On 7/4/19 6:12 PM, Jon Masters wrote:
I think we have identified the root cause of the 32-bit builder
issue.
Many thanks to Paul and Peter for assistance in debugging. Here's my
write-up, and we'll work with the vendor on a suitable mitigation to
workaround any errata:
https://medium.com/@jonmasters_84473/debugging-a-32-bit-fedora-arm-builde...
The hardware vendor have reproduced what I believe to be an errata.
Meanwhile, I've made a test kernel that forces CONFIG_HIGHPTE to off:
https://koji.fedoraproject.org/koji/taskinfo?taskID=36328838
With this kernel, you still get LPAE but leaf level PTEs are not
allocated from high memory any longer. This is because I believe the
errata to be caused by stage 1 page table walks in the guest trapping to
stage 2 (hypervisor) for e.g. Access bit updates on the host. When those
occur, I believe there is a truncation of the guest IPA (guest memory)
address to 32-bits, but only for page table entry walks. Normal
translation faults I think are unaffected by this (TBC).
Normally, we don't allocate PGDs (high level page table pieces) from
high memory (we allocate those from kernel memory caches) but we DO
allocate PTEs specifically from what might be high memory. Except when
we force CONFIG_HIGHPTE to off. The patch I'm using is attached.
It's currently being tested. If it works, I'm curious for input on
temporarily carrying this in Fedora. In theory it means an LPAE system
could starve for PTEs if it has many many processes running, but in
practice I'm willing to bet LPAE is mostly used by Fedora for the 32-bit
builders and that few people would actually complain if we did this.
Jon.
--
Computer Architect | Sent with my Fedora powered laptop