Transport layer abstraction

Hi all,

We had an internal discussion about XPC use with clangd and the result is that XPC-based transport layer in clangd is a hard requirement. This effectively means that the adapter approach is not feasible and I am looking into the transport layer abstraction Sam proposed earlier (https://reviews.llvm.org/D49389). Sorry for the trouble!

There is a difficulty with designing an abstraction over transport layer due to XPC API requiring client code to hand control over. It's necessary for XPC server to call xpc_main() with a message handling callback:
xpc_main(connection_handler);
But XPC is running it's own message loop inside xpc_main() which is supposed to never return (in case of failure exit() is called) which makes it difficult to have common interface for JSON RPC and XPC.
Unless I am mistaken that means that the message loop (Transport::loop(), ClangdLSPServer::run()) can't be part of generic code.

I am curious what are other people's thoughts on this?

I tried to come up with some reasonable abstraction but I am probably already anchored by my previous work (https://reviews.llvm.org/D48559) - all my design ideas tend to go in the direction of abstraction over JsonOutput and custom dispatchers (StdIoDispatcher/XPCDispatcher).

Cheers,

Jan

Hi all,

We had an internal discussion about XPC use with clangd and the result is that XPC-based transport layer in clangd is a hard requirement. This effectively means that the adapter approach is not feasible and I am looking into the transport layer abstraction Sam proposed earlier (https://reviews.llvm.org/D49389). Sorry for the trouble!

No problem! (And sorry if I’ve been a bit unresponsive!)

There is a difficulty with designing an abstraction over transport layer due to XPC API requiring client code to hand control over. It’s necessary for XPC server to call xpc_main() with a message handling callback:
xpc_main(connection_handler);
But XPC is running it’s own message loop inside xpc_main() which is supposed to never return (in case of failure exit() is called) which makes it difficult to have common interface for JSON RPC and XPC.
Unless I am mistaken that means that the message loop (Transport::loop(), ClangdLSPServer::run()) can’t be part of generic code.

I am curious what are other people’s thoughts on this?

That makes sense to me, but I don’t think a common interface is impossible.
I noticed this in your patch, and in D49389 the idea (which I didn’t spell out) was that the XPC transport’s loop() would call xpc_main and never return.
Transport::loop() is specific to a transport, it doesn’t need to be part of generic code. ClangdLSPServer::run() ends up calling Transport::loop(), so it could be common code.

It’d look something like:

static Transport::MessageHandler *IncomingXPCMessages; // XPC → clangd
static xpc_connection_t *XPCConnection; // clangd → XPC

XPCTransport::loop(MessageHandler &M) {

IncomingXPCMessages = &M;
xpc_main(&XPCEntrypoint);
}

static void XPCEntrypoint(xpc_connection_t Conn) {
// I don’t remember how an XPC server looks
// i guess you’re reading editor->clangd messages in a loop?
for (;:wink: {
// lets say you parsed a notification
std::string name = …;

json::Value args = xpcToJSON(…);
IncomingXPCMessages->notify(name, args); // send the notification to clangd
}
}

XPCTransport::notify(StringRef name, json::Value args) {

xpc_object_t XPCNotification = encodeNotificationForXPCSomehow(name, jsonToXPC(args))
xpc_connection_send(*XPCConnection, XPCNotification);
}

WDYT?

Hi Sam,

Thanks for the response.

I think I understand your intentions better now. The Transport abstraction looks good enough to be used to me so let me just explain what I though are imperfections that I know about (no big deals though). If these are ok with you I’d go ahead and work on it. Another thing is that you guys already have another transport layer implementation so if this abstraction fits nicely 2 out of 3 implementations with XPC being kind of little less aligned I’d consider it still pragmatic to use it.

Transport::MessageHandler

I just wonder if this isn't actually the “reverse” of desired interface. It seems to me that MessageHandler is abstracting clangd implementation details when used in specific Transport implementation. For example it would help if we ever decide to swap clangd for another LSP implementation or clangd-ng. But it doesn’t abstract the transport layer itself. What benefit does it grant instead of calling for example ClangdLSPServer::notify() directly?

