http: drop custom http client logic
Eight and a half years ago, as my starter bug on code.google.com, I
investigated a mysterious "broken pipe" error from seemingly random
clients[0]. That investigation revealed a tragic story: the Python
standard library's httplib was (and remains) barely functional. During
large POSTs, if a server responds early with an error (even a
permission denied error!) the client only notices that the server
closed the connection and everything breaks. Such server behavior is
implicitly legal under RFC 2616 (the latest HTTP RFC as of when I was
last working on this), and my understanding is that later RFCs have
made it explicitly legal to respond early with any status code outside
the 2xx range.
I embarked, probably foolishly, on a journey to write a new http
library with better overall behavior. The http library appears to work
well in most cases, but it can get confused in the presence of
proxies, and it depends on select(2) which limits its utility if a lot
of file descriptors are open. I haven't touched the http library in
almost two years, and in the interim the Python community has
discovered a better way[1] of writing network code. In theory some day
urllib3 will have its own home-grown http library built on h11[2], or
we could do that. Either way, it's time to declare our current
confusingly-named "http2" client logic and move on. I do hope to
revisit this some day: it's still garbage that we can't even respond
with a 401 or 403 without reading the entire POST body from the
client, but the goalposts on writing a new http client library have
moved substantially. We're almost certainly better off just switching
to requests and eventually picking up their http fixes than trying to
live with something that realistically only we'll ever use. Another
approach would be to write an adapter so that Mercurial can use pycurl
if it's installed. Neither of those approaches seem like they should
be investigated prior to a release of Mercurial that works on Python
3: that's where the mindshare is going to be for any improvements to
the state of the http client art.
0: http://web.archive.org/web/20130501031801/http://code.google.com/p/support/issues/detail?id=2716
1: http://sans-io.readthedocs.io/
2: https://github.com/njsmith/h11
Differential Revision: https://phab.mercurial-scm.org/D2444
===================
Mercurial Rust Code
===================
This directory contains various Rust code for the Mercurial project.
The top-level ``Cargo.toml`` file defines a workspace containing
all primary Mercurial crates.
Building
========
To build the Rust components::
$ cargo build
If you prefer a non-debug / release configuration::
$ cargo build --release
Features
--------
The following Cargo features are available:
localdev (default)
Produce files that work with an in-source-tree build.
In this mode, the build finds and uses a ``python2.7`` binary from
``PATH``. The ``hg`` binary assumes it runs from ``rust/target/<target>hg``
and it finds Mercurial files at ``dirname($0)/../../../``.
Build Mechanism
---------------
The produced ``hg`` binary is *bound* to a CPython installation. The
binary links against and loads a CPython library that is discovered
at build time (by a ``build.rs`` Cargo build script). The Python
standard library defined by this CPython installation is also used.
Finding the appropriate CPython installation to use is done by
the ``python27-sys`` crate's ``build.rs``. Its search order is::
1. ``PYTHON_SYS_EXECUTABLE`` environment variable.
2. ``python`` executable on ``PATH``
3. ``python2`` executable on ``PATH``
4. ``python2.7`` executable on ``PATH``
Additional verification of the found Python will be performed by our
``build.rs`` to ensure it meets Mercurial's requirements.
Details about the build-time configured Python are built into the
produced ``hg`` binary. This means that a built ``hg`` binary is only
suitable for a specific, well-defined role. These roles are controlled
by Cargo features (see above).
Running
=======
The ``hgcli`` crate produces an ``hg`` binary. You can run this binary
via ``cargo run``::
$ cargo run --manifest-path hgcli/Cargo.toml
Or directly::
$ target/debug/hg
$ target/release/hg
You can also run the test harness with this binary::
$ ./run-tests.py --with-hg ../rust/target/debug/hg
.. note::
Integration with the test harness is still preliminary. Remember to
``cargo build`` after changes because the test harness doesn't yet
automatically build Rust code.