Re: [RFC patch 00/15] entry: Provide generic implementation for host and guest entry/exit work

From: Christian Borntraeger
Date: Tue Sep 24 2019 - 02:50:32 EST




On 19.09.19 17:03, Thomas Gleixner wrote:
> When working on a way to move out the posix cpu timer expiry out of the
> timer interrupt context, I noticed that KVM is not handling pending task
> work before entering a guest. A quick hack was to add that to the x86 KVM
> handling loop. The discussion ended with a request to make this a generic
> infrastructure possible with also moving the per arch implementations of
> the enter from and return to user space handling generic.
>
> https://lore.kernel.org/r/89E42BCC-47A8-458B-B06A-D6A20D20512C@xxxxxxxxxxxxxx
>
> You asked for it, so don't complain that you have to review it :)
>
> The series implements the syscall enter/exit and the general exit to
> userspace work handling along with the pre guest enter functionality.
>
> The series converts x86 and ARM64. x86 is fully tested including selftests
> etc. ARM64 is only compile tested for now as my only ARM64 testbox is not
> available right now.

It seems that s390x would also need to look into TIF_NOTIFY_PENDING before
entering a KVM guest. Given that the s390x entry path is still in assembler
this might not be something to do quickly.

Would it make sense to actually start with a minimal solution (e.g. one that
provides notify_resume_pending like your original patch) as a fix. That would
also be simple to backport. And then we can do the proper rework on top.

Or do we consider anything that depends on TIF_NOTIFY_PENDING before entering
a guest as not important enough for stable?
After all the vcpu_run ioctl almost never returns to userspace and nothing
was obviously broken.

Another question: Are there callbacks due to TIF_NOTIFY_PENDING that should
NOT happen as long as we stay in the vpcu loop?