Gregory Szorc <gregory.szorc@gmail.com> [Wed, 14 Mar 2018 16:51:34 -0700] rev 37057
wireproto: add request IDs to frames
One of my primary goals with the new wire protocol is to make
operations faster and enable both client and server-side
operations to scale to multiple CPU cores.
One of the ways we make server interactions faster is by reducing
the number of round trips to that server.
With the existing wire protocol, the "batch" command facilitates
executing multiple commands from a single request payload. The way
it works is the requests for multiple commands are serialized. The
server executes those commands sequentially then serializes all
their results. As an optimization for reducing round trips, this
is very effective. The technical implementation, however, is pretty
bad and suffers from a number of deficiencies. For example, it
creates a new place where authorization to run a command must be
checked. (The lack of this checking in older Mercurial releases
was CVE-2018-1000132.)
The principles behind the "batch" command are sound. However, the
execution is not. Therefore, I want to ditch "batch" in the
new wire protocol and have protocol level support for issuing
multiple requests in a single round trip.
This commit introduces support in the frame-based wire protocol to
facilitate this. We do this by adding a "request ID" to each frame.
If a server sees frames associated with different "request IDs," it
handles them as separate requests. All of this happening possibly
as part of the same message from client to server (the same request
body in the case of HTTP).
We /could/ model the exchange the way pipelined HTTP requests do,
where the server processes requests in order they are issued and
received. But this artifically constrains scalability. A better
model is to allow multi-requests to be executed concurrently and
for responses to be sent and handled concurrently. So the
specification explicitly allows this. There is some work to be done
around specifying dependencies between multi-requests. We take
the easy road for now and punt on this problem, declaring that
if order is important, clients must not issue the request until
responses to dependent requests have been received.
This commit focuses on the boilerplate of implementing the request
ID. The server reactor still can't manage multiple, in-flight
request IDs. This will be addressed in a subsequent commit.
Because the wire semantics have changed, we bump the version of the
media type.
Differential Revision: https://phab.mercurial-scm.org/D2869
Gregory Szorc <gregory.szorc@gmail.com> [Wed, 14 Mar 2018 14:01:16 -0700] rev 37056
wireproto: buffer output frames when in half duplex mode
Previously, when told that a response was ready, the server reactor
would instruct the caller to send frames immediately. This was OK
as an initial implementation. But it would not work for half-duplex
connections where the sender can't receive until all data has been
transmitted - such as httplib based clients.
In this commit, we teach the reactor that output frames should
be buffered until end of input is seen. This required a new
event to inform the reactor of end of input. The result from that
event will instruct the consumer to send all buffered frames.
The HTTP server is buffered by default.
This change effectively hides the complexity of buffering within
the reactor so that transports need not be concerned about it.
This helps keep the transports "dumb" and will make implementing
multiple requests-responses per atomic exchange (like an HTTP
request) much simpler.
Differential Revision: https://phab.mercurial-scm.org/D2860