Home | History | Annotate | Download | only in blockfile
      1 // Copyright (c) 2011 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 // For a general description of the files used by the cache see file_format.h.
      6 //
      7 // A block file is a file designed to store blocks of data of a given size. It
      8 // is able to store data that spans from one to four consecutive "blocks", and
      9 // it grows as needed to store up to approximately 65000 blocks. It has a fixed
     10 // size header used for book keeping such as tracking free of blocks on the
     11 // file. For example, a block-file for 1KB blocks will grow from 8KB when
     12 // totally empty to about 64MB when completely full. At that point, data blocks
     13 // of 1KB will be stored on a second block file that will store the next set of
     14 // 65000 blocks. The first file contains the number of the second file, and the
     15 // second file contains the number of a third file, created when the second file
     16 // reaches its limit. It is important to remember that no matter how long the
     17 // chain of files is, any given block can be located directly by its address,
     18 // which contains the file number and starting block inside the file.
     19 
     20 #ifndef NET_DISK_CACHE_BLOCKFILE_DISK_FORMAT_BASE_H_
     21 #define NET_DISK_CACHE_BLOCKFILE_DISK_FORMAT_BASE_H_
     22 
     23 #include "base/basictypes.h"
     24 #include "net/base/net_export.h"
     25 
     26 namespace disk_cache {
     27 
     28 typedef uint32 CacheAddr;
     29 
     30 const uint32 kBlockVersion2 = 0x20000;  // Version 2.0.
     31 const uint32 kBlockCurrentVersion = 0x30000;  // Version 3.0.
     32 
     33 const uint32 kBlockMagic = 0xC104CAC3;
     34 const int kBlockHeaderSize = 8192;  // Two pages: almost 64k entries
     35 const int kMaxBlocks = (kBlockHeaderSize - 80) * 8;
     36 const int kNumExtraBlocks = 1024;  // How fast files grow.
     37 
     38 // Bitmap to track used blocks on a block-file.
     39 typedef uint32 AllocBitmap[kMaxBlocks / 32];
     40 
     41 // A block-file is the file used to store information in blocks (could be
     42 // EntryStore blocks, RankingsNode blocks or user-data blocks).
     43 // We store entries that can expand for up to 4 consecutive blocks, and keep
     44 // counters of the number of blocks available for each type of entry. For
     45 // instance, an entry of 3 blocks is an entry of type 3. We also keep track of
     46 // where did we find the last entry of that type (to avoid searching the bitmap
     47 // from the beginning every time).
     48 // This Structure is the header of a block-file:
     49 struct BlockFileHeader {
     50   uint32          magic;
     51   uint32          version;
     52   int16           this_file;    // Index of this file.
     53   int16           next_file;    // Next file when this one is full.
     54   int32           entry_size;   // Size of the blocks of this file.
     55   int32           num_entries;  // Number of stored entries.
     56   int32           max_entries;  // Current maximum number of entries.
     57   int32           empty[4];     // Counters of empty entries for each type.
     58   int32           hints[4];     // Last used position for each entry type.
     59   volatile int32  updating;     // Keep track of updates to the header.
     60   int32           user[5];
     61   AllocBitmap     allocation_map;
     62 };
     63 
     64 COMPILE_ASSERT(sizeof(BlockFileHeader) == kBlockHeaderSize, bad_header);
     65 
     66 // Sparse data support:
     67 // We keep a two level hierarchy to enable sparse data for an entry: the first
     68 // level consists of using separate "child" entries to store ranges of 1 MB,
     69 // and the second level stores blocks of 1 KB inside each child entry.
     70 //
     71 // Whenever we need to access a particular sparse offset, we first locate the
     72 // child entry that stores that offset, so we discard the 20 least significant
     73 // bits of the offset, and end up with the child id. For instance, the child id
     74 // to store the first megabyte is 0, and the child that should store offset
     75 // 0x410000 has an id of 4.
     76 //
     77 // The child entry is stored the same way as any other entry, so it also has a
     78 // name (key). The key includes a signature to be able to identify children
     79 // created for different generations of the same resource. In other words, given
     80 // that a given sparse entry can have a large number of child entries, and the
     81 // resource can be invalidated and replaced with a new version at any time, it
     82 // is important to be sure that a given child actually belongs to certain entry.
     83 //
     84 // The full name of a child entry is composed with a prefix ("Range_"), and two
     85 // hexadecimal 64-bit numbers at the end, separated by semicolons. The first
     86 // number is the signature of the parent key, and the second number is the child
     87 // id as described previously. The signature itself is also stored internally by
     88 // the child and the parent entries. For example, a sparse entry with a key of
     89 // "sparse entry name", and a signature of 0x052AF76, may have a child entry
     90 // named "Range_sparse entry name:052af76:4", which stores data in the range
     91 // 0x400000 to 0x4FFFFF.
     92 //
     93 // Each child entry keeps track of all the 1 KB blocks that have been written
     94 // to the entry, but being a regular entry, it will happily return zeros for any
     95 // read that spans data not written before. The actual sparse data is stored in
     96 // one of the data streams of the child entry (at index 1), while the control
     97 // information is stored in another stream (at index 2), both by parents and
     98 // the children.
     99 
    100 // This structure contains the control information for parent and child entries.
    101 // It is stored at offset 0 of the data stream with index 2.
    102 // It is possible to write to a child entry in a way that causes the last block
    103 // to be only partialy filled. In that case, last_block and last_block_len will
    104 // keep track of that block.
    105 struct SparseHeader {
    106   int64 signature;          // The parent and children signature.
    107   uint32 magic;             // Structure identifier (equal to kIndexMagic).
    108   int32 parent_key_len;     // Key length for the parent entry.
    109   int32 last_block;         // Index of the last written block.
    110   int32 last_block_len;     // Lenght of the last written block.
    111   int32 dummy[10];
    112 };
    113 
    114 // The SparseHeader will be followed by a bitmap, as described by this
    115 // structure.
    116 struct SparseData {
    117   SparseHeader header;
    118   uint32 bitmap[32];        // Bitmap representation of known children (if this
    119                             // is a parent entry), or used blocks (for child
    120                             // entries. The size is fixed for child entries but
    121                             // not for parents; it can be as small as 4 bytes
    122                             // and as large as 8 KB.
    123 };
    124 
    125 // The number of blocks stored by a child entry.
    126 const int kNumSparseBits = 1024;
    127 COMPILE_ASSERT(sizeof(SparseData) == sizeof(SparseHeader) + kNumSparseBits / 8,
    128                Invalid_SparseData_bitmap);
    129 
    130 }  // namespace disk_cache
    131 
    132 #endif  // NET_DISK_CACHE_BLOCKFILE_DISK_FORMAT_BASE_H_
    133