From f25c0d11e1b640e3c7e0addb66a1ff50730be508 Mon Sep 17 00:00:00 2001 From: Illia Volochii Date: Fri, 5 Dec 2025 16:40:41 +0200 Subject: [PATCH] Merge commit from fork * Prevent decompression bomb for zstd in Python 3.14 * Add experimental `decompress_iter` for Brotli * Update changes for Brotli * Add `GzipDecoder.decompress_iter` * Test https://github.com/python-hyper/brotlicffi/pull/207 * Pin Brotli * Add `decompress_iter` to all decoders and make tests pass * Pin brotlicffi to an official release * Revert changes to response.py * Add `max_length` parameter to all `decompress` methods * Fix the `test_brotlipy` session * Unset `_data` on gzip error * Add a test for memory usage * Test more methods * Fix the test for `stream` * Cover more lines with tests * Add more coverage * Make `read1` a bit more efficient * Fix PyPy tests for Brotli * Revert an unnecessarily moved check * Add some comments * Leave just one `self._obj.decompress` call in `GzipDecoder` * Refactor test params * Test reads with all data already in the decompressor * Prevent needless copying of data decoded with `max_length` * Rename the changed test * Note that responses of unknown length should be streamed too * Add a changelog entry * Avoid returning a memory view from `BytesQueueBuffer` * Add one more note to the changelog entry CVE: CVE-2025-66471 Upstream-Status: Backport [https://github.com/urllib3/urllib3/commit/c19571de34c47de3a766541b041637ba5f716ed7] Signed-off-by: Jiaying Song --- docs/advanced-usage.rst | 3 +- docs/user-guide.rst | 4 +- pyproject.toml | 5 +- src/urllib3/response.py | 278 ++++++++++++++++++++++++++++++++++------ 4 files changed, 246 insertions(+), 44 deletions(-) diff --git a/docs/advanced-usage.rst b/docs/advanced-usage.rst index 36a51e67..a12c7143 100644 --- a/docs/advanced-usage.rst +++ b/docs/advanced-usage.rst @@ -66,7 +66,8 @@ When using ``preload_content=True`` (the default setting) the response body will be read immediately into memory and the HTTP connection will be released back into the pool without manual intervention. -However, when dealing with large responses it's often better to stream the response +However, when dealing with responses of large or unknown length, +it's often better to stream the response content using ``preload_content=False``. Setting ``preload_content`` to ``False`` means that urllib3 will only read from the socket when data is requested. diff --git a/docs/user-guide.rst b/docs/user-guide.rst index 5c78c8af..1d9d0bbd 100644 --- a/docs/user-guide.rst +++ b/docs/user-guide.rst @@ -145,8 +145,8 @@ to a byte string representing the response content: print(resp.data) # b"\xaa\xa5H?\x95\xe9\x9b\x11" -.. note:: For larger responses, it's sometimes better to :ref:`stream ` - the response. +.. note:: For responses of large or unknown length, it's sometimes better to + :ref:`stream ` the response. Using io Wrappers with Response Content ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ diff --git a/pyproject.toml b/pyproject.toml index 1fe82937..58a2c2db 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -40,8 +40,8 @@ dynamic = ["version"] [project.optional-dependencies] brotli = [ - "brotli>=1.0.9; platform_python_implementation == 'CPython'", - "brotlicffi>=0.8.0; platform_python_implementation != 'CPython'" + "brotli>=1.2.0; platform_python_implementation == 'CPython'", + "brotlicffi>=1.2.0.0; platform_python_implementation != 'CPython'" ] zstd = [ "zstandard>=0.18.0", @@ -95,6 +95,7 @@ filterwarnings = [ '''default:ssl\.PROTOCOL_TLSv1_1 is deprecated:DeprecationWarning''', '''default:ssl\.PROTOCOL_TLSv1_2 is deprecated:DeprecationWarning''', '''default:ssl NPN is deprecated, use ALPN instead:DeprecationWarning''', + '''default:Brotli >= 1.2.0 is required to prevent decompression bombs\.:urllib3.exceptions.DependencyWarning''', '''default:Async generator 'quart\.wrappers\.response\.DataBody\.__aiter__\.\._aiter' was garbage collected.