clear excl monitor on context switches and data aborts

[kib]

Summary

I cannot find a mention in the documentation that ERET automatically clears exclusive monitors. If I am right, not clearing the monitor could result in the aliasing of the stxr success. Practically, this is not a likely outcome since selecting new thread includes taking at least thread locks, but I believe that doing clrex could avoid surprises.

Also, after the data abort handler call, excl monitor is in the undefined state, better clear it.

[kib]

Ok, so I found a statement that ERET clears the local monitor, but I now think that the reference to ERET in my description is bogus. At least now, it seems that context switch in kernel could happen without neccesary return from the exception frame, e.g. by intr_event_handle()->critical_exit()->mi_switch() path.

So I still think that the change is correct.

[emaste]

sys/arm64/include/cpufunc.h

117 ↗

(On Diff #8761)

I don't like the ? in this comment - we don't know?

[kib]

After discussion with jhb, I was convinced that we always lock a thread when doing context switch. This removes the need for clrex in swtch.S. Still, I think that the data abort handler must clear local monitor.

[kib]

sys/arm64/include/cpufunc.h

117 ↗

(On Diff #8810)

Indeed, I do not know would compiler move the instruction around accesses without the "memory" constraint, which makes the asm block a compiler barrier. I think that it is more likely needed than not.

[andrew]

sys/arm64/arm64/trap.c

158 ↗

(On Diff #8810)

It would pay to add a comment here explaining the exclusive monitor state is unknown on data abort.