wireproto: support for receiving multiple requests Now that we have request IDs on each frame and a specification that allows multiple requests to be issued simultaneously, possibly interleaved, let's teach the server to deal with that. Instead of tracking the state for *the* active command request, we instead track the state of each receiving command by its request ID. The multiple states in our state machine for processing each command's state has been collapsed into a single state for "receiving commands." Tests have been added so our branch coverage covers all meaningful branches. However, we did lose some logical coverage. The implementation of this new feature opens up the door to a server having partial command requests when end of input is reached. We will probably want a mechanism to deal with partial requests. For now, I've tracked that as a known issue in the class docstring. I've also noted an abuse vector that becomes a little bit easier to exploit with this feature. Differential Revision: https://phab.mercurial-scm.org/D2870

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

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

wireproto: define and implement responses in framing protocol Previously, we only had client-side frame types defined. This commit defines and implements basic support for server-side frame types. We introduce two frame types - one for representing the raw bytes result of a command and another for representing error results. The types are quite primitive and behavior will expand over time. But you have to start somewhere. Our server reactor gains methods to react to an intent to send a response. Again, following the "sans I/O" pattern, the reactor doesn't actually send the data. Instead, it gives the caller a generator to frames that it can send out over the wire. Differential Revision: https://phab.mercurial-scm.org/D2858

wireproto: implement basic command dispatching for HTTPv2 Now that we can ingest frames and decode them to requests to run commands, we are able to actually run those commands. So this commit starts to implement that. There are numerous shortcomings. We can't operate on commands with "*" arguments. We can only emit bytesresponse results. We don't yet issue a response in the unified framing protocol. But it's a start. Differential Revision: https://phab.mercurial-scm.org/D2857

wireproto: nominally don't expose "batch" to version 2 wire transports The unified frame-based protocol will (eventually) support multiple requests per client transmission. This means that the [very hacky] "batch" command has no purpose existing in this protocol. This commit marks the command as applying to v1 transports only. But because SSHv2 == SSHv1 currently, we had to hack it back in for the SSHv2 transport. Bleh. Tests changed because the capabilities string changed. The order of tokens in the string is not important. Differential Revision: https://phab.mercurial-scm.org/D2856

wireproto: implement basic frame reading and processing We just implemented support for writing frames. Now let's implement support for reading them. The bulk of the new code is for a class that maintains the state of a server. Essentially, you construct an instance, feed frames to it, and it tells you what you should do next. The design is inspired by the "sans I/O" movement and the reactor pattern. We don't want to perform I/O or any major blocking event during frame ingestion because this arbitrarily limits ways that server pieces can be implemented. For example, it makes it much harder to swap in an alternate implementation based on asyncio or do crazy things like have requests dispatch to other processes. We do still implement readframe() which does I/O. But it is decoupled from the server reactor. And important parsing of frame headers is a standalone function. So I/O is only needed to obtain frame data. Because testing server-side ingest is useful and difficult on running servers, we create a new "debugreflect" endpoint that will echo back to the client what was received and how it was interpreted. This could be useful for a server admin, someone implementing a client. But immediately, it is useful for testing: we're able to demonstrate that frames are parsed correctly and turned into requests to run commands without having to implement command dispatch on the server! In addition, we implement Python level unit tests for the reactor. This is vastly more efficient than sending requests to the "debugreflect" endpoint and vastly more powerful for advanced testing. Differential Revision: https://phab.mercurial-scm.org/D2852

