fs/squashfs/cache.c

   1 /*
   2  * Squashfs - a compressed read only filesystem for Linux
   3  *
   4  * Copyright (c) 2002, 2003, 2004, 2005, 2006, 2007, 2008
   5  * Phillip Lougher <phillip@squashfs.org.uk>
   6  *
   7  * This program is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU General Public License
   9  * as published by the Free Software Foundation; either version 2,
  10  * or (at your option) any later version.
  11  *
  12  * This program is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  * GNU General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU General Public License
  18  * along with this program; if not, write to the Free Software
  19  * Foundation, 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
  20  *
  21  * cache.c
  22  */
  23
  24 /*
  25  * Blocks in Squashfs are compressed.  To avoid repeatedly decompressing
  26  * recently accessed data Squashfs uses two small metadata and fragment caches.
  27  *
  28  * This file implements a generic cache implementation used for both caches,
  29  * plus functions layered ontop of the generic cache implementation to
  30  * access the metadata and fragment caches.
  31  *
  32  * To avoid out of memory and fragmentation issues with vmalloc the cache
  33  * uses sequences of kmalloced PAGE_CACHE_SIZE buffers.
  34  *
  35  * It should be noted that the cache is not used for file datablocks, these
  36  * are decompressed and cached in the page-cache in the normal way.  The
  37  * cache is only used to temporarily cache fragment and metadata blocks
  38  * which have been read as as a result of a metadata (i.e. inode or
  39  * directory) or fragment access.  Because metadata and fragments are packed
  40  * together into blocks (to gain greater compression) the read of a particular
  41  * piece of metadata or fragment will retrieve other metadata/fragments which
  42  * have been packed with it, these because of locality-of-reference may be read
  43  * in the near future. Temporarily caching them ensures they are available for
  44  * near future access without requiring an additional read and decompress.
  45  */
  46
  47 #include <linux/fs.h>
  48 #include <linux/vfs.h>
  49 #include <linux/slab.h>
  50 #include <linux/vmalloc.h>
  51 #include <linux/sched.h>
  52 #include <linux/spinlock.h>
  53 #include <linux/wait.h>
  54 #include <linux/pagemap.h>
  55
  56 #include "squashfs_fs.h"
  57 #include "squashfs_fs_sb.h"
  58 #include "squashfs.h"
  59 #include "page_actor.h"
  60
  61 /*
  62  * Look-up block in cache, and increment usage count.  If not in cache, read
  63  * and decompress it from disk.
  64  */
  65 struct squashfs_cache_entry *squashfs_cache_get(struct super_block *sb,
  66         struct squashfs_cache *cache, u64 block, int length)
  67 {
  68         int i, n;
  69         struct squashfs_cache_entry *entry;
  70
  71         spin_lock(&cache->lock);
  72
  73         while (1) {
  74                 for (i = cache->curr_blk, n = 0; n < cache->entries; n++) {
  75                         if (cache->entry[i].block == block) {
  76                                 cache->curr_blk = i;
  77                                 break;
  78                         }
  79                         i = (i + 1) % cache->entries;
  80                 }
  81
  82                 if (n == cache->entries) {
  83                         /*
  84                          * Block not in cache, if all cache entries are used
  85                          * go to sleep waiting for one to become available.
  86                          */
  87                         if (cache->unused == 0) {
  88                                 cache->num_waiters++;
  89                                 spin_unlock(&cache->lock);
  90                                 wait_event(cache->wait_queue, cache->unused);
  91                                 spin_lock(&cache->lock);
  92                                 cache->num_waiters--;
  93                                 continue;
  94                         }
  95
  96                         /*
  97                          * At least one unused cache entry.  A simple
  98                          * round-robin strategy is used to choose the entry to
  99                          * be evicted from the cache.
