Hi Marc,
On 2020/3/5 4:33, Marc Zyngier wrote:
This (now shorter) series expands the existing GICv4 support to deal
with the new GICv4.1 architecture, which comes with a set of major
improvements compared to v4.0:
- One architectural doorbell per vcpu, instead of one doorbell per VLPI
- Doorbell entirely managed by the HW, with an "at most once" delivery
guarantee per non-residency phase and only when requested by the
hypervisor
- A shared memory scheme between ITSs and redistributors, allowing for an
optimised residency sequence (the use of VMOVP becomes less frequent)
- Support for direct virtual SGI delivery (the injection path still involves
the hypervisor), at the cost of losing the active state on SGIs. It
shouldn't be a big deal, but some guest operating systems might notice
(Linux definitely won't care).
On the other hand, public documentation is not available yet, so that's a
bit annoying...
The series is roughly organised in 3 parts:
(0) Fixes
(1) v4.1 doorbell management
(2) Virtual SGI support
(3) Plumbing of virtual SGIs in KVM
Notes:
- The whole thing is tested on a FVP model, which can be obtained
free of charge on ARM's developer website. It requires you to
create an account, unfortunately... You'll need a fix for the
devicetree that is in the kernel tree (should be merged before
the 5.6 release).
- This series has uncovered a behaviour that looks like a HW bug on
the Cavium ThunderX (aka TX1) platform. I'd very much welcome some
clarification from the Marvell/Cavium folks on Cc, as well as an
official erratum number if this happens to be an actual bug.
[v3 update]
People have ignored for two months now, and it is fairly obvious
that support for this machine is slowly bit-rotting. Maybe I'll
drop the patch and instead start the process of removing all TX1
support from the kernel (we'd certainly be better off without it).
[v4 update]
TX1 is now broken in mainline, and nobody cares. Make of this what
you want.
- I'm extremely grateful for Zenghui Yu's huge effort in carefully
reviewing this rather difficult series (if we ever get to meet
face to face, drinks are definitely on me!).
It's a pleasure to review this work and it's pretty useful for
understanding how Linux works as a GICv4.1-capable hypervisor.
Yay, cheers ;-)!
I'll go through the v4.1 spec one more time before the final
review of this series, as we still have plenty of time to do
some reviews (and even some tests) before the 5.7 MW.
- Unless someone cries wolf, I plan to take this into 5.7.
Good news!
Thanks,
Zenghui