ClangdLSPServer::run()

IIUIC purpose of this method is to have abstraction over different Transport implementations with common strategies for program return value (ShutdownRequestReceived) and error handling (return value of Transport.loop()). Yet since we can’t do this for XPC (as xpc_main doesn’t return) we’d have to duplicate the code (~10 LOC, no big deal) but it would mean that it wouldn’t be always possible to just blindly swap implementations of this interface.
In other words it’s a violation of Liskov substitution principle which in itself doesn’t necessarily mean it’s a big deal - I am just wondering if it makes sense to have such abstraction.

This is basically what I meant that it’s difficult to have common abstraction over the “input direction” of transport layer (since Transport::MessageHandler looks more like an abstraction over ClangdLSPServer input handling) and because of that I’d consider different dispatchers instead of Transport::loop().

WDYT?

Cheers,

Jan

P. S. For the record (if it would matter in the future) XPC interface is callback-based:

void XPCEntrypoint(xpc_connection_t connection) {
  xpc_connection_set_event_handler(connection,
    ^(xpc_object_t message) {/* C block handling the message */ }
  );
  xpc_connection_resume(connection);
}

int main() {
  xpc_main(XPCEntrypoint);
}

Hi Sam,

Thanks for the response.

I think I understand your intentions better now. The Transport abstraction looks good enough to be used to me so let me just explain what I though are imperfections that I know about (no big deals though). If these are ok with you I’d go ahead and work on it. Another thing is that you guys already have another transport layer implementation so if this abstraction fits nicely 2 out of 3 implementations with XPC being kind of little less aligned I’d consider it still pragmatic to use it.

So Transport isn’t implemented or even really fully fleshed out. We should design it (or something else) to bridge the JSON/XPC divide well, though perfect is always hard.

Our “other transport layer” is really a different protocol, it doesn’t use ClangdLSPServer but wraps ClangdServer instead. The protocol happens to follow LSP fairly closely (at least in spirit), but doesn’t try to reuse any of our LSP code.

Not sure whether this was the right call, but they’re not likely to switch to any Transport abstraction we add.
This is certainly an option for XCode too, but then the most obvious path is write a wrapper around ClangdServer, and ignore ClangdLSPServer/ClangdMain.

Transport::MessageHandler

I just wonder if this isn’t actually the “reverse” of desired interface. It seems to me that MessageHandler is abstracting clangd implementation details when used in specific Transport implementation. For example it would help if we ever decide to swap clangd for another LSP implementation or clangd-ng. But it doesn’t abstract the transport layer itself. What benefit does it grant instead of calling for example ClangdLSPServer::notify() directly?

Yeah, this is mostly just ad-hoc decoupling, not real abstraction. It buys us some things:

  • transports can be tested without involving the behavior of all of clangd
  • ClangdLSPServer’s notify() can be private (using interface inheritance, or a nested class, or something)
  • transports don’t have to depend on ClangdLSPServer, if we decide to start caring more about layering (interesting for out-of-tree transports!)

ClangdLSPServer::run()

IIUIC purpose of this method is to have abstraction over different Transport implementations with common strategies for program return value (ShutdownRequestReceived) and error handling (return value of Transport.loop()).

Well, the purpose of the current ClangdLSPServer::run() is set up the dispatcher right, and to put the error handling code somewhere.
Maybe it’s totally redundant and main should just call Transport.loop().

(FWIW I think the error handling on exit isn’t an important part of the protocol and we could take some shortcuts, especially in an XPC context).

Yet since we can’t do this for XPC (as xpc_main doesn’t return) we’d have to duplicate the code (~10 LOC, no big deal) but it would mean that it wouldn’t be always possible to just blindly swap implementations of this interface.
In other words it’s a violation of Liskov substitution principle which in itself doesn’t necessarily mean it’s a big deal - I am just wondering if it makes sense to have such abstraction.

