885                 mem = (VMALLOC_END - VMALLOC_START) * DM_BUFIO_VMALLOC_PERCENT / 100;
1886 #endif
1887
1888         dm_bufio_default_cache_size = mem;
1889
1890         mutex_lock(&dm_bufio_clients_lock);
1891         __cache_size_refresh();
1892         mutex_unlock(&dm_bufio_clients_lock);
1893
1894         dm_bufio_wq = alloc_workqueue("dm_bufio_cache", WQ_MEM_RECLAIM, 0);
1895         if (!dm_bufio_wq)
1896                 return -ENOMEM;
1897
1898         INIT_DELAYED_WORK(&dm_bufio_work, work_fn);
1899         queue_delayed_work(dm_bufio_wq, &dm_bufio_work,
1900                            DM_BUFIO_WORK_TIMER_SECS * HZ);
1901
1902         return 0;
1903 }
1904
1905 /*
1906  * This is called once when unloading the dm_bufio module.
1907  */
1908 static void __exit dm_bufio_exit(void)
1909 {
1910         int bug = 0;
1911         int i;
1912
1913         cancel_delayed_work_sync(&dm_bufio_work);
1914         destroy_workqueue(dm_bufio_wq);
1915
1916         for (i = 0; i < ARRAY_SIZE(dm_bufio_caches); i++)
1917                 kmem_cache_destroy(dm_bufio_caches[i]);
1918
1919         for (i = 0; i < ARRAY_SIZE(dm_bufio_cache_names); i++)
1920                 kfree(dm_bufio_cache_names[i]);
1921
1922         if (dm_bufio_client_count) {
1923                 DMCRIT("%s: dm_bufio_client_count leaked: %d",
1924                         __func__, dm_bufio_client_count);
1925                 bug = 1;
1926         }
1927
1928         if (dm_bufio_current_allocated) {
1929                 DMCRIT("%s: dm_bufio_current_allocated leaked: %lu",
1930                         __func__, dm_bufio_current_allocated);
1931                 bug = 1;
1932         }
1933
1934         if (dm_bufio_allocated_get_free_pages) {
1935                 DMCRIT("%s: dm_bufio_allocated_get_free_pages leaked: %lu",
1936                        __func__, dm_bufio_allocated_get_free_pages);
1937                 bug = 1;
1938         }
1939
1940         if (dm_bufio_allocated_vmalloc) {
1941                 DMCRIT("%s: dm_bufio_vmalloc leaked: %lu",
1942                        __func__, dm_bufio_allocated_vmalloc);
1943                 bug = 1;
1944         }
1945
1946         BUG_ON(bug);
1947 }
1948
1949 module_init(dm_bufio_init)
1950 module_exit(dm_bufio_exit)
1951
1952 module_param_named(max_cache_size_bytes, dm_bufio_cache_size, ulong, S_IRUGO | S_IWUSR);
1953 MODULE_PARM_DESC(max_cache_size_bytes, "Size of metadata cache");
1954
1955 module_param_named(max_age_seconds, dm_bufio_max_age, uint, S_IRUGO | S_IWUSR);
1956 MODULE_PARM_DESC(max_age_seconds, "Max age of a buffer in seconds");
1957
1958 module_param_named(retain_bytes, dm_bufio_retain_bytes, ulong, S_IRUGO | S_IWUSR);
1959 MODULE_PARM_DESC(retain_bytes, "Try to keep at least this many bytes cached in memory");
1960
1961 module_param_named(peak_allocated_bytes, dm_bufio_peak_allocated, ulong, S_IRUGO | S_IWUSR);
1962 MODULE_PARM_DESC(peak_allocated_bytes, "Tracks the maximum allocated memory");
1963
1964 module_param_named(allocated_kmem_cache_bytes, dm_bufio_allocated_kmem_cache, ulong, S_IRUGO);
1965 MODULE_PARM_DESC(allocated_kmem_cache_bytes, "Memory allocated with kmem_cache_alloc");
1966
1967 module_param_named(allocated_get_free_pages_bytes, dm_bufio_allocated_get_free_pages, ulong, S_IRUGO);
1968 MODULE_PARM_DESC(allocated_get_free_pages_bytes, "Memory allocated with get_free_pages");
1969
1970 module_param_named(allocated_vmalloc_bytes, dm_bufio_allocated_vmalloc, ulong, S_IRUGO);
1971 MODULE_PARM_DESC(allocated_vmalloc_bytes, "Memory allocated with vmalloc");
1972
1973 module_param_named(current_allocated_bytes, dm_bufio_current_allocated, ulong, S_IRUGO);
1974 MODULE_PARM_DESC(current_allocated_bytes, "Memory currently used by the cache");
1975
1976 MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
1977 MODULE_DESCRIPTION(DM_NAME " buffered I/O library");
1978 MODULE_LICENSE("GPL");