panic!ing twice crashes PG with "Illegal instruction"

[auterium]

OS

Ubuntu 18.04 (Windows Subsystem for Linux)

Code

panicking example package

Issue

Rust panic!() is caught and handled properly on the first call, but if called again, the server crashes:

Here's PG logs:

That little extra weird stuff

This also happens when using error!() macro, but it requires 3 executions:

Here;s PG logs:

[bluejekyll]

Error, fatal and panic all cause PG to abort the transaction, where panic will cause a full restart. The illegal obstruction is concerning though.

Is this a double panic? Or is it a panic, and then a different transaction also panicking?

[auterium]

Unless psql is not setting AUTOCOMMIT then yes, they are separate transactions panicking

[bluejekyll]

This is happening on Windows? I wonder if this behavior can be reproduced on a vanilla Linux box.

[auterium]

Same error happens in a Docker container running PG10

[bluejekyll]

I haven’t looked into this, yet, but what is the expected behavior of Postgres when it panics? It might do exactly this. It would be interesting if we had a test case that showed this with and without pg-extend. That would require a vanilla C extension I think.

[auterium]

Looking at PG source code, I found this:

The important part is:

If elevel >= ERROR, the call will not return; we try to inform the compiler of that via pg_unreachable()

So, if the log level is ERROR or higher, there will be a different behaviour (no return). Perhaps this might explain the weird behaviour?

[bluejekyll]

Yes. That’s what I was trying to explain. Error and higher log statements have the side effect of killing the transaction/process/server, depending on which you use.

It’s really ugly imo, but that’s the way it works.