wireproto: define and implement protocol for issuing requests The existing HTTP and SSH wire protocols suffer from a host of flaws and shortcomings. I've been wanting to rewrite the protocol for a while now. Supporting partial clone - which will require new wire protocol commands and capabilities - and other advanced server functionality will be much easier if we start from a clean slate and don't have to be constrained by limitations of the existing wire protocol. This commit starts to introduce a new data exchange format for use over the wire protocol. The new protocol is built on top of "frames," which are atomic units of metadata + data. Frames will make it easier to implement proxies and other mechanisms that want to inspect data without having to maintain state. The existing frame metadata is very minimal and it will evolve heavily. (We will eventually support things like concurrent requests, out-of-order responses, compression, side-channels for status updates, etc. Some of these will require additions to the frame header.) Another benefit of frames is that all reads are of a fixed size. A reader works by consuming a frame header, extracting the payload length, then reading that many bytes. No lookahead, buffering, or memory reallocations are needed. The new protocol attempts to be transport agnostic. I want all that's required to use the new protocol to be a pair of unidirectional, half-duplex pipes. (Yes, we will eventually make use of full-duplex pipes, but that's for another commit.) Notably, when the SSH transport switches to this new protocol, stderr will be unused. This is by design: the lack of stderr on HTTP harms protocol behavior there. By shoehorning everything into a pair of pipes, we can have more consistent behavior across transports. We currently only define the client side parts of the new protocol, specifically the bits for requesting that a command run. This keeps the new code and feature small and somewhat easy to review. We add support to `hg debugwireproto` for writing frames into HTTP request bodies. Our tests that issue commands to the new HTTP endpoint have been updated to transmit frames. The server bits haven't been touched to consume the frames yet. This will occur in the next commit... Astute readers may notice that the command name is transmitted in both the HTTP request URL and the command request frame. This is partially a kludge from me initially implementing the frame-based protocol for SSH first. But it is also a feature: I intend to eventually support issuing multiple commands per HTTP request. This will allow us to replace the abomination that is the "batch" wire protocol command with a protocol-level mechanism for performing multi-dispatch. Because I want the frame-based protocol to be as similar as possible across transports, I'd rather we (redundantly) include the command name in the frame than differ behavior between transports that have out-of-band routing information (like HTTP) readily available. Differential Revision: https://phab.mercurial-scm.org/D2851

wireproto: define content negotiation for HTTPv2 HTTP messages communicate their media types and what media types they can understand via the Content-Type and Accept header, respectively. While I don't want the wire protocol to lean too heavily on HTTP because I'm aiming for the wire protocol to be as transport agnostic as possible, it is nice to play by the spec if possible. This commit defines our media negotiation mechanism for version 2 of the HTTP protocol. Essentially, we mandate the use of a new media type and how clients and servers should react to various headers or lack thereof. The name of the media type is a placeholder. We purposefully don't yet define the format of the new media type because that's a lot of work. I feel pretty strongly that we should use Content-Type. I feel less strongly about Accept. I think it is reasonable for servers to return the media type that was submitted to them. So we may strike this header before the protocol is finished... Differential Revision: https://phab.mercurial-scm.org/D2850

hgweb: also set Content-Type header Our HTTP/WSGI server may convert the Content-Type HTTP request header to the CONTENT_TYPE WSGI environment key and not set HTTP_CONTENT_TYPE. Other WSGI server implementations do this, so I think the behavior is acceptable. So assuming this HTTP request header could get "lost" by the WSGI server, let's restore it on the request object like we do for Content-Length. FWIW, the WSGI server may also *invent* a Content-Type value. The default behavior of Python's RFC 822 message class returns a default media type if Content-Type isn't defined. This is kind of annoying. But RFC 7231 section 3.1.1.5 does say the recipient may assume a media type of application/octet-stream. Python's defaults are for text/plain (given we're using an RFC 822 parser). But whatever. Differential Revision: https://phab.mercurial-scm.org/D2849

wireproto: require POST for all HTTPv2 requests Wire protocol version 1 transfers argument data via request headers by default. This has historically caused problems because servers institute limits on the length of individual HTTP headers as well as the total size of all request headers. Mercurial servers can advertise the maximum length of an individual header. But there's no guarantee any intermediate HTTP agents will accept headers up to that length. In the existing wire protocol, server operators typically also key off the HTTP request method to implement authentication. For example, GET requests translate to read-only requests and can be allowed. But read-write commands must use POST and require authentication. This has typically worked because the only wire protocol commands that use POST modify the repo (e.g. the "unbundle" command). There is an experimental feature to enable clients to transmit argument data via POST request bodies. This is technically a better and more robust solution. But we can't enable it by default because of servers assuming POST means write access. In version 2 of the wire protocol, the permissions of a request are encoded in the URL. And with it being a new protocol in a new URL space, we're not constrained by backwards compatibility requirements. This commit adopts the technically superior mechanism of using HTTP request bodies to send argument data by requiring POST for all commands. Strictly speaking, it may be possible to send request bodies on GET requests. But my experience is that not all HTTP stacks support this. POST pretty much always works. Using POST for read-only operations does sacrifice some RESTful design purity. But this API cares about practicality, not about being in Roy T. Fielding's REST ivory tower. There's a chance we may relax this restriction in the future. But for now, I want to see how far we can get with a POST only API. Differential Revision: https://phab.mercurial-scm.org/D2837

