Respect requested message limits within a single MessageProducer in BinderTransport.

[akandratovich]

Add check for request message count in binder grpc inbound stream.

Without check if messages are available and requested, producer provides everything that is available ignoring the request count. It leads to call listener API violation - it's possible that more messages will be provided than requested.

[linux-foundation-easycla]

The committers listed above are authorized under a signed CLA.

• ✅ login: akandratovich / name: Andrei Kandratovich (8daca09, 0b4f165)

[markb74]

Looks good to me, but wait for ejona since the new test affects other transport as well.

[ejona86]

I believe flowControlPushBack() would have ordinarily caught this bug, but that test assumes end-to-end flow control which isn't in-place for Binder so the test is disabled. This new test is inherently timing-sensitive (just like flowControlPushBack()) and might not catch bugs. @akandratovich, did you confirm the test fails before the bug fix?

[markb74]

@ejona86: I verified it failed before the fix, but that's a good point about timing.

It might help to have the client halfClose, and then awaitHalfClose on the server.

[markb74]

Sorry, awaitHalfClose won't help, but adding a doPingPong() call should help.

[ejona86]

Ah, yes. doPingPong() is how I solved that before.

[akandratovich]

I believe flowControlPushBack() would have ordinarily caught this bug, but that test assumes end-to-end flow control which isn't in-place for Binder so the test is disabled. This new test is inherently timing-sensitive (just like flowControlPushBack()) and might not catch bugs. @akandratovich, did you confirm the test fails before the bug fix?

I tried enabling flowControlPushBack(). It feels a bit heuristic for me and I actually managed to tweak buffers enough to make it passing, but ~5% flaky (without fix it fails 100%). But the test is supposed to validate over-rpc flow control and binder doesn't do that. So we decided to make a new test with less scope instead of trying to cheat with flowControlPushBack().