On 2022/12/02 8:13, Marc Zyngier wrote:
On Thu, 01 Dec 2022 17:26:08 +0000,
Akihiko Odaki <akihiko.odaki@xxxxxxxxxx> wrote:
On 2022/12/01 20:06, Marc Zyngier wrote:
On Thu, 01 Dec 2022 10:49:11 +0000,
Akihiko Odaki <akihiko.odaki@xxxxxxxxxx> wrote:
Thanks for looking into this.
M2 MacBook Air has mismatched CCSIDR associativity bits, which makes the
bits a KVM vCPU sees inconsistent when migrating.
Can you describe the actual discrepancy? Is that an issue between the
two core types? In which case, nothing says that these two cluster
should have the same cache topology.
Yes, the processor has big.LITTLE configuration.
On the processor, the valid CSSELR values are 0 (L1D), 1 (L1I), 3
(L2D). For each CSSELR values, each cluster has:
- 0x700FE03A, 0x203FE01A, 0x70FFE07B
- 0x701FE03A, 0x203FE02A, 0x73FFE07B
This is a perfectly valid configuration. The architecture doesn't
place any limitation on how different or identical the cache
hierarchies are from the PoV of each CPU. Actually, most big-little
systems show similar differences across their clusters.
It also makes QEMU fail restoring the vCPU registers because QEMU saves
and restores all of the registers including CCSIDRs, and if the vCPU
migrated among physical CPUs between saving and restoring, it tries to
restore CCSIDR values that mismatch with the current physical CPU, which
causes EFAULT.
Well, QEMU will have plenty of other problems, starting with MIDRs,
which always reflect the physical one. In general, KVM isn't well
geared for VMs spanning multiple CPU types. It is improving, but there
is a long way to go.
On M2 MacBook Air, I have seen no other difference in standard ID
registers and CCSIDRs are exceptions. Perhaps Apple designed this way
so that macOS's Hypervisor can freely migrate vCPU, but I can't assure
that without more analysis. This is still enough to migrate vCPU
running Linux at least.
I guess that MacOS hides more of the underlying HW than KVM does. And
KVM definitely doesn't hide the MIDR_EL1 registers, which *are*
different between the two clusters.
It seems KVM stores a MIDR value of a CPU and reuse it as "invariant"
value for ioctls while it exposes the MIDR value each physical CPU owns
to vCPU.
This may be a problem worth fixing. My understanding is that while there
is no serious application which requires vCPU migration among physical
clusters, crosvm uses KVM on big.LITTLE processors by pinning vCPU to
physical CPU, and it is a real-world application which needs to be
supported.
For an application like crosvm, you would expect the vCPU thread gets
the MIDR value of the physical CPU which the thread is pinned to when it
calls ioctl, but it can get one of another arbitrary CPU in reality.
Fixing this problem will pose two design questions:
1. Should it expose a value consistent among clusters?
For example, we can change the KVM initialization code so that it
initializes VPIDR with the value stored as "invariant". This would help
migrating vCPU among clusters, but if you pin each vCPU thread to a
distinct phyiscal CPU, you may rather want the vCPU to see the MIDR
value specific to each physical CPU and to apply quirks or tuning
parameters according to the value.
2. Should it be invariant or variable?
Fortunately making it variable is easy. Arm provides VPIDR_EL1 register
to specify the value exposed as MPIDR_EL0 so there is no trapping cost.
...or we may just say the value of MPIDR_EL0 (and possibly other
"invariant" registers) exposed via ioctl are useless and deprecated.
Trap CCSIDRs if there are CCSIDR value msimatches, and override the
associativity bits when handling the trap.
TBH, I'd rather we stop reporting this stuff altogether.
There is nothing a correctly written arm64 guest should do with any of
this (this is only useful for set/way CMOs, which non-secure SW should
never issue). It would be a lot better to expose a virtual topology
(one set, one way, one level). It would also save us from the CCSIDRX
silliness.
The only complexity would be to still accept different topologies from
userspace so that we can restore a VM saved before this virtual
topology.
Another (minor) concern is that trapping relevant registers may cost
too much. Currently KVM traps CSSELR and CCSIDR accesses with
HCR_TID2, but HCR_TID2 also affects CTR_EL0.
It will have an additional impact (JITs, for example, will take a hit
if they don't cache that value), but this is pretty easy to mitigate
if it proves to have too much of an impact. We already have a bunch of
fast-paths for things that we want to emulate more efficiently, and
CTR_EL0 could be one of them,
Although I'm not sure if the register is referred frequently, Arm
introduced FEAT_EVT to trap CSSELR and CSSIDR but not CTR_EL0 so
there may be some case where trapping CTR_EL0 is not
tolerated. Perhaps Arm worried that a userspace application may read
CTR_EL0 frequently.
FEAT_EVT is one of these "let's add random traps" extensions,
culminating in FEAT_FGT. Having FEAT_EVT would make it more efficient,
but we need to support this for all revisions of the architecture.
So let's first build on top of HCR_EL2.TID2, and only then once we
have an idea of the overhead add support for HCR_EL2.TID4 for the
systems that have FEAT_EVT.
That sounds good, I'll write a new series according to this idea.
Regards,
Akihiko Odaki
If you think the concern on VM restoration you mentioned and the
trapping overhead is tolerable, I'll write a new, much smaller patch
accordingly.
That would great, thanks. There are a number of gotchas around that
(like the 32bit stuff that already plays the emulation game), but this
is the right time to start and have something in 6.3 if you keep to it!
M.
_______________________________________________
kvmarm mailing list
kvmarm@xxxxxxxxxxxxxxxxxxxxx
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
