Robert O'Callahan

> would it be possible to add a mode to rr (or is there one already?) that just runs the instrumented command in chaos mode & doesn't do anything else...

https://github.com/rr-debugger/rr/blob/815eae2f4de45c990c92dcf12da818dcf816b7ea/src/Scheduler.cc#L362 https://robert.ocallahan.org/2016/02/introducing-rr-chaos-mode.html > How would Pernosco aide in diagnosing this cross-process IPC issue more effectively than naiive gdb debugging? If this bug is some weird kernel issue and you already...

Are you using edge-triggered epoll or level-triggered?

That could explain why you have data in the buffer and yet epoll did not unblock. Are you sure your design here is correct? Keep in mind that written data...

> Maybe I need to see if I can write a trivial program to reproduce this using `rr`. That would be great if you can. One thing you could do...

I added a trivial rr test for edge-triggered epoll here: https://github.com/rr-debugger/rr/commit/fd5178713bfa8d25af7b9858309a542a0631aed0 It passes with kernel 5.14. I wonder about the syscallbuf code here. It calls `epoll_wait()` with a zero timeout....

> Maybe an `EINTR` return followed by a retry doesn't work for edge-triggered `epoll_wait()`s? Seems to me that if the `epoll_wait()` returns `EINTR` then there should have been no edge-trigger...

> The trace is a bit big so I put it on Google Drive (can switch to Dropbox if that's more convenient): https://drive.google.com/file/d/1skzdhNHOALN-dEuHSFB1p8RFqIVNRzLI/view?usp=sharing I sent an access request. What version...

I just built capnp master and it did not install `lexer.capnp.h` :-(

ok I hacked things enough to build and run it locally.