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