*:ResourceWarning''', # https://github.com/pallets/quart/issues/301 '''default:unclosed file <_io\.BufferedWriter name='/dev/null'>:ResourceWarning''', # https://github.com/SeleniumHQ/selenium/issues/13328 ] diff --git a/src/urllib3/response.py b/src/urllib3/response.py index b8e8565c..4304133e 100644 --- a/src/urllib3/response.py +++ b/src/urllib3/response.py @@ -49,6 +49,7 @@ from .connection import BaseSSLError, HTTPConnection, HTTPException from .exceptions import ( BodyNotHttplibCompatible, DecodeError, + DependencyWarning, HTTPError, IncompleteRead, InvalidChunkLength, @@ -68,7 +69,11 @@ log = logging.getLogger(__name__) class ContentDecoder: - def decompress(self, data: bytes) -> bytes: + def decompress(self, data: bytes, max_length: int = -1) -> bytes: + raise NotImplementedError() + + @property + def has_unconsumed_tail(self) -> bool: raise NotImplementedError() def flush(self) -> bytes: @@ -78,30 +83,57 @@ class ContentDecoder: class DeflateDecoder(ContentDecoder): def __init__(self) -> None: self._first_try = True - self._data = b"" + self._first_try_data = b"" + self._unfed_data = b"" self._obj = zlib.decompressobj() - def decompress(self, data: bytes) -> bytes: - if not data: + def decompress(self, data: bytes, max_length: int = -1) -> bytes: + data = self._unfed_data + data + self._unfed_data = b"" + if not data and not self._obj.unconsumed_tail: return data + original_max_length = max_length + if original_max_length < 0: + max_length = 0 + elif original_max_length == 0: + # We should not pass 0 to the zlib decompressor because 0 is + # the default value that will make zlib decompress without a + # length limit. + # Data should be stored for subsequent calls. + self._unfed_data = data + return b"" + # Subsequent calls always reuse `self._obj`. zlib requires + # passing the unconsumed tail if decompression is to continue. if not self._first_try: - return self._obj.decompress(data) + return self._obj.decompress( + self._obj.unconsumed_tail + data, max_length=max_length + ) - self._data += data + # First call tries with RFC 1950 ZLIB format. + self._first_try_data += data try: - decompressed = self._obj.decompress(data) + decompressed = self._obj.decompress(data, max_length=max_length) if decompressed: self._first_try = False - self._data = None # type: ignore[assignment] + self._first_try_data = b"" return decompressed + # On failure, it falls back to RFC 1951 DEFLATE format. except zlib.error: self._first_try = False self._obj = zlib.decompressobj(-zlib.MAX_WBITS) try: - return self.decompress(self._data) + return self.decompress( + self._first_try_data, max_length=original_max_length + ) finally: - self._data = None # type: ignore[assignment] + self._first_try_data = b"" + + @property + def has_unconsumed_tail(self) -> bool: + return bool(self._unfed_data) or ( + bool(self._obj.unconsumed_tail) and not self._first_try + ) def flush(self) -> bytes: return self._obj.flush() @@ -117,27 +149,61 @@ class GzipDecoder(ContentDecoder): def __init__(self) -> None: self._obj = zlib.decompressobj(16 + zlib.MAX_WBITS) self._state = GzipDecoderState.FIRST_MEMBER + self._unconsumed_tail = b"" - def decompress(self, data: bytes) -> bytes: + def decompress(self, data: bytes, max_length: int = -1) -> bytes: ret = bytearray() - if self._state == GzipDecoderState.SWALLOW_DATA or not data: + if self._state == GzipDecoderState.SWALLOW_DATA: return bytes(ret) + + if max_length == 0: + # We should not pass 0 to the zlib decompressor because 0 is + # the default value that will make zlib decompress without a + # length limit. + # Data should be stored for subsequent calls. + self._unconsumed_tail += data + return b"" + + # zlib requires passing the unconsumed tail to the subsequent + # call if decompression is to continue. + data = self._unconsumed_tail + data + if not data and self._obj.eof: + return bytes(ret) + while True: try: - ret += self._obj.decompress(data) + ret += self._obj.decompress( + data, max_length=max(max_length - len(ret), 0) + ) except zlib.error: previous_state = self._state # Ignore data after the first error self._state = GzipDecoderState.SWALLOW_DATA + self._unconsumed_tail = b"" if previous_state == GzipDecoderState.OTHER_MEMBERS: # Allow trailing garbage acceptable in other gzip clients return bytes(ret) raise - data = self._obj.unused_data + + self._unconsumed_tail = data = ( + self._obj.unconsumed_tail or self._obj.unused_data + ) + if max_length > 0 and len(ret) >= max_length: + break + if not data: return bytes(ret) - self._state = GzipDecoderState.OTHER_MEMBERS - self._obj = zlib.decompressobj(16 + zlib.MAX_WBITS) + # When the end of a gzip member is reached, a new decompressor + # must be created for unused (possibly future) data. + if self._obj.eof: + self._state = GzipDecoderState.OTHER_MEMBERS + self._obj = zlib.decompressobj(16 + zlib.MAX_WBITS) + + return bytes(ret) + + @property + def