So I’m not sure how big a deal this is. It would indeed be cleaner not to have a loop() that sometimes returns and sometimes doesn’t. Pragmatically, I think we can arrange to handle it calling exit instead, I guess.

How do you want clangd shutdown to work in an XPC context? Reading the XPC docs suggest services just run until they get SIGKILL. In which case a Transport::loop() that never returns seems completely appropriate to me: the transport dictates that the stream of events never ends.

This is basically what I meant that it’s difficult to have common abstraction over the “input direction” of transport layer (since Transport::MessageHandler looks more like an abstraction over ClangdLSPServer input handling) and because of that I’d consider different dispatchers instead of Transport::loop().

I’m not sure what distinction is being drawn here, I think we’ll end up with some function like loop() that returns in JSON and not with XPC.
Where is my logic confused?

  • JSONRPCDispatcher does some JSON stuff and some generic “message handling” stuff
  • we want the “message handling” stuff to be common
  • the JSON stuff needs to be split from the message handling stuff, so it can be replaced by XPC (in the sketch I called this part “Transport”)
  • we want a common abstraction (e.g. interface) over the XPC and JSON-specific code
  • XPC needs somewhere to call xpc_main(), which doesn’t return
  • JSON needs every function to return, because it supports clean exit
  • therefore some XPC function doesn’t return, when the corresponding JSON one does
    There are ways around this I guess (call xpc_main on a thread? have the JSON code clean up and call exit()?) but they seem bad worse than just having a transport that never stops looping…

Hi Sam,

Thanks for the response.

I think I understand your intentions better now. The Transport abstraction looks good enough to be used to me so let me just explain what I though are imperfections that I know about (no big deals though). If these are ok with you I’d go ahead and work on it. Another thing is that you guys already have another transport layer implementation so if this abstraction fits nicely 2 out of 3 implementations with XPC being kind of little less aligned I’d consider it still pragmatic to use it.

So Transport isn’t implemented or even really fully fleshed out. We should design it (or something else) to bridge the JSON/XPC divide well, though perfect is always hard.

Our “other transport layer” is really a different protocol, it doesn’t use ClangdLSPServer but wraps ClangdServer instead. The protocol happens to follow LSP fairly closely (at least in spirit), but doesn’t try to reuse any of our LSP code.

Not sure whether this was the right call, but they’re not likely to switch to any Transport abstraction we add.

Got it.

This is certainly an option for XCode too, but then the most obvious path is write a wrapper around ClangdServer, and ignore ClangdLSPServer/ClangdMain.

True. We should be good with just custom transport layer implementation though.

Transport::MessageHandler

I just wonder if this isn’t actually the “reverse” of desired interface. It seems to me that MessageHandler is abstracting clangd implementation details when used in specific Transport implementation. For example it would help if we ever decide to swap clangd for another LSP implementation or clangd-ng. But it doesn’t abstract the transport layer itself. What benefit does it grant instead of calling for example ClangdLSPServer::notify() directly?

Yeah, this is mostly just ad-hoc decoupling, not real abstraction. It buys us some things:

  • transports can be tested without involving the behavior of all of clangd

Oh, I totally missed that. I am sold now!

  • ClangdLSPServer’s notify() can be private (using interface inheritance, or a nested class, or something)
  • transports don’t have to depend on ClangdLSPServer, if we decide to start caring more about layering (interesting for out-of-tree transports!)

Also a fair point.

ClangdLSPServer::run()

IIUIC purpose of this method is to have abstraction over different Transport implementations with common strategies for program return value (ShutdownRequestReceived) and error handling (return value of Transport.loop()).

Well, the purpose of the current ClangdLSPServer::run() is set up the dispatcher right, and to put the error handling code somewhere.
Maybe it’s totally redundant and main should just call Transport.loop().

(FWIW I think the error handling on exit isn’t an important part of the protocol and we could take some shortcuts, especially in an XPC context).