 100                          */
 101                         i = cache->next_blk;
 102                         for (n = 0; n < cache->entries; n++) {
 103                                 if (cache->entry[i].refcount == 0)
 104                                         break;
 105                                 i = (i + 1) % cache->entries;
 106                         }
 107
 108                         cache->next_blk = (i + 1) % cache->entries;
 109                         entry = &cache->entry[i];
 110
 111                         /*
 112                          * Initialise chosen cache entry, and fill it in from
 113                          * disk.
 114                          */
 115                         cache->unused--;
 116                         entry->block = block;
 117                         entry->refcount = 1;
 118                         entry->pending = 1;
 119                         entry->num_waiters = 0;
 120                         entry->error = 0;
 121                         spin_unlock(&cache->lock);
 122
 123                         entry->length = squashfs_read_data(sb, block, length,
 124                                 &entry->next_index, entry->actor);
 125
 126                         spin_lock(&cache->lock);
 127
 128                         if (entry->length < 0)
 129                                 entry->error = entry->length;
 130
 131                         entry->pending = 0;
 132
 133                         /*
 134                          * While filling this entry one or more other processes
 135                          * have looked it up in the cache, and have slept
 136                          * waiting for it to become available.
 137                          */
 138                         if (entry->num_waiters) {
 139                                 spin_unlock(&cache->lock);
 140                                 wake_up_all(&entry->wait_queue);
 141                         } else
 142                                 spin_unlock(&cache->lock);
 143
 144                         goto out;
 145                 }
 146
 147                 /*
 148                  * Block already in cache.  Increment refcount so it doesn't
 149                  * get reused until we're finished with it, if it was
 150                  * previously unused there's one less cache entry available
 151                  * for reuse.
 152                  */
 153                 entry = &cache->entry[i];
 154                 if (entry->refcount == 0)
 155                         cache->unused--;
 156                 entry->refcount++;
 157
 158                 /*
 159                  * If the entry is currently being filled in by another process
 160                  * go to sleep waiting for it to become available.
 161                  */
 162                 if (entry->pending) {
 163                         entry->num_waiters++;
 164                         spin_unlock(&cache->lock);
 165                         wait_event(entry->wait_queue, !entry->pending);
 166                 } else
 167                         spin_unlock(&cache->lock);
 168
 169                 goto out;
 170         }
 171
 172 out:
 173         TRACE("Got %s %d, start block %lld, refcount %d, error %d\n",
 174                 cache->name, i, entry->block, entry->refcount, entry->error);
 175
 176         if (entry->error)
 177                 ERROR("Unable to read %s cache entry [%llx]\n", cache->name,
 178                                                         block);
 179         return entry;
 180 }
 181
 182
 183 /*
 184  * Release cache entry, once usage count is zero it can be reused.
 185  */
 186 void squashfs_cache_put(struct squashfs_cache_entry *entry)
 187 {
 188         struct squashfs_cache *cache = entry->cache;
 189
 190         spin_lock(&cache->lock);
 191         entry->refcount--;
 192         if (entry->refcount == 0) {
 193                 cache->unused++;
 194                 /*
 195                  * If there's any processes waiting for a block to become
 196                  * available, wake one up.
 197                  */
 198                 if (cache->num_waiters) {
 199                         spin_unlock(&cache->lock);
 200                         wake_up(&cache->wait_queue);
 201                         return;
 202                 }
 203         }
 204         spin_unlock(&cache->lock);
 205 }
 206
 207 /*
 208  * Delete cache reclaiming all kmalloced buffers.
 209  */
 210 void squashfs_cache_delete(struct squashfs_cache *cache)
 211 {
 212         int i;
 213
 214         if (cache == NULL)
 215                 return;
 216
 217         for (i = 0; i < cache->entries; i++) {
 218                 if (cache->entry[i].page)
 219                         free_page_array(cache->entry[i].page, cache->pages);
 220                 kfree(cache->entry[i].actor);
 221         }
 222
 223         kfree(cache->entry);
 224         kfree(cache);
 225 }
 226
 227
 228 /*
 229  * Initialise cache allocating the specified number of entries, each of
 230  * size block_size.  To avoid vmalloc fragmentation issues each entry
 231  * is allocated as a sequence of kmalloced PAGE_CACHE_SIZE buffers.