has_unconsumed_tail(self) -> bool: + return bool(self._unconsumed_tail) def flush(self) -> bytes: return self._obj.flush() @@ -152,9 +218,35 @@ if brotli is not None: def __init__(self) -> None: self._obj = brotli.Decompressor() if hasattr(self._obj, "decompress"): - setattr(self, "decompress", self._obj.decompress) + setattr(self, "_decompress", self._obj.decompress) else: - setattr(self, "decompress", self._obj.process) + setattr(self, "_decompress", self._obj.process) + + # Requires Brotli >= 1.2.0 for `output_buffer_limit`. + def _decompress(self, data: bytes, output_buffer_limit: int = -1) -> bytes: + raise NotImplementedError() + + def decompress(self, data: bytes, max_length: int = -1) -> bytes: + try: + if max_length > 0: + return self._decompress(data, output_buffer_limit=max_length) + else: + return self._decompress(data) + except TypeError: + # Fallback for Brotli/brotlicffi/brotlipy versions without + # the `output_buffer_limit` parameter. + warnings.warn( + "Brotli >= 1.2.0 is required to prevent decompression bombs.", + DependencyWarning, + ) + return self._decompress(data) + + @property + def has_unconsumed_tail(self) -> bool: + try: + return not self._obj.can_accept_more_data() + except AttributeError: + return False def flush(self) -> bytes: if hasattr(self._obj, "flush"): @@ -168,16 +260,46 @@ if HAS_ZSTD: def __init__(self) -> None: self._obj = zstd.ZstdDecompressor().decompressobj() - def decompress(self, data: bytes) -> bytes: - if not data: + def decompress(self, data: bytes, max_length: int = -1) -> bytes: + if not data and not self.has_unconsumed_tail: return b"" - data_parts = [self._obj.decompress(data)] - while self._obj.eof and self._obj.unused_data: + if self._obj.eof: + data = self._obj.unused_data + data + self._obj = zstd.ZstdDecompressor() + part = self._obj.decompress(data, max_length=max_length) + length = len(part) + data_parts = [part] + # Every loop iteration is supposed to read data from a separate frame. + # The loop breaks when: + # - enough data is read; + # - no more unused data is available; + # - end of the last read frame has not been reached (i.e., + # more data has to be fed). + while ( + self._obj.eof + and self._obj.unused_data + and (max_length < 0 or length < max_length) + ): unused_data = self._obj.unused_data - self._obj = zstd.ZstdDecompressor().decompressobj() - data_parts.append(self._obj.decompress(unused_data)) + if not self._obj.needs_input: + self._obj = zstd.ZstdDecompressor() + part = self._obj.decompress( + unused_data, + max_length=(max_length - length) if max_length > 0 else -1, + ) + if part_length := len(part): + data_parts.append(part) + length += part_length + elif self._obj.needs_input: + break return b"".join(data_parts) + @property + def has_unconsumed_tail(self) -> bool: + return not (self._obj.needs_input or self._obj.eof) or bool( + self._obj.unused_data + ) + def flush(self) -> bytes: ret = self._obj.flush() # note: this is a no-op if not self._obj.eof: @@ -210,10 +332,35 @@ class MultiDecoder(ContentDecoder): def flush(self) -> bytes: return self._decoders[0].flush() - def decompress(self, data: bytes) -> bytes: - for d in reversed(self._decoders): - data = d.decompress(data) - return data + def decompress(self, data: bytes, max_length: int = -1) -> bytes: + if max_length <= 0: + for d in reversed(self._decoders): + data = d.decompress(data) + return data + + ret = bytearray() + # Every while loop iteration goes through all decoders once. + # It exits when enough data is read or no more data can be read. + # It is possible that the while loop iteration does not produce + # any data because we retrieve up to `max_length` from every + # decoder, and the amount of bytes may be insufficient for the + # next decoder to produce enough/any output. + while True: + any_data = False + for d in reversed(self._decoders): + data = d.decompress(data, max_length=max_length - len(ret)) + if data: + any_data = True + # We should not break when no data is returned because + # next decoders may produce data even with empty input. + ret += data + if not any_data or len(ret) >= max_length: + return bytes(ret) + data = b"" + + @property + def has_unconsumed_tail(self) -> bool: + return any(d.has_unconsumed_tail for d in self._decoders) def _get_decoder(mode: str) -> ContentDecoder: @@ -246,9 +393,6 @@ class BytesQueueBuffer: * self.buffer, which contains the full data * the largest chunk that