• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Copyright (c) 2006-2009 The Chromium Authors. All rights reserved.
2 // Use of this source code is governed by a BSD-style license that can be
3 // found in the LICENSE file.
4 
5 // The cache is stored on disk as a collection of block-files, plus an index
6 // file plus a collection of external files.
7 //
8 // Any data blob bigger than kMaxBlockSize (net/addr.h) will be stored on a
9 // separate file named f_xxx where x is a hexadecimal number. Shorter data will
10 // be stored as a series of blocks on a block-file. In any case, CacheAddr
11 // represents the address of the data inside the cache.
12 //
13 // The index file is just a simple hash table that maps a particular entry to
14 // a CacheAddr value. Linking for a given hash bucket is handled internally
15 // by the cache entry.
16 //
17 // The last element of the cache is the block-file. A block file is a file
18 // designed to store blocks of data of a given size. It is able to store data
19 // that spans from one to four consecutive "blocks", and it grows as needed to
20 // store up to approximately 65000 blocks. It has a fixed size header used for
21 // book keeping such as tracking free of blocks on the file. For example, a
22 // block-file for 1KB blocks will grow from 8KB when totally empty to about 64MB
23 // when completely full. At that point, data blocks of 1KB will be stored on a
24 // second block file that will store the next set of 65000 blocks. The first
25 // file contains the number of the second file, and the second file contains the
26 // number of a third file, created when the second file reaches its limit. It is
27 // important to remember that no matter how long the chain of files is, any
28 // given block can be located directly by its address, which contains the file
29 // number and starting block inside the file.
30 //
31 // A new cache is initialized with four block files (named data_0 through
32 // data_3), each one dedicated to store blocks of a given size. The number at
33 // the end of the file name is the block file number (in decimal).
34 //
35 // There are two "special" types of blocks: an entry and a rankings node. An
36 // entry keeps track of all the information related to the same cache entry,
37 // such as the key, hash value, data pointers etc. A rankings node keeps track
38 // of the information that is updated frequently for a given entry, such as its
39 // location on the LRU lists, last access time etc.
40 //
41 // The files that store internal information for the cache (blocks and index)
42 // are at least partially memory mapped. They have a location that is signaled
43 // every time the internal structures are modified, so it is possible to detect
44 // (most of the time) when the process dies in the middle of an update.
45 //
46 // In order to prevent dirty data to be used as valid (after a crash), every
47 // cache entry has a dirty identifier. Each running instance of the cache keeps
48 // a separate identifier (maintained on the "this_id" header field) that is used
49 // to mark every entry that is created or modified. When the entry is closed,
50 // and all the data can be trusted, the dirty flag is cleared from the entry.
51 // When the cache encounters an entry whose identifier is different than the one
52 // being currently used, it means that the entry was not properly closed on a
53 // previous run, so it is discarded.
54 
55 #ifndef NET_DISK_CACHE_DISK_FORMAT_H_
56 #define NET_DISK_CACHE_DISK_FORMAT_H_
57 
58 #include "base/basictypes.h"
59 
60 namespace disk_cache {
61 
62 typedef uint32 CacheAddr;
63 
64 const int kIndexTablesize = 0x10000;
65 const uint32 kIndexMagic = 0xC103CAC3;
66 const uint32 kCurrentVersion = 0x20000;  // Version 2.0.
67 
68 struct LruData {
69   int32     pad1[2];
70   int32     filled;          // Flag to tell when we filled the cache.
71   int32     sizes[5];
72   CacheAddr heads[5];
73   CacheAddr tails[5];
74   CacheAddr transaction;     // In-flight operation target.
75   int32     operation;       // Actual in-flight operation.
76   int32     operation_list;  // In-flight operation list.
77   int32     pad2[7];
78 };
79 
80 // Header for the master index file.
81 struct IndexHeader {
82   uint32      magic;
83   uint32      version;
84   int32       num_entries;   // Number of entries currently stored.
85   int32       num_bytes;     // Total size of the stored data.
86   int32       last_file;     // Last external file created.
87   int32       this_id;       // Id for all entries being changed (dirty flag).
