1
# Copyright (C) 2005, 2006 by Canonical Ltd
2
# Written by Robert Collins <robert.collins@canonical.com>
4
# This program is free software; you can redistribute it and/or modify
5
# it under the terms of the GNU General Public License as published by
6
# the Free Software Foundation; either version 2 of the License, or
7
# (at your option) any later version.
9
# This program is distributed in the hope that it will be useful,
10
# but WITHOUT ANY WARRANTY; without even the implied warranty of
11
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12
# GNU General Public License for more details.
14
# You should have received a copy of the GNU General Public License
15
# along with this program; if not, write to the Free Software
16
# Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
18
"""Bzrlib specific gzip tunings. We plan to feed these to the upstream gzip."""
20
# make GzipFile faster:
22
from gzip import U32, LOWU32, FEXTRA, FCOMMENT, FNAME, FHCRC
27
# we want a \n preserved, break on \n only splitlines.
30
__all__ = ["GzipFile"]
33
class GzipFile(gzip.GzipFile):
34
"""Knit tuned version of GzipFile.
36
This is based on the following lsprof stats:
37
python 2.4 stock GzipFile write:
38
58971 0 5644.3090 2721.4730 gzip:193(write)
39
+58971 0 1159.5530 1159.5530 +<built-in method compress>
40
+176913 0 987.0320 987.0320 +<len>
41
+58971 0 423.1450 423.1450 +<zlib.crc32>
42
+58971 0 353.1060 353.1060 +<method 'write' of 'cStringIO.
45
58971 0 4477.2590 2103.1120 bzrlib.knit:1250(write)
46
+58971 0 1297.7620 1297.7620 +<built-in method compress>
47
+58971 0 406.2160 406.2160 +<zlib.crc32>
48
+58971 0 341.9020 341.9020 +<method 'write' of 'cStringIO.
50
+58971 0 328.2670 328.2670 +<len>
53
Yes, its only 1.6 seconds, but they add up.
56
def _add_read_data(self, data):
58
# temp var for len(data) and switch to +='s.
61
self.crc = zlib.crc32(data, self.crc)
63
self.extrasize += len_data
66
def _read(self, size=1024):
67
# various optimisations:
68
# reduces lsprof count from 2500 to
69
# 8337 calls in 1272, 365 internal
70
if self.fileobj is None:
71
raise EOFError, "Reached EOF"
74
# If the _new_member flag is set, we have to
75
# jump to the next member, if there is one.
77
# First, check if we're at the end of the file;
78
# if so, it's time to stop; no more members to read.
79
next_header_bytes = self.fileobj.read(10)
80
if next_header_bytes == '':
81
raise EOFError, "Reached EOF"
84
self._read_gzip_header(next_header_bytes)
85
self.decompress = zlib.decompressobj(-zlib.MAX_WBITS)
86
self._new_member = False
88
# Read a chunk of data from the file
89
buf = self.fileobj.read(size)
91
# If the EOF has been reached, flush the decompression object
92
# and mark this object as finished.
95
self._add_read_data(self.decompress.flush())
96
assert len(self.decompress.unused_data) >= 8, "what does flush do?"
98
# tell the driving read() call we have stuffed all the data
100
raise EOFError, 'Reached EOF'
102
self._add_read_data(self.decompress.decompress(buf))
104
if self.decompress.unused_data != "":
105
# Ending case: we've come to the end of a member in the file,
106
# so seek back to the start of the data for the next member which
107
# is the length of the decompress objects unused data - the first
108
# 8 bytes for the end crc and size records.
110
# so seek back to the start of the unused data, finish up
111
# this member, and read a new gzip header.
112
# (The number of bytes to seek back is the length of the unused
113
# data, minus 8 because those 8 bytes are part of this member.
114
seek_length = len (self.decompress.unused_data) - 8
116
# we read too much data
117
self.fileobj.seek(-seek_length, 1)
118
elif seek_length < 0:
119
# we haven't read enough to check the checksum.
120
assert -8 < seek_length, "too great a seek."
121
buf = self.fileobj.read(-seek_length)
122
self.decompress.decompress(buf)
124
# Check the CRC and file size, and set the flag so we read
125
# a new member on the next call
127
self._new_member = True
130
"""tuned to reduce function calls and eliminate file seeking:
132
reduces lsprof count from 800 to 288
134
avoid U32 call by using struct format L
137
# We've read to the end of the file, so we should have 8 bytes of
138
# unused data in the decompressor. If we dont, there is a corrupt file.
139
# We use these 8 bytes to calculate the CRC and the recorded file size.
140
# We then check the that the computed CRC and size of the
141
# uncompressed data matches the stored values. Note that the size
142
# stored is the true file size mod 2**32.
143
crc32, isize = struct.unpack("<LL", self.decompress.unused_data[0:8])
144
# note that isize is unsigned - it can exceed 2GB
145
if crc32 != U32(self.crc):
146
raise IOError, "CRC check failed %d %d" % (crc32, U32(self.crc))
147
elif isize != LOWU32(self.size):
148
raise IOError, "Incorrect length of data produced"
150
def _read_gzip_header(self, bytes=None):
151
"""Supply bytes if the minimum header size is already read.