we will copy in get() - - The worst case scenario is a single chunk, in which case we'll make a full copy of - the data inside get(). """ def __init__(self) -> None: @@ -270,6 +414,10 @@ class BytesQueueBuffer: elif n < 0: raise ValueError("n should be > 0") + if len(self.buffer[0]) == n and isinstance(self.buffer[0], bytes): + self._size -= n + return self.buffer.popleft() + fetched = 0 ret = io.BytesIO() while fetched < n: @@ -473,7 +621,11 @@ class BaseHTTPResponse(io.IOBase): self._decoder = _get_decoder(content_encoding) def _decode( - self, data: bytes, decode_content: bool | None, flush_decoder: bool + self, + data: bytes, + decode_content: bool | None, + flush_decoder: bool, + max_length: int | None = None, ) -> bytes: """ Decode the data passed in and potentially flush the decoder. @@ -486,9 +638,12 @@ class BaseHTTPResponse(io.IOBase): ) return data + if max_length is None or flush_decoder: + max_length = -1 + try: if self._decoder: - data = self._decoder.decompress(data) + data = self._decoder.decompress(data, max_length=max_length) self._has_decoded_content = True except self.DECODER_ERROR_CLASSES as e: content_encoding = self.headers.get("content-encoding", "").lower() @@ -953,6 +1108,14 @@ class HTTPResponse(BaseHTTPResponse): elif amt is not None: cache_content = False + if self._decoder and self._decoder.has_unconsumed_tail: + decoded_data = self._decode( + b"", + decode_content, + flush_decoder=False, + max_length=amt - len(self._decoded_buffer), + ) + self._decoded_buffer.put(decoded_data) if len(self._decoded_buffer) >= amt: return self._decoded_buffer.get(amt) @@ -960,7 +1123,11 @@ class HTTPResponse(BaseHTTPResponse): flush_decoder = amt is None or (amt != 0 and not data) - if not data and len(self._decoded_buffer) == 0: + if ( + not data + and len(self._decoded_buffer) == 0 + and not (self._decoder and self._decoder.has_unconsumed_tail) + ): return data if amt is None: @@ -977,7 +1144,12 @@ class HTTPResponse(BaseHTTPResponse): ) return data - decoded_data = self._decode(data, decode_content, flush_decoder) + decoded_data = self._decode( + data, + decode_content, + flush_decoder, + max_length=amt - len(self._decoded_buffer), + ) self._decoded_buffer.put(decoded_data) while len(self._decoded_buffer) < amt and data: @@ -985,7 +1157,12 @@ class HTTPResponse(BaseHTTPResponse): # For example, the GZ file header takes 10 bytes, we don't want to read # it one byte at a time data = self._raw_read(amt) - decoded_data = self._decode(data, decode_content, flush_decoder) + decoded_data = self._decode( + data, + decode_content, + flush_decoder, + max_length=amt - len(self._decoded_buffer), + ) self._decoded_buffer.put(decoded_data) data = self._decoded_buffer.get(amt) @@ -1020,6 +1197,20 @@ class HTTPResponse(BaseHTTPResponse): "Calling read1(decode_content=False) is not supported after " "read1(decode_content=True) was called." ) + if ( + self._decoder + and self._decoder.has_unconsumed_tail + and (amt is None or len(self._decoded_buffer) < amt) + ): + decoded_data = self._decode( + b"", + decode_content, + flush_decoder=False, + max_length=( + amt - len(self._decoded_buffer) if amt is not None else None + ), + ) + self._decoded_buffer.put(decoded_data) if len(self._decoded_buffer) > 0: if amt is None: return self._decoded_buffer.get_all() @@ -1035,7 +1226,9 @@ class HTTPResponse(BaseHTTPResponse): self._init_decoder() while True: flush_decoder = not data - decoded_data = self._decode(data, decode_content, flush_decoder) + decoded_data = self._decode( + data, decode_content, flush_decoder, max_length=amt + ) self._decoded_buffer.put(decoded_data) if decoded_data or flush_decoder: break @@ -1066,7 +1259,11 @@ class HTTPResponse(BaseHTTPResponse): if self.chunked and self.supports_chunked_reads(): yield from self.read_chunked(amt, decode_content=decode_content) else: - while not is_fp_closed(self._fp) or len(self._decoded_buffer) > 0: + while ( + not is_fp_closed(self._fp) + or len(self._decoded_buffer) > 0 + or (self._decoder and self._decoder.has_unconsumed_tail) + ): data = self.read(amt=amt, decode_content=decode_content) if data: @@ -1218,7 +1415,10 @@ class HTTPResponse(BaseHTTPResponse): break chunk = self._handle_chunk(amt) decoded = self._decode( - chunk, decode_content=decode_content, flush_decoder=False + chunk, + decode_content=decode_content, + flush_decoder=False, + max_length=amt, ) if decoded: yield decoded -- 2.34.1