-
-
Notifications
You must be signed in to change notification settings - Fork 2k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
HttpPayloadParser fails with "Not enough data for satisfy transfer length header" on chunked transfer encoding if the data is split exactly where trailers could occur #4630
Comments
This happens on the split Also note that while |
…s when the buffer ends in an unfortunate spot aio-libs/aiohttp#4630
Could you please submit a PR with a test for that? Also, are you relying on the cext version of http_parser or pure-python? |
I'd love to, but it wasn't immediately obvious to me how this needs to be fixed since the parser is not straightforward to understand due to the state transitions and is entirely undocumented, and I don't currently have time to dig into this deeper. The |
@JustAnotherArchivist Could you provide PR with the test only, without fixing the bug ? |
@socketpair Done: #4651 While writing this, I realised that the error does not occur if there is a trailer field and the split happens before the closing CRLF, but it does happen if the split is between that CR and LF. I also added test cases for both of these. |
PR #4651 by @JustAnotherArchivist This change adds tests that demonstrate the failures described in #4630. They are marked as xfail so that they don't affect the CI status. Once the issue is fixed, they'll reported as XPASS and pytest will fail, which would be a signal that it's time to remove the xfail markers keeping the contents of the tests to prevent regressions. (ref: https://pganssle-talks.github.io/xfail-lightning) Co-Authored-By: Sviatoslav Sydorenko <[email protected]>
FTR the PR with tests is in. Once the defect is fixed, |
HttpPayloadParser waits for trailers indefinitely even if there are no trailers at the response. This happens when only the last CRLF or the last LF are sent via separate TCP segment. When the connection is keep alive and if this bug occurs then users experience response timeout. But this problem is not exposed when keep alive is disabled because .feed_eof is called.
If the last CRLF or only the LF are received via separate TCP segment, HTTPPayloadParser misjudges that trailers should come after 0\r\n in the chunked response body. In this case, HttpPayloadParser starts waiting for trailers, but the only remaining data to be received is CRLF. Thus, HttpPayloadParser waits trailers indefinitely and this incurs TimeoutError in user code. However, if the connection is keep alive disabled, this problem is not reproduced because the server shutdown the connection explicitly after sending all data. If the connection is closed .feed_eof is called and it helps HttpPayloadParser finish its waiting.
If the last CRLF or only the LF are received via separate TCP segment, HTTPPayloadParser misjudges that trailers should come after 0\r\n in the chunked response body. In this case, HttpPayloadParser starts waiting for trailers, but the only remaining data to be received is CRLF. Thus, HttpPayloadParser waits trailers indefinitely and this incurs TimeoutError in user code. However, if the connection is keep alive disabled, this problem is not reproduced because the server shutdown the connection explicitly after sending all data. If the connection is closed .feed_eof is called and it helps HttpPayloadParser finish its waiting.
If the last CRLF or only the LF are received via separate TCP segment, HTTPPayloadParser misjudges that trailers should come after 0\r\n in the chunked response body. In this case, HttpPayloadParser starts waiting for trailers, but the only remaining data to be received is CRLF. Thus, HttpPayloadParser waits trailers indefinitely and this incurs TimeoutError in user code. However, if the connection is keep alive disabled, this problem is not reproduced because the server shutdown the connection explicitly after sending all data. If the connection is closed .feed_eof is called and it helps HttpPayloadParser finish its waiting.