88   CacheAddr   stats;         // Storage for usage data.
89   int32       table_len;     // Actual size of the table (0 == kIndexTablesize).
90   int32       crash;         // Signals a previous crash.
91   int32       experiment;    // Id of an ongoing test.
92   uint64      create_time;   // Creation time for this set of files.
93   int32       pad[52];
94   LruData     lru;           // Eviction control data.
IndexHeaderIndexHeader95   IndexHeader() {
96     memset(this, 0, sizeof(*this));
97     magic = kIndexMagic;
98     version = kCurrentVersion;
99   };
100 };
101 
102 // The structure of the whole index file.
103 struct Index {
104   IndexHeader header;
105   CacheAddr   table[kIndexTablesize];  // Default size. Actual size controlled
106                                        // by header.table_len.
107 };
108 
109 // Main structure for an entry on the backing storage. If the key is longer than
110 // what can be stored on this structure, it will be extended on consecutive
111 // blocks (adding 256 bytes each time), up to 4 blocks (1024 - 32 - 1 chars).
112 // After that point, the whole key will be stored as a data block or external
113 // file.
114 struct EntryStore {
115   uint32      hash;               // Full hash of the key.
116   CacheAddr   next;               // Next entry with the same hash or bucket.
117   CacheAddr   rankings_node;      // Rankings node for this entry.
118   int32       reuse_count;        // How often is this entry used.
119   int32       refetch_count;      // How often is this fetched from the net.
120   int32       state;              // Current state.
121   uint64      creation_time;
122   int32       key_len;
123   CacheAddr   long_key;           // Optional address of a long key.
124   int32       data_size[4];       // We can store up to 4 data streams for each
125   CacheAddr   data_addr[4];       // entry.
126   uint32      flags;              // Any combination of EntryFlags.
127   int32       pad[5];
128   char        key[256 - 24 * 4];  // null terminated
129 };
130 
131 COMPILE_ASSERT(sizeof(EntryStore) == 256, bad_EntyStore);
132 const int kMaxInternalKeyLength = 4 * sizeof(EntryStore) -
133                                   offsetof(EntryStore, key) - 1;
134 
135 // Possible states for a given entry.
136 enum EntryState {
137   ENTRY_NORMAL = 0,
138   ENTRY_EVICTED,    // The entry was recently evicted from the cache.
139   ENTRY_DOOMED      // The entry was doomed.
140 };
141 
142 // Flags that can be applied to an entry.
143 enum EntryFlags {
144   PARENT_ENTRY = 1,         // This entry has children (sparse) entries.
145   CHILD_ENTRY = 1 << 1      // Child entry that stores sparse data.
146 };
147 
148 #pragma pack(push, 4)
149 // Rankings information for a given entry.
150 struct RankingsNode {
151   uint64      last_used;        // LRU info.
152   uint64      last_modified;    // LRU info.
153   CacheAddr   next;             // LRU list.
154   CacheAddr   prev;             // LRU list.
155   CacheAddr   contents;         // Address of the EntryStore.
156   int32       dirty;            // The entry is being modifyied.
157   int32       dummy;            // Old files may have a pointer here.
158 };
159 #pragma pack(pop)
160 
161 COMPILE_ASSERT(sizeof(RankingsNode) == 36, bad_RankingsNode);
162 
163 const uint32 kBlockMagic = 0xC104CAC3;
164 const int kBlockHeaderSize = 8192;  // Two pages: almost 64k entries
165 const int kMaxBlocks = (kBlockHeaderSize - 80) * 8;
166 
167 // Bitmap to track used blocks on a block-file.
168 typedef uint32 AllocBitmap[kMaxBlocks / 32];
169 
170 // A block-file is the file used to store information in blocks (could be
171 // EntryStore blocks, RankingsNode blocks or user-data blocks).
172 // We store entries that can expand for up to 4 consecutive blocks, and keep
173 // counters of the number of blocks available for each type of entry. For
174 // instance, an entry of 3 blocks is an entry of type 3. We also keep track of
175 // where did we find the last entry of that type (to avoid searching the bitmap
176 // from the beginning every time).
177 // This Structure is the header of a block-file:
178 struct BlockFileHeader {
179   uint32          magic;
180   uint32          version;
181   int16           this_file;    // Index of this file.
