On 4/25/2024 6:29 PM, Zhao Liu wrote:
On Thu, Apr 25, 2024 at 04:40:10PM +0800, Xiaoyao Li wrote:
Date: Thu, 25 Apr 2024 16:40:10 +0800
From: Xiaoyao Li <xiaoyao.li@xxxxxxxxx>
Subject: Re: [PATCH for-9.1 0/7] target/i386/kvm: Cleanup the kvmclock
feature name
On 4/25/2024 3:17 PM, Zhao Liu wrote:
Hi Xiaoyao,
On Wed, Apr 24, 2024 at 11:57:11PM +0800, Xiaoyao Li wrote:
Date: Wed, 24 Apr 2024 23:57:11 +0800
From: Xiaoyao Li <xiaoyao.li@xxxxxxxxx>
Subject: Re: [PATCH for-9.1 0/7] target/i386/kvm: Cleanup the kvmclock
feature name
On 3/29/2024 6:19 PM, Zhao Liu wrote:
From: Zhao Liu <zhao1.liu@xxxxxxxxx>
Hi list,
This series is based on Paolo's guest_phys_bits patchset [1].
Currently, the old and new kvmclocks have the same feature name
"kvmclock" in FeatureWordInfo[FEAT_KVM].
When I tried to dig into the history of this unusual naming and fix it,
I realized that Tim was already trying to rename it, so I picked up his
renaming patch [2] (with a new commit message and other minor changes).
13 years age, the same name was introduced in [3], and its main purpose
is to make it easy for users to enable/disable 2 kvmclocks. Then, in
2012, Don tried to rename the new kvmclock, but the follow-up did not
address Igor and Eduardo's comments about compatibility.
Tim [2], not long ago, and I just now, were both puzzled by the naming
one after the other.
The commit message of [3] said the reason clearly:
When we tweak flags involving this value - specially when we use "-",
we have to act on both.
So you are trying to change it to "when people want to disable kvmclock,
they need to use '-kvmclock,-kvmclock2' instead of '-kvmclock'"
IMHO, I prefer existing code and I don't see much value of differentiating
them. If the current code puzzles you, then we can add comment to explain.
It's enough to just enable kvmclock2 for Guest; kvmclock (old) is
redundant in the presence of kvmclock2.
So operating both feature bits at the same time is not a reasonable
choice, we should only keep kvmclock2 for Guest. It's possible because
the oldest linux (v4.5) which QEMU i386 supports has kvmclock2.
who said the oldest Linux QEMU supports is from 4.5? what about 2.x kernel?
For Host (docs/system/target-i386.rst).
Besides, not only the Linux guest, whatever guest OS is, it will be broken
if the guest is using kvmclock and QEMU starts to drop support of kvmclock.
I'm not aware of any minimum version requirements for Guest supported
by KVM, but there are no commitment.
the common commitment is at least keeping backwards compatibility.
So, again, hard NAK to drop the support of kvmclock. It breaks existing
guests that use kvmclock. You cannot force people to upgrade their existing
VMs to use kvmclock2 instead of kvmclock.
I agree, legacy kvmclock can be left out, if the old kernel does not
support kvmclock2 and strongly requires kvmclock, it can be enabled
using 9.1 and earlier machines or legacy-kvmclock, as long as Host still
supports it.
What's the gap in handling it this way? especially considering that
kvmclock2 was introduced in v2.6.35, and earlier kernels are no longer
maintained. The availability of the PV feature requires compatibility
for both Host and Guest.
Anyway, the above discussion here is about future plans, and this series
does not prevent any Guest from ignoring kvmclock2 in favor of kvmclock.
What this series is doing, i.e. separating the current two kvmclock and
ensuring CPUID compatibility via legacy-kvmclock, could balance the
compatibility requirements of the ancient (unmaintained kernel) with
the need for future feature changes.
You introduce a user-visible change that people need to use
"-kvmclock,-kvmclock2" to totally disable kvmclock.
The only difference between kvmclock and kvmclock2 is the MSR index. And
from users' perspective, they don't care this difference. The existing
usage is simple. When I want kvmclock, use "+kvmclock". When I don't
want it, use "-kvmclock".
You are complicating things and I don't see a strong reason that we have
to do it.
Thanks,
Zhao
