dumbdbm.py revision 0320464583957258031ebee1b8897343211e4339
1"""A dumb and slow but simple dbm clone. 2 3For database spam, spam.dir contains the index (a text file), 4spam.bak *may* contain a backup of the index (also a text file), 5while spam.dat contains the data (a binary file). 6 7XXX TO DO: 8 9- seems to contain a bug when updating... 10 11- reclaim free space (currently, space once occupied by deleted or expanded 12items is never reused) 13 14- support concurrent access (currently, if two processes take turns making 15updates, they can mess up the index) 16 17- support efficient access to large databases (currently, the whole index 18is read when the database is opened, and some updates rewrite the whole index) 19 20- support opening for read-only (flag = 'm') 21 22""" 23 24import os as _os 25import __builtin__ 26import UserDict 27 28_open = __builtin__.open 29 30_BLOCKSIZE = 512 31 32error = IOError # For anydbm 33 34class _Database(UserDict.DictMixin): 35 36 # The on-disk directory and data files can remain in mutually 37 # inconsistent states for an arbitrarily long time (see comments 38 # at the end of __setitem__). This is only repaired when _commit() 39 # gets called. One place _commit() gets called is from __del__(), 40 # and if that occurs at program shutdown time, module globals may 41 # already have gotten rebound to None. Since it's crucial that 42 # _commit() finish successfully, we can't ignore shutdown races 43 # here, and _commit() must not reference any globals. 44 _os = _os # for _commit() 45 _open = _open # for _commit() 46 47 def __init__(self, filebasename, mode): 48 self._mode = mode 49 50 # The directory file is a text file. Each line looks like 51 # "%r, (%d, %d)\n" % (key, pos, siz) 52 # where key is the string key, pos is the offset into the dat 53 # file of the associated value's first byte, and siz is the number 54 # of bytes in the associated value. 55 self._dirfile = filebasename + _os.extsep + 'dir' 56 57 # The data file is a binary file pointed into by the directory 58 # file, and holds the values associated with keys. Each value 59 # begins at a _BLOCKSIZE-aligned byte offset, and is a raw 60 # binary 8-bit string value. 61 self._datfile = filebasename + _os.extsep + 'dat' 62 self._bakfile = filebasename + _os.extsep + 'bak' 63 64 # The index is an in-memory dict, mirroring the directory file. 65 self._index = None # maps keys to (pos, siz) pairs 66 67 # Mod by Jack: create data file if needed 68 try: 69 f = _open(self._datfile, 'r') 70 except IOError: 71 f = _open(self._datfile, 'w', self._mode) 72 f.close() 73 self._update() 74 75 # Read directory file into the in-memory index dict. 76 def _update(self): 77 self._index = {} 78 try: 79 f = _open(self._dirfile) 80 except IOError: 81 pass 82 else: 83 for line in f: 84 key, pos_and_siz_pair = eval(line) 85 self._index[key] = pos_and_siz_pair 86 f.close() 87 88 # Write the index dict to the directory file. The original directory 89 # file (if any) is renamed with a .bak extension first. If a .bak 90 # file currently exists, it's deleted. 91 def _commit(self): 92 # CAUTION: It's vital that _commit() succeed, and _commit() can 93 # be called from __del__(). Therefore we must never reference a 94 # global in this routine. 95 try: 96 self._os.unlink(self._bakfile) 97 except self._os.error: 98 pass 99 100 try: 101 self._os.rename(self._dirfile, self._bakfile) 102 except self._os.error: 103 pass 104 105 f = self._open(self._dirfile, 'w', self._mode) 106 for key, pos_and_siz_pair in self._index.iteritems(): 107 f.write("%r, %r\n" % (key, pos_and_siz_pair)) 108 f.close() 109 110 def __getitem__(self, key): 111 pos, siz = self._index[key] # may raise KeyError 112 f = _open(self._datfile, 'rb') 113 f.seek(pos) 114 dat = f.read(siz) 115 f.close() 116 return dat 117 118 # Append val to the data file, starting at a _BLOCKSIZE-aligned 119 # offset. The data file is