mcount_return_trampoline: Don't disable interrupts

[mkroening]

This fixes #8 for me.

[tlambertz]

Huh yeah, seems I missed that one. Now wondering how it ever worked?
The latest commit was done a few months after my thesis was complete and I was cleaning up my working directory, might have not tested it fully?

If i recall correctly, the reason for the changes in the latest commit were some changes in libhermit which lead to frequent crashes (tracing interrupt-based filesystem writes?). Something with interrupt at inopportune moments, trashing the stack/flags register.

[mkroening]

Ah, I see. That makes sense. I have only tested with the rust example, and that one always crashes for me on the latest commit without this fix, as explained in #8.

[mkroening]

Ah, nevermind, I did not test this change correctly. I'll rework this.

[mkroening]

Ah, sorry. I must have been not very concentrated. Anyway, I found the issue now. The stack has been handled correctly.

The problem is calling cli in user space, which is not possible. When you tested this in libhermit you might have tested kernel space only. Removing cli makes the example work for me. Why would we need to disable interrupts there in the first place? Any interrupt handler would make sure not to clobber any registers, wouldn't it?

[tlambertz]

I think one issue I tried to fix was that an interrupt was arriving while the kernel was in mcount(). The interrupt in turn called mcount(), which then overwrote the return pointer of the first mcount() with the trampoline, which broke things.

[mkroening]

Ah, so then the problem is the access to RETSTACK? I was trying to reproduce the issue to be sure, but I can't get RustyHermit to build with rftrace.