153
:param bytes: 10 bytes of header data.
155
"""starting cost: 300 in 3998
156
15998 reads from 3998 calls
160
bytes = self.fileobj.read(10)
162
if magic != '\037\213':
163
raise IOError, 'Not a gzipped file'
164
method = ord(bytes[2:3])
166
raise IOError, 'Unknown compression method'
167
flag = ord(bytes[3:4])
168
# modtime = self.fileobj.read(4) (bytes [4:8])
169
# extraflag = self.fileobj.read(1) (bytes[8:9])
170
# os = self.fileobj.read(1) (bytes[9:10])
171
# self.fileobj.read(6)
174
# Read & discard the extra field, if present
175
xlen = ord(self.fileobj.read(1))
176
xlen = xlen + 256*ord(self.fileobj.read(1))
177
self.fileobj.read(xlen)
179
# Read and discard a null-terminated string containing the filename
181
s = self.fileobj.read(1)
182
if not s or s=='\000':
185
# Read and discard a null-terminated string containing a comment
187
s = self.fileobj.read(1)
188
if not s or s=='\000':
191
self.fileobj.read(2) # Read & discard the 16-bit header CRC
193
def readline(self, size=-1):
194
"""Tuned to remove buffer length calls in _unread and...
196
also removes multiple len(c) calls, inlines _unread,
197
total savings - lsprof 5800 to 5300
200
8176 calls to read() in 1684
201
changing the min chunk size to 200 halved all the cache misses
202
leading to a drop to:
204
4168 call to read() in 1646
205
- i.e. just reduced the function call overhead. May be worth
208
if size < 0: size = sys.maxint
210
readsize = min(200, size) # Read from the file in small chunks
213
return "".join(bufs) # Return resulting line
216
c = self.read(readsize)
217
# number of bytes read
221
# We set i=size to break out of the loop under two
222
# conditions: 1) there's no newline, and the chunk is
223
# larger than size, or 2) there is a newline, but the
224
# resulting line would be longer than 'size'.
225
if i==-1 and len_c > size: i=size-1
226
elif size <= i: i = size -1
228
if i >= 0 or c == '':
229
# if i>= 0 we have a newline or have triggered the above
230
# if size is not None condition.
231
# if c == '' its EOF.
232
bufs.append(c[:i+1]) # Add portion of last chunk
233
# -- inlined self._unread --
234
## self._unread(c[i+1:], len_c - i) # Push back rest of chunk
235
self.extrabuf = c[i+1:] + self.extrabuf
236
self.extrasize = len_c - i + self.extrasize
237
self.offset -= len_c - i
238
# -- end inlined self._unread --
239
return ''.join(bufs) # Return resulting line
241
# Append chunk to list, decrease 'size',
244
readsize = min(size, readsize * 2)
246
def readlines(self, sizehint=0):
247
# optimise to avoid all the buffer manipulation
248
# lsprof changed from:
249
# 4168 calls in 5472 with 32000 calls to readline()
252
# Negative numbers result in reading all the lines
255
content = self.read(sizehint)
256
return bzrlib.osutils.split_lines(content)
258
def _unread(self, buf, len_buf=None):
259
"""tuned to remove unneeded len calls.
261
because this is such an inner routine in readline, and readline is
262
in many inner loops, this has been inlined into readline().
264
The len_buf parameter combined with the reduction in len calls dropped
265
the lsprof ms count for this routine on my test data from 800 to 200 -
270
self.extrabuf = buf + self.extrabuf
271
self.extrasize = len_buf + self.extrasize
272
self.offset -= len_buf
274
def write(self, data):
275
if self.mode != gzip.WRITE:
277
raise IOError(errno.EBADF, "write() on read-only GzipFile object")
279
if self.fileobj is None:
280
raise ValueError, "write() on closed GzipFile object"
283
self.size = self.size + data_len
284
self.crc = zlib.crc32(data, self.crc)
285
self.fileobj.write( self.compress.compress(data) )
286
self.offset += data_len
288
def writelines(self, lines):
289
# profiling indicated a significant overhead
290
# calling write for each line.
291
# this batch call is a lot faster :).
292
# (4 seconds to 1 seconds for the sample upgrades I was testing).
293
self.write(''.join(lines))
added
removed
bzrlib/tuned_gzip.py
