sshpeer: defer pipe buffering and stderr sidechannel binding The doublepipe and bufferedinputpipe types facilitate polling multiple pipes without blocking and for automatically forwarding output from the SSH server's stderr pipe to the ui as "remote: " output. This all happens automatically and callers don't need to worry about reading from multiple pipes. An upcoming change to version 2 of the SSH wire protocol will eliminate the use of stderr and move side-channel output into the "main" pipe. The SSH wire protocol will use a pair of unidirectional pipes - just like the HTTP protocol. In this future world, the doublepipe primitive isn't necessary because the stderr pipe won't be used. To prepare for eventually not using doublepipe, we delay the construction of this primitive from immediately after connection establishment to inside construction of the peer instance. The handshake occurs between these two events. So we had to teach the handshake code to read from stderr so any stderr output from the server is still attended to early in the connection lifetime. Differential Revision: https://phab.mercurial-scm.org/D2383

sshpeer: make pipe polling code more explicit "hasbuffer" is a property on our special bufferedinputpipe class. When reading this code, I thought it might have had something special to do properties on built-in types. But "hasbuffer" doesn't appear in the CPython code base for either 2.7 or 3.7, so the answer is no. Let's make the code more explicit about the fact that it deals with our special bufferedinputpipe type. Differential Revision: https://phab.mercurial-scm.org/D2382

tests: store protocol payload in files Upcoming changes to version 2 of the SSH protocol will introduce binary components to the protocol. It will be easier to eliminate trailing newlines and use binary in the tests if the protocol payload is being generated by Python. So use inline Python to write payloads to files and pipe those files to server processes instead of shell strings/variables. Differential Revision: https://phab.mercurial-scm.org/D2381

sshpeer: return framed file object when needed Currently, wireproto.wirepeer has a default implementation of _submitbatch() and sshv1peer has a very similar implementation. The main difference is that sshv1peer is aware of the total amount of bytes it can read whereas the default implementation reads the stream until no more data is returned. The default implementation works for HTTP, since there is a known end to HTTP responses (either Content-Length or 0 sized chunk). This commit teaches sshv1peer to use our just-introduced "cappedreader" class for wrapping a file object to limit the number of bytes that can be read. We do this by introducing an argument to specify whether the response is framed. If set, we returned a cappedreader instance instead of the raw pipe. _call() always has framed responses. So we set this argument unconditionally and then .read() the entirety of the result. Strictly speaking, we don't need to use cappedreader in this case and can inline frame decoding/read logic. But I like when things are consistent. The overhead should be negligible. _callstream() and _callcompressable() are special: whether framing is used depends on the specific command. So, we define a set of commands that have framed response. It currently only contains "batch." As a result of this change, the one-off implementation of _submitbatch() in sshv1peer can be removed since it is now safe to .read() the response's file object until end of stream. cappedreader takes care of not overrunning the frame. Differential Revision: https://phab.mercurial-scm.org/D2380

sshpeer: move logic for sending a request into a new function The **args being used to pass arbitrary command arguments is limiting because it makes it harder to control behavior of the function. We factor most of _callstream() into a new function that doesn't use **args. Differential Revision: https://phab.mercurial-scm.org/D2379

sshpeer: rename _recv and _send to _readframed and _writeframed Because it is reading and writing a chunk of data with a well-defined size. "recv" and "send" make it sound like things are a direct proxy to the underlying pipe, which they aren't. Differential Revision: https://phab.mercurial-scm.org/D2378

util: add a file object proxy that can read at most N bytes Sometimes we have data of a known size within a stream. For performance reasons, we don't want to pre-read this data (we want to allow consumers to read on demand). For simplicitly reasons, we don't want callers to necessarily know their data is coming from within an outer stream and there is a limit to how much they should read. The class introduced by this commit provides a very simple proxy around an underlying file object that allows the consumer to .read() up to N bytes from the file object. Attempts to read past this many bytes results in a simulated EOF. Differential Revision: https://phab.mercurial-scm.org/D2377