On 4/27/2024 11:04 AM, Mingwei Zhang wrote:
On Fri, Apr 26, 2024 at 12:46 PM Sean Christopherson <seanjc@xxxxxxxxxx> wrote:
On Fri, Apr 26, 2024, Kan Liang wrote:
Optimization 4
allows the host side to immediately profiling this part instead of
waiting for vcpu to reach to PMU context switch locations. Doing so
will generate more accurate results.
If so, I think the 4 is a must to have. Otherwise, it wouldn't honer the
definition of the exclude_guest. Without 4, it brings some random blind
spots, right?
+1, I view it as a hard requirement. It's not an optimization, it's about
accuracy and functional correctness.
Well. Does it have to be a _hard_ requirement? no? The irq handler
triggered by "perf record -a" could just inject a "state". Instead of
immediately preempting the guest PMU context, perf subsystem could
allow KVM defer the context switch when it reaches the next PMU
context switch location.
This is the same as the preemption kernel logic. Do you want me to
stop the work immediately? Yes (if you enable preemption), or No, let
me finish my job and get to the scheduling point.
Implementing this might be more difficult to debug. That's my real
concern. If we do not enable preemption, the PMU context switch will
only happen at the 2 pairs of locations. If we enable preemption, it
could happen at any time.
IMO I don't prefer to add a switch to enable/disable the preemption. I
think current implementation is already complicated enough and
unnecessary to introduce an new parameter to confuse users. Furthermore,
the switch could introduce an uncertainty and may mislead the perf user
to read the perf stats incorrectly. As for debug, it won't bring any
difference as long as no host event is created.
What _is_ an optimization is keeping guest state loaded while KVM is in its
run loop, i.e. initial mediated/passthrough PMU support could land upstream with
unconditional switches at entry/exit. The performance of KVM would likely be
unacceptable for any production use cases, but that would give us motivation to
finish the job, and it doesn't result in random, hard to diagnose issues for
userspace.
That's true. I agree with that.
Do we want to preempt that? I think it depends. For regular cloud
usage, we don't. But for any other usages where we want to prioritize
KVM/VMM profiling over guest vPMU, it is useful.
My current opinion is that optimization 4 is something nice to have.
But we should allow people to turn it off just like we could choose to
disable preempt kernel.
The exclude_guest means everything but the guest. I don't see a reason
why people want to turn it off and get some random blind spots.