182   int16           next_file;    // Next file when this one is full.
183   int32           entry_size;   // Size of the blocks of this file.
184   int32           num_entries;  // Number of stored entries.
185   int32           max_entries;  // Current maximum number of entries.
186   int32           empty[4];     // Counters of empty entries for each type.
187   int32           hints[4];     // Last used position for each entry type.
188   volatile int32  updating;     // Keep track of updates to the header.
189   int32           user[5];
190   AllocBitmap     allocation_map;
BlockFileHeaderBlockFileHeader191   BlockFileHeader() {
192     memset(this, 0, sizeof(BlockFileHeader));
193     magic = kBlockMagic;
194     version = kCurrentVersion;
195   };
196 };
197 
198 COMPILE_ASSERT(sizeof(BlockFileHeader) == kBlockHeaderSize, bad_header);
199 
200 // Sparse data support:
201 // We keep a two level hierarchy to enable sparse data for an entry: the first
202 // level consists of using separate "child" entries to store ranges of 1 MB,
203 // and the second level stores blocks of 1 KB inside each child entry.
204 //
205 // Whenever we need to access a particular sparse offset, we first locate the
206 // child entry that stores that offset, so we discard the 20 least significant
207 // bits of the offset, and end up with the child id. For instance, the child id
208 // to store the first megabyte is 0, and the child that should store offset
209 // 0x410000 has an id of 4.
210 //
211 // The child entry is stored the same way as any other entry, so it also has a
212 // name (key). The key includes a signature to be able to identify children
213 // created for different generations of the same resource. In other words, given
214 // that a given sparse entry can have a large number of child entries, and the
215 // resource can be invalidated and replaced with a new version at any time, it
216 // is important to be sure that a given child actually belongs to certain entry.
217 //
218 // The full name of a child entry is composed with a prefix ("Range_"), and two
219 // hexadecimal 64-bit numbers at the end, separated by semicolons. The first
220 // number is the signature of the parent key, and the second number is the child
221 // id as described previously. The signature itself is also stored internally by
222 // the child and the parent entries. For example, a sparse entry with a key of
223 // "sparse entry name", and a signature of 0x052AF76, may have a child entry
224 // named "Range_sparse entry name:052af76:4", which stores data in the range
225 // 0x400000 to 0x4FFFFF.
226 //
227 // Each child entry keeps track of all the 1 KB blocks that have been written
228 // to the entry, but being a regular entry, it will happily return zeros for any
229 // read that spans data not written before. The actual sparse data is stored in
230 // one of the data streams of the child entry (at index 1), while the control
231 // information is stored in another stream (at index 2), both by parents and
232 // the children.
233 
234 // This structure contains the control information for parent and child entries.
235 // It is stored at offset 0 of the data stream with index 2.
236 // It is possible to write to a child entry in a way that causes the last block
237 // to be only partialy filled. In that case, last_block and last_block_len will
238 // keep track of that block.
239 struct SparseHeader {
240   int64 signature;          // The parent and children signature.
241   uint32 magic;             // Structure identifier (equal to kIndexMagic).
242   int32 parent_key_len;     // Key length for the parent entry.
243   int32 last_block;         // Index of the last written block.
244   int32 last_block_len;     // Lenght of the last written block.
245   int32 dummy[10];
246 };
247 
248 // The SparseHeader will be followed by a bitmap, as described by this
249 // structure.
250 struct SparseData {
251   SparseHeader header;
252   uint32 bitmap[32];        // Bitmap representation of known children (if this
253                             // is a parent entry), or used blocks (for child
254                             // entries. The size is fixed for child entries but
255                             // not for parents; it can be as small as 4 bytes
256                             // and as large as 8 KB.
257 };
258 
259 // The number of blocks stored by a child entry.
260 const int kNumSparseBits = 1024;
261 COMPILE_ASSERT(sizeof(SparseData) == sizeof(SparseHeader) + kNumSparseBits / 8,
262                Invalid_SparseData_bitmap);
263 
264 }  // namespace disk_cache
265 
266 #endif  // NET_DISK_CACHE_DISK_FORMAT_H_
267