 232  */
 233 struct squashfs_cache *squashfs_cache_init(char *name, int entries,
 234         int block_size)
 235 {
 236         int i;
 237         struct squashfs_cache *cache = kzalloc(sizeof(*cache), GFP_KERNEL);
 238
 239         if (cache == NULL) {
 240                 ERROR("Failed to allocate %s cache\n", name);
 241                 return NULL;
 242         }
 243
 244         cache->entry = kcalloc(entries, sizeof(*(cache->entry)), GFP_KERNEL);
 245         if (cache->entry == NULL) {
 246                 ERROR("Failed to allocate %s cache\n", name);
 247                 goto cleanup;
 248         }
 249
 250         cache->curr_blk = 0;
 251         cache->next_blk = 0;
 252         cache->unused = entries;
 253         cache->entries = entries;
 254         cache->block_size = block_size;
 255         cache->pages = block_size >> PAGE_CACHE_SHIFT;
 256         cache->pages = cache->pages ? cache->pages : 1;
 257         cache->name = name;
 258         cache->num_waiters = 0;
 259         spin_lock_init(&cache->lock);
 260         init_waitqueue_head(&cache->wait_queue);
 261
 262         for (i = 0; i < entries; i++) {
 263                 struct squashfs_cache_entry *entry = &cache->entry[i];
 264
 265                 init_waitqueue_head(&cache->entry[i].wait_queue);
 266                 entry->cache = cache;
 267                 entry->block = SQUASHFS_INVALID_BLK;
 268                 entry->page = alloc_page_array(cache->pages, GFP_KERNEL);
 269                 if (!entry->page) {
 270                         ERROR("Failed to allocate %s cache entry\n", name);
 271                         goto cleanup;
 272                 }
 273                 entry->actor = squashfs_page_actor_init(entry->page,
 274                         cache->pages, 0, NULL);
 275                 if (entry->actor == NULL) {
 276                         ERROR("Failed to allocate %s cache entry\n", name);
 277                         goto cleanup;
 278                 }
 279         }
 280
 281         return cache;
 282
 283 cleanup:
 284         squashfs_cache_delete(cache);
 285         return NULL;
 286 }
 287
 288
 289 /*
 290  * Copy up to length bytes from cache entry to buffer starting at offset bytes
 291  * into the cache entry.  If there's not length bytes then copy the number of
 292  * bytes available.  In all cases return the number of bytes copied.
 293  */
 294 int squashfs_copy_data(void *buffer, struct squashfs_cache_entry *entry,
 295                 int offset, int length)
 296 {
 297         int remaining = length;
 298
 299         if (length == 0)
 300                 return 0;
 301         else if (buffer == NULL)
 302                 return min(length, entry->length - offset);
 303
 304         while (offset < entry->length) {
 305                 void *buff = kmap_atomic(entry->page[offset / PAGE_CACHE_SIZE])
 306                              + (offset % PAGE_CACHE_SIZE);
 307                 int bytes = min_t(int, entry->length - offset,
 308                                 PAGE_CACHE_SIZE - (offset % PAGE_CACHE_SIZE));
 309
 310                 if (bytes >= remaining) {
 311                         memcpy(buffer, buff, remaining);
 312                         kunmap_atomic(buff);
 313                         remaining = 0;
 314                         break;
 315                 }
 316
 317                 memcpy(buffer, buff, bytes);
 318                 kunmap_atomic(buff);
 319                 buffer += bytes;
 320                 remaining -= bytes;
 321                 offset += bytes;
 322         }
 323
 324         return length - remaining;
 325 }
 326
 327
 328 /*
 329  * Read length bytes from metadata position <block, offset> (block is the
 330  * start of the compressed block on disk, and offset is the offset into
 331  * the block once decompressed).  Data is packed into consecutive blocks,
 332  * and length bytes may require reading more than one block.