Yet since we can’t do this for XPC (as xpc_main doesn’t return) we’d have to duplicate the code (~10 LOC, no big deal) but it would mean that it wouldn’t be always possible to just blindly swap implementations of this interface.
In other words it’s a violation of Liskov substitution principle which in itself doesn’t necessarily mean it’s a big deal - I am just wondering if it makes sense to have such abstraction.

So I’m not sure how big a deal this is. It would indeed be cleaner not to have a loop() that sometimes returns and sometimes doesn’t. Pragmatically, I think we can arrange to handle it calling exit instead, I guess.

How do you want clangd shutdown to work in an XPC context? Reading the XPC docs suggest services just run until they get SIGKILL. In which case a Transport::loop() that never returns seems completely appropriate to me: the transport dictates that the stream of events never ends.

Check if the client request is exit notification in the XPC message handling C block and conditionally call exit(). Would just need to get access to ClangdLSPServer::ShutdownRequestReceived from there.

This is basically what I meant that it’s difficult to have common abstraction over the “input direction” of transport layer (since Transport::MessageHandler looks more like an abstraction over ClangdLSPServer input handling) and because of that I’d consider different dispatchers instead of Transport::loop().

I’m not sure what distinction is being drawn here, I think we’ll end up with some function like loop() that returns in JSON and not with XPC.
Where is my logic confused?

  • JSONRPCDispatcher does some JSON stuff and some generic “message handling” stuff
  • we want the “message handling” stuff to be common
  • the JSON stuff needs to be split from the message handling stuff, so it can be replaced by XPC (in the sketch I called this part “Transport”)
  • we want a common abstraction (e.g. interface) over the XPC and JSON-specific code
  • XPC needs somewhere to call xpc_main(), which doesn’t return
  • JSON needs every function to return, because it supports clean exit
  • therefore some XPC function doesn’t return, when the corresponding JSON one does
    There are ways around this I guess (call xpc_main on a thread? have the JSON code clean up and call exit()?) but they seem bad worse than just having a transport that never stops looping…

My point of view was that rather than having part of abstraction that is working only for 1 out of 2 implementations it is supposed to abstract I would consider removing it from the interface (less code, less complexity) and keeping it as part of JSON-RPC specific code. But I see the benefit of not having duplicated callbacks to dispatcher registration so I am happy to go with the Transport::loop().

Cheers,

Jan

Hi Jan,

Wondering how your work here is going :slight_smile:
I revived the Transport idea and put together https://reviews.llvm.org/D53286 where everything seems to work.
(There’s some cruft left in JSONRPCDispatcher that can be removed afterwards).

I think this is worthwhile independent of XPC (e.g. for testing ClangdLSPServer) but of course it has to be compatible with what you need.
Let me know if you see any blockers here or if something has changed!

Cheers, Sam

Hi Sam,

My work is unfortunately stuck at the moment - I have been enjoying a sinus infection for almost two weeks now, expecting to go back to work in another week.

I put together crude implementation (lots of TODO everywhere and definitely not finished) using your old proposal. If that’s of any interest you can find it here:
https://reviews.llvm.org/D53290

Since this is very important for us I am definitely going to work on transport layer when I am back at work. I will take a proper look and try to “merge” our approaches.

Cheers,

Jan

Ouch, sorry to hear that! Hope it clears up soon!

I took a look at your patch and I don’t think there’s any divergence here.
D53286 only covers a subset of the work, and the interface changes look much the same in both patches.

The bit that’s left out is moving the remaining JSONRPCDispatcher logic into ClangdLSPServer or similar - your patch is more complete in this respect. But my hope is to try to land this in smaller pieces, so I might try to get reviews for D53286, and then come back for the rest.
By the time you’re back, we should be a lot closer to what you need than we are today. Sound OK?

Get well soon!

Oops, I completely missed your reply. Sorry about that!

I am finally back to work. Just read through https://reviews.llvm.org/D53387 and really like it!

I am resuming my work on Transport layer today or tomorrow.

Cheers,

Jan