first padded with NUL bytes (if needed) 120 # to get to an aligned offset. Return pair 121 # (starting offset of val, len(val)) 122 def _addval(self, val): 123 f = _open(self._datfile, 'rb+') 124 f.seek(0, 2) 125 pos = int(f.tell()) 126 npos = ((pos + _BLOCKSIZE - 1) // _BLOCKSIZE) * _BLOCKSIZE 127 f.write('\0'*(npos-pos)) 128 pos = npos 129 f.write(val) 130 f.close() 131 return (pos, len(val)) 132 133 # Write val to the data file, starting at offset pos. The caller 134 # is responsible for ensuring that there's enough room starting at 135 # pos to hold val, without overwriting some other value. Return 136 # pair (pos, len(val)). 137 def _setval(self, pos, val): 138 f = _open(self._datfile, 'rb+') 139 f.seek(pos) 140 f.write(val) 141 f.close() 142 return (pos, len(val)) 143 144 # key is a new key whose associated value starts in the data file 145 # at offset pos and with length siz. Add an index record to 146 # the in-memory index dict, and append one to the directory file. 147 def _addkey(self, key, pos_and_siz_pair): 148 self._index[key] = pos_and_siz_pair 149 f = _open(self._dirfile, 'a', self._mode) 150 f.write("%r, %r\n" % (key, pos_and_siz_pair)) 151 f.close() 152 153 def __setitem__(self, key, val): 154 if not type(key) == type('') == type(val): 155 raise TypeError, "keys and values must be strings" 156 if key not in self._index: 157 self._addkey(key, self._addval(val)) 158 else: 159 # See whether the new value is small enough to fit in the 160 # (padded) space currently occupied by the old value. 161 pos, siz = self._index[key] 162 oldblocks = (siz + _BLOCKSIZE - 1) // _BLOCKSIZE 163 newblocks = (len(val) + _BLOCKSIZE - 1) // _BLOCKSIZE 164 if newblocks <= oldblocks: 165 self._index[key] = self._setval(pos, val) 166 else: 167 # The new value doesn't fit in the (padded) space used 168 # by the old value. The blocks used by the old value are 169 # forever lost. 170 self._index[key] = self._addval(val) 171 172 # Note that _index may be out of synch with the directory 173 # file now: _setval() and _addval() don't update the directory 174 # file. This also means that the on-disk directory and data 175 # files are in a mutually inconsistent state, and they'll 176 # remain that way until _commit() is called. Note that this 177 # is a disaster (for the database) if the program crashes 178 # (so that _commit() never gets called). 179 180 def __delitem__(self, key): 181 # The blocks used by the associated value are lost. 182 del self._index[key] 183 # XXX It's unclear why we do a _commit() here (the code always 184 # XXX has, so I'm not changing it). _setitem__ doesn't try to 185 # XXX keep the directory file in synch. Why should we? Or 186 # XXX why shouldn't __setitem__? 187 self._commit() 188 189 def keys(self): 190 return self._index.keys() 191 192 def has_key(self, key): 193 return key in self._index 194 195 def __contains__(self, key): 196 return key in self._index 197 198 def iterkeys(self): 199 return self._index.iterkeys() 200 __iter__ = iterkeys 201 202 def __len__(self): 203 return len(self._index) 204 205 def close(self): 206 self._commit() 207 self._index = None 208 self._datfile = self._dirfile = self._bakfile = None 209 210 def __del__(self): 211 if self._index is not None: 212 self._commit() 213 214 215 216def open(file, flag=None, mode=0666): 217 """Open the database file, filename, and return corresponding object. 218 219 The flag argument, used to control how the database is opened in the 220 other DBM implementations, is ignored in the dumbdbm module; the 221 database is always opened for update, and will be created if it does 222 not exist. 223 224 The optional mode argument is the UNIX mode of the file, used only when 225 the database has to be created. It defaults to octal code 0666 (and 226 will be modified by the prevailing umask). 227 228 """ 229 # flag argument is currently ignored 230 return _Database(file, mode) 231