 333  */
 334 int squashfs_read_metadata(struct super_block *sb, void *buffer,
 335                 u64 *block, int *offset, int length)
 336 {
 337         struct squashfs_sb_info *msblk = sb->s_fs_info;
 338         int bytes, res = length;
 339         struct squashfs_cache_entry *entry;
 340
 341         TRACE("Entered squashfs_read_metadata [%llx:%x]\n", *block, *offset);
 342
 343         while (length) {
 344                 entry = squashfs_cache_get(sb, msblk->block_cache, *block, 0);
 345                 if (entry->error) {
 346                         res = entry->error;
 347                         goto error;
 348                 } else if (*offset >= entry->length) {
 349                         res = -EIO;
 350                         goto error;
 351                 }
 352
 353                 bytes = squashfs_copy_data(buffer, entry, *offset, length);
 354                 if (buffer)
 355                         buffer += bytes;
 356                 length -= bytes;
 357                 *offset += bytes;
 358
 359                 if (*offset == entry->length) {
 360                         *block = entry->next_index;
 361                         *offset = 0;
 362                 }
 363
 364                 squashfs_cache_put(entry);
 365         }
 366
 367         return res;
 368
 369 error:
 370         squashfs_cache_put(entry);
 371         return res;
 372 }
 373
 374
 375 /*
 376  * Look-up in the fragmment cache the fragment located at <start_block> in the
 377  * filesystem.  If necessary read and decompress it from disk.
 378  */
 379 struct squashfs_cache_entry *squashfs_get_fragment(struct super_block *sb,
 380                                 u64 start_block, int length)
 381 {
 382         struct squashfs_sb_info *msblk = sb->s_fs_info;
 383
 384         return squashfs_cache_get(sb, msblk->fragment_cache, start_block,
 385                 length);
 386 }
 387
 388
 389 /*
 390  * Read and decompress the datablock located at <start_block> in the
 391  * filesystem.  The cache is used here to avoid duplicating locking and
 392  * read/decompress code.
 393  */
 394 struct squashfs_cache_entry *squashfs_get_datablock(struct super_block *sb,
 395                                 u64 start_block, int length)
 396 {
 397         struct squashfs_sb_info *msblk = sb->s_fs_info;
 398
 399         return squashfs_cache_get(sb, msblk->read_page, start_block, length);
 400 }
 401
 402
 403 /*
 404  * Read a filesystem table (uncompressed sequence of bytes) from disk
 405  */
 406 void *squashfs_read_table(struct super_block *sb, u64 block, int length)
 407 {
 408         int pages = (length + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 409         struct page **page;
 410         void *buff;
 411         int res;
 412         struct squashfs_page_actor *actor;
 413
 414         page = alloc_page_array(pages, GFP_KERNEL);
 415         if (!page)
 416                 return ERR_PTR(-ENOMEM);
 417
 418         actor = squashfs_page_actor_init(page, pages, length, NULL);
 419         if (actor == NULL) {
 420                 res = -ENOMEM;
 421                 goto failed;
 422         }
 423
 424         res = squashfs_read_data(sb, block, length |
 425                 SQUASHFS_COMPRESSED_BIT_BLOCK, NULL, actor);
 426
 427         if (res < 0)
 428                 goto failed2;
 429
 430         buff = kmalloc(length, GFP_KERNEL);
 431         if (!buff)
 432                 goto failed2;
 433         squashfs_actor_to_buf(actor, buff, length);
 434         squashfs_page_actor_free(actor, 0);
 435         free_page_array(page, pages);
 436         return buff;
 437
 438 failed2:
 439         squashfs_page_actor_free(actor, 0);
 440 failed:
 441         free_page_array(page, pages);
 442         return ERR_PTR(res);
 443 }