wireproto: define permissions-based routing of HTTPv2 wire protocol Now that we have a scaffolding for serving version 2 of the HTTP protocol, let's start implementing it. A good place to start is URL routing and basic request processing semantics. We can focus on content types, capabilities detect, etc later. Version 2 of the HTTP wire protocol encodes the needed permissions of the request in the URL path. The reasons for this are documented in the added documentation. In short, a) it makes it really easy and fail proof for server administrators to implement path-based authentication and b) it will enable clients to realize very early in a server exchange that authentication will be required to complete the operation. This latter point avoids all kinds of complexity and problems, like dealing with Expect: 100-continue and clients finding out later during `hg push` that they need to provide authentication. This will avoid the current badness where clients send a full bundle, get an HTTP 403, provide authentication, then retransmit the bundle. In order to implement command checking, we needed to implement a protocol handler for the new wire protocol. Our handler is just small enough to run the code we've implemented. Tests for the defined functionality have been added. I very much want to refactor the permissions checking code and define a better response format. But this can be done later. Nothing is covered by backwards compatibility at this point. Differential Revision: https://phab.mercurial-scm.org/D2836

wireproto: support /api/* URL space for exposing APIs I will soon be introducing a new version of the HTTP wire protocol. One of the things I want to change with it is the URL routing. I want to rely on URL paths to define endpoints rather than the "cmd" query string argument. That should be pretty straightforward. I was thinking about what URL space to reserve for the new protocol. We /could/ put everything at a top-level path. e.g. /wireproto/* or /http-v2-wireproto/*. However, these constrain us a bit because they assume there will only be 1 API: version 2 of the HTTP wire protocol. I think there is room to grow multiple APIs. For example, there may someday be a proper JSON API to query or even manipulate the repository. And I don't think we should have to create a new top-level URL space for each API nor should we attempt to shoehorn each future API into the same shared URL space: that would just be too chaotic. This commits reserves the /api/* URL space for all our future API needs. Essentially, all requests to /api/* get routed to a new WSGI handler. By default, it 404's the entire URL space unless the "api server" feature is enabled. When enabled, requests to "/api" list available APIs. URLs of the form /api/<name>/* are reserved for a particular named API. Behavior within each API is left up to that API. So, we can grow new APIs easily without worrying about URL space conflicts. APIs can be registered by adding entries to a global dict. This allows extensions to provide their own APIs should they choose to do so. This is probably a premature feature. But IMO the code is easier to read if we're not dealing with API-specific behavior like config option querying inline. To prove it works, we implement a very basic API for version 2 of the HTTP wire protocol. It does nothing of value except facilitate testing of the /api/* URL space. We currently emit plain text responses for all /api/* endpoints. There's definitely room to look at Accept and other request headers to vary the response format. But we have to start somewhere. Differential Revision: https://phab.mercurial-scm.org/D2834

url: support suppressing Accept header Sending this header automatically could interfere with future testing and client behavior. Let's add a knob to disable the behavior. We don't have a control for User-Agent because urllib will send it if we don't set something. I don't feel like hacking into the bowels of urllib to figure out how to suppress that. UA shouldn't be used for anything meaningful. So it shouldn't pose any problems beyond non-determinism (since the header has the Mercurial version in it). Differential Revision: https://phab.mercurial-scm.org/D2843

util: don't log low-level I/O calls for HTTP peer `hg debugwireproto` is useful for testing HTTP interactions. Possibly more useful than `get-with-headers.py`. But one thing that makes it annoying for mid-level tests is that it logs all API calls, such as readline(). This makes output - especially headers - overly verbose. This commit teaches our file and socket observers to not log API calls on functions dealing with data. We change the behavior of `hg debugwireproto` to enable this mode by default. --debug can be added to restore the previous behavior. As the test changes demonstrate, this makes tests much easier to read. As a bonus, it also removes some required (glob) over lengths in system call results. One thing that's lacking is knowing which side sent data. But we can fix this in a follow-up once it becomes a problem. Differential Revision: https://phab.mercurial-scm.org/D2842