I'm seeing unhandled level 2 translation fault errors when running my experiments. As near as I can tell, they occur on ARM only when trying to load a value from memory, where the address if offset from an address that was previously loaded with an adrp instruction. After migration, however, it appears that the initial address is not reloaded into a register appropriately.
I've only been able to reproduce this bug when I run 50+ applications at once, which makes it difficult for me to provide useful debugging information. I've attached what I can provide:
The binaries that produce the bug
The output of -mllvm -debug-only=stacktransform -mllvm -debug-only=-insert-stackmaps for each binary
The stacktransform.log output (note, it's highly interspersed due to running so many applications at once)
The kernel log showing the faulting programs (also highly interspersed).
Here's a concise snippet of the kernel log showing the error:
I'm seeing
unhandled level 2 translation fault
errors when running my experiments. As near as I can tell, they occur on ARM only when trying to load a value from memory, where the address if offset from an address that was previously loaded with anadrp
instruction. After migration, however, it appears that the initial address is not reloaded into a register appropriately.I've only been able to reproduce this bug when I run 50+ applications at once, which makes it difficult for me to provide useful debugging information. I've attached what I can provide:
-mllvm -debug-only=stacktransform -mllvm -debug-only=-insert-stackmaps
for each binaryHere's a concise snippet of the kernel log showing the error:
Everything is in the attached tarball.
2018-02-26_NPB3.3-SER-C_adrp-info.tar.gz