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