PR aio-libs#4651 by @JustAnotherArchivist This change adds tests that demonstrate the failures described in aio-libs#4630. They are marked as xfail so that they don't affect the CI status. Once the issue is fixed, they'll reported as XPASS and pytest will fail, which would be a signal that it's time to remove the xfail markers keeping the contents of the tests to prevent regressions. (ref: https://pganssle-talks.github.io/xfail-lightning) Co-Authored-By: Sviatoslav Sydorenko <[email protected]>
* Parse the last CRLF of chunked response correctly (#4630) If the last CRLF or only the LF are received via separate TCP segment, HTTPPayloadParser misjudges that trailers should come after 0\r\n in the chunked response body. In this case, HttpPayloadParser starts waiting for trailers, but the only remaining data to be received is CRLF. Thus, HttpPayloadParser waits trailers indefinitely and this incurs TimeoutError in user code. However, if the connection is keep alive disabled, this problem is not reproduced because the server shutdown the connection explicitly after sending all data. If the connection is closed .feed_eof is called and it helps HttpPayloadParser finish its waiting. Co-authored-by: JustAnotherArchivist <[email protected]> Co-authored-by: Sviatoslav Sydorenko <[email protected]> Co-authored-by: Andrew Svetlov <[email protected]>
* Parse the last CRLF of chunked response correctly (#4630) If the last CRLF or only the LF are received via separate TCP segment, HTTPPayloadParser misjudges that trailers should come after 0\r\n in the chunked response body. In this case, HttpPayloadParser starts waiting for trailers, but the only remaining data to be received is CRLF. Thus, HttpPayloadParser waits trailers indefinitely and this incurs TimeoutError in user code. However, if the connection is keep alive disabled, this problem is not reproduced because the server shutdown the connection explicitly after sending all data. If the connection is closed .feed_eof is called and it helps HttpPayloadParser finish its waiting. Co-authored-by: JustAnotherArchivist <[email protected]> Co-authored-by: Sviatoslav Sydorenko <[email protected]> Co-authored-by: Andrew Svetlov <[email protected]>
PR #4651 by @JustAnotherArchivist This change adds tests that demonstrate the failures described in #4630. They are marked as xfail so that they don't affect the CI status. Once the issue is fixed, they'll reported as XPASS and pytest will fail, which would be a signal that it's time to remove the xfail markers keeping the contents of the tests to prevent regressions. (ref: https://pganssle-talks.github.io/xfail-lightning) Co-Authored-By: Sviatoslav Sydorenko <[email protected]>
* Parse the last CRLF of chunked response correctly (#4630) If the last CRLF or only the LF are received via separate TCP segment, HTTPPayloadParser misjudges that trailers should come after 0\r\n in the chunked response body. In this case, HttpPayloadParser starts waiting for trailers, but the only remaining data to be received is CRLF. Thus, HttpPayloadParser waits trailers indefinitely and this incurs TimeoutError in user code. However, if the connection is keep alive disabled, this problem is not reproduced because the server shutdown the connection explicitly after sending all data. If the connection is closed .feed_eof is called and it helps HttpPayloadParser finish its waiting. Co-authored-by: JustAnotherArchivist <[email protected]> Co-authored-by: Sviatoslav Sydorenko <[email protected]> Co-authored-by: Andrew Svetlov <[email protected]>
This fixes py-yarl in pkgsrc being too new for py-aiohttp. 3.7.0 (2020-10-24) ================== Features -------- - Response headers are now prepared prior to running ``on_response_prepare`` hooks, directly before headers are sent to the client. `#1958 <https://github.com/aio-libs/aiohttp/issues/1958>`_ - Add a ``quote_cookie`` option to ``CookieJar``, a way to skip quotation wrapping of cookies containing special characters. `#2571 <https://github.com/aio-libs/aiohttp/issues/2571>`_ - Call ``AccessLogger.log`` with the current exception available from ``sys.exc_info()``. `#3557 <https://github.com/aio-libs/aiohttp/issues/3557>`_ - `web.UrlDispatcher.add_routes` and `web.Application.add_routes` return a list of registered `AbstractRoute` instances. `AbstractRouteDef.register` (and all subclasses) return a list of registered resources registered resource. `#3866 <https://github.com/aio-libs/aiohttp/issues/3866>`_ - Added properties of default ClientSession params to ClientSession class so it is available for introspection `#3882 <https://github.com/aio-libs/aiohttp/issues/3882>`_ - Don't cancel web handler on peer disconnection, raise `OSError` on reading/writing instead. `#4080 <https://github.com/aio-libs/aiohttp/issues/4080>`_ - Implement BaseRequest.get_extra_info() to access a protocol transports' extra info. `#4189 <https://github.com/aio-libs/aiohttp/issues/4189>`_ - Added `ClientSession.timeout` property. `#4191 <https://github.com/aio-libs/aiohttp/issues/4191>`_ - allow use of SameSite in cookies. `#4224 <https://github.com/aio-libs/aiohttp/issues/4224>`_ - Use ``loop.sendfile()`` instead of custom implementation if available. `#4269 <https://github.com/aio-libs/aiohttp/issues/4269>`_ - Apply SO_REUSEADDR to test server's socket. `#4393 <https://github.com/aio-libs/aiohttp/issues/4393>`_ - Use .raw_host instead of slower .host in client API `#4402 <https://github.com/aio-libs/aiohttp/issues/4402>`_ - Allow configuring the buffer size of input stream by passing ``read_bufsize`` argument. `#4453 <https://github.com/aio-libs/aiohttp/issues/4453>`_ - Pass tests on Python 3.8 for Windows. `#4513 <https://github.com/aio-libs/aiohttp/issues/4513>`_ - Add `method` and `url` attributes to `TraceRequestChunkSentParams` and `TraceResponseChunkReceivedParams`. `#4674 <https://github.com/aio-libs/aiohttp/issues/4674>`_ - Add ClientResponse.ok property for checking status code under 400. `#4711 <https://github.com/aio-libs/aiohttp/issues/4711>`_ - Don't ceil timeouts that are smaller than 5 seconds. `#4850 <https://github.com/aio-libs/aiohttp/issues/4850>`_ - TCPSite now listens by default on all interfaces instead of just IPv4 when `None` is passed in as the host. `#4894 <https://github.com/aio-libs/aiohttp/issues/4894>`_ - Bump ``http_parser`` to 2.9.4 `#5070 <https://github.com/aio-libs/aiohttp/issues/5070>`_ Bugfixes -------- - Fix keepalive connections not being closed in time `#3296 <https://github.com/aio-libs/aiohttp/issues/3296>`_ - Fix failed websocket handshake leaving connection hanging. `#3380 <https://github.com/aio-libs/aiohttp/issues/3380>`_ - Fix tasks cancellation order on exit. The run_app task needs to be cancelled first for cleanup hooks to run with all tasks intact. `#3805 <https://github.com/aio-libs/aiohttp/issues/3805>`_ - Don't start heartbeat until _writer is set `#4062 <https://github.com/aio-libs/aiohttp/issues/4062>`_ - Fix handling of multipart file uploads without a content type. `#4089 <https://github.com/aio-libs/aiohttp/issues/4089>`_ - Preserve view handler function attributes across middlewares `#4174 <https://github.com/aio-libs/aiohttp/issues/4174>`_ - Fix the string representation of ``ServerDisconnectedError``. `#4175 <https://github.com/aio-libs/aiohttp/issues/4175>`_ - Raising RuntimeError when trying to get encoding from not read body `#4214 <https://github.com/aio-libs/aiohttp/issues/4214>`_ - Remove warning messages from noop. `#4282 <https://github.com/aio-libs/aiohttp/issues/4282>`_ - Raise ClientPayloadError if FormData re-processed. `#4345 <https://github.com/aio-libs/aiohttp/issues/4345>`_ - Fix a warning about unfinished task in ``web_protocol.py`` `#4408 <https://github.com/aio-libs/aiohttp/issues/4408>`_ - Fixed 'deflate' compression. According to RFC 2616 now. `#4506 <https://github.com/aio-libs/aiohttp/issues/4506>`_ - Fixed OverflowError on platforms with 32-bit time_t `#4515 <https://github.com/aio-libs/aiohttp/issues/4515>`_ - Fixed request.body_exists returns wrong value for methods without body. `#4528 <https://github.com/aio-libs/aiohttp/issues/4528>`_ - Fix connecting to link-local IPv6 addresses. `#4554 <https://github.com/aio-libs/aiohttp/issues/4554>`_ - Fix a problem with connection waiters that are never awaited. `#4562 <https://github.com/aio-libs/aiohttp/issues/4562>`_ - Always make sure transport is not closing before reuse a connection. Reuse a protocol based on keepalive in headers is unreliable. For example, uWSGI will not support keepalive even it serves a HTTP 1.1 request, except explicitly configure uWSGI with a ``--http-keepalive`` option. Servers designed like uWSGI could cause aiohttp intermittently raise a ConnectionResetException when the protocol poll runs out and some protocol is reused. `#4587 <https://github.com/aio-libs/aiohttp/issues/4587>`_ - Handle the last CRLF correctly even if it is received via separate TCP segment. `#4630 <https://github.com/aio-libs/aiohttp/issues/4630>`_ - Fix the register_resource function to validate route name before splitting it so that route name can include python keywords. `#4691 <https://github.com/aio-libs/aiohttp/issues/4691>`_ - Improve typing annotations for ``web.Request``, ``aiohttp.ClientResponse`` and ``multipart`` module. `#4736 <https://github.com/aio-libs/aiohttp/issues/4736>`_ - Fix resolver task is not awaited when connector is cancelled `#4795 <https://github.com/aio-libs/aiohttp/issues/4795>`_ - Fix a bug "Aiohttp doesn't return any error on invalid request methods" `#4798 <https://github.com/aio-libs/aiohttp/issues/4798>`_ - Fix HEAD requests for static content. `#4809 <https://github.com/aio-libs/aiohttp/issues/4809>`_ - Fix incorrect size calculation for memoryview `#4890 <https://github.com/aio-libs/aiohttp/issues/4890>`_ - Add HTTPMove to _all__. `#4897 <https://github.com/aio-libs/aiohttp/issues/4897>`_ - Fixed the type annotations in the ``tracing`` module. `#4912 <https://github.com/aio-libs/aiohttp/issues/4912>`_ - Fix typing for multipart ``__aiter__``. `#4931 <https://github.com/aio-libs/aiohttp/issues/4931>`_ - Fix for race condition on connections in BaseConnector that leads to exceeding the connection limit. `#4936 <https://github.com/aio-libs/aiohttp/issues/4936>`_ - Add forced UTF-8 encoding for ``application/rdap+json`` responses. `#4938 <https://github.com/aio-libs/aiohttp/issues/4938>`_ - Fix inconsistency between Python and C http request parsers in parsing pct-encoded URL. `#4972 <https://github.com/aio-libs/aiohttp/issues/4972>`_ - Fix connection closing issue in HEAD request. `#5012 <https://github.com/aio-libs/aiohttp/issues/5012>`_ - Fix type hint on BaseRunner.addresses (from ``List[str]`` to ``List[Any]``) `#5086 <https://github.com/aio-libs/aiohttp/issues/5086>`_ - Make `web.run_app()` more responsive to Ctrl+C on Windows for Python < 3.8. It slightly increases CPU load as a side effect. `#5098 <https://github.com/aio-libs/aiohttp/issues/5098>`_ Improved Documentation ---------------------- - Fix example code in client quick-start `#3376 <https://github.com/aio-libs/aiohttp/issues/3376>`_ - Updated the docs so there is no contradiction in ``ttl_dns_cache`` default value `#3512 <https://github.com/aio-libs/aiohttp/issues/3512>`_ - Add 'Deploy with SSL' to docs. `#4201 <https://github.com/aio-libs/aiohttp/issues/4201>`_ - Change typing of the secure argument on StreamResponse.set_cookie from ``Optional[str]`` to ``Optional[bool]`` `#4204 <https://github.com/aio-libs/aiohttp/issues/4204>`_ - Changes ``ttl_dns_cache`` type from int to Optional[int]. `#4270 <https://github.com/aio-libs/aiohttp/issues/4270>`_ - Simplify README hello word example and add a documentation page for people coming from requests. `#4272 <https://github.com/aio-libs/aiohttp/issues/4272>`_ - Improve some code examples in the documentation involving websockets and starting a simple HTTP site with an AppRunner. `#4285 <https://github.com/aio-libs/aiohttp/issues/4285>`_ - Fix typo in code example in Multipart docs `#4312 <https://github.com/aio-libs/aiohttp/issues/4312>`_ - Fix code example in Multipart section. `#4314 <https://github.com/aio-libs/aiohttp/issues/4314>`_ - Update contributing guide so new contributors read the most recent version of that guide. Update command used to create test coverage reporting. `#4810 <https://github.com/aio-libs/aiohttp/issues/4810>`_ - Spelling: Change "canonize" to "canonicalize". `#4986 <https://github.com/aio-libs/aiohttp/issues/4986>`_ - Add ``aiohttp-sse-client`` library to third party usage list. `#5084 <https://github.com/aio-libs/aiohttp/issues/5084>`_ Misc ---- - `#2856 <https://github.com/aio-libs/aiohttp/issues/2856>`_, `#4218 <https://github.com/aio-libs/aiohttp/issues/4218>`_, `#4250 <https://github.com/aio-libs/aiohttp/issues/4250>`_
🐞 Describe the bug
If the response data is fed into
HttpPayloadParser.feed_data
in a particular way, the parser is unable to successfully parse chunked data. Specifically, this happens when one call to the function contains the last0\r\n
chunk but the following\r\n
is supplied in a separate call.💡 To Reproduce
I added print statements here to debug what exactly aiohttp is returning compared to the simple
feed_data(b'4\r\nasdf\r\n0\r\n\r\n')
call (which works fine).💡 Expected behavior
The parser is able to process the data, and the last line produces
b'asdf'
.📋 Logs/tracebacks
📋 Your version of the Python
📋 Your version of the aiohttp/yarl/multidict distributions
📋 Additional context
Discovered with aiohttp 2.3.10 due to errors in qwarc, which definitely uses aiohttp in weird, undocumented, and unsupported ways. But I believe the error could happen also in normal aiohttp usage if the data returned from the server is just the right size, namely two bytes over a multiple of the internal buffer size.
The text was updated successfully, but these errors were encountered: