On 7/16/2021 5:07 PM, Jim Mattson wrote:
On Fri, Jul 16, 2021 at 12:00 PM Liang, Kan <kan.liang@xxxxxxxxxxxxxxx> wrote:
On 7/16/2021 1:02 PM, Jim Mattson wrote:
On Fri, Jul 16, 2021 at 1:54 AM Zhu Lingshan <lingshan.zhu@xxxxxxxxx> wrote:
The guest Precise Event Based Sampling (PEBS) feature can provide an
architectural state of the instruction executed after the guest instruction
that exactly caused the event. It needs new hardware facility only available
on Intel Ice Lake Server platforms. This patch set enables the basic PEBS
feature for KVM guests on ICX.
We can use PEBS feature on the Linux guest like native:
# echo 0 > /proc/sys/kernel/watchdog (on the host)
# perf record -e instructions:ppp ./br_instr a
# perf record -c 100000 -e instructions:pp ./br_instr a
To emulate guest PEBS facility for the above perf usages,
we need to implement 2 code paths:
1) Fast path
This is when the host assigned physical PMC has an identical index as the
virtual PMC (e.g. using physical PMC0 to emulate virtual PMC0).
This path is used in most common use cases.
2) Slow path
This is when the host assigned physical PMC has a different index from the
virtual PMC (e.g. using physical PMC1 to emulate virtual PMC0) In this case,
KVM needs to rewrite the PEBS records to change the applicable counter indexes
to the virtual PMC indexes, which would otherwise contain the physical counter
index written by PEBS facility, and switch the counter reset values to the
offset corresponding to the physical counter indexes in the DS data structure.
The previous version [0] enables both fast path and slow path, which seems
a bit more complex as the first step. In this patchset, we want to start with
the fast path to get the basic guest PEBS enabled while keeping the slow path
disabled. More focused discussion on the slow path [1] is planned to be put to
another patchset in the next step.
Compared to later versions in subsequent steps, the functionality to support
host-guest PEBS both enabled and the functionality to emulate guest PEBS when
the counter is cross-mapped are missing in this patch set
(neither of these are typical scenarios).
I'm not sure exactly what scenarios you're ruling out here. In our
environment, we always have to be able to support host-level
profiling, whether or not the guest is using the PMU (for PEBS or
anything else). Hence, for our *basic* vPMU offering, we only expose
two general purpose counters to the guest, so that we can keep two
general purpose counters for the host. In this scenario, I would
expect cross-mapped counters to be common. Are we going to be able to
use this implementation?
Let's say we have 4 GP counters in HW.
Do you mean that the host owns 2 GP counters (counter 0 & 1) and the
guest own the other 2 GP counters (counter 2 & 3) in your envirinment?
We did a similar implementation in V1, but the proposal has been denied.
https://lore.kernel.org/kvm/20200306135317.GD12561@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx/
It's the other way around. AFAIK, there is no architectural way to
specify that only counters 2 and 3 are available, so we have to give
the guest counters 0 and 1.
How about the host? Can the host see all 4 counters?
For the current proposal, both guest and host can see all 4 GP counters.
The counters are shared.
I don't understand how that can work. If the host programs two
counters, how can you give the guest four counters?
The guest cannot know the availability of the counters. It may requires
a counter (e.g., counter 0) which may has been used by the host. Host
may provides another counter (e.g., counter 1) to the guest. This is the
case described in the slow path. For this case, we have to modify the
guest PEBS record. Because the counter index in the PEBS record is 1,
while the guest perf driver expects 0.
If we reserve counters 0 and 1 for the guest, this is not a problem
(assuming we tell the guest it only has two counters). If we don't
statically partition the counters, I don't see how you can ensure that
the guest behaves as architected. For example, what do you do when the
guest programs four counters and the host programs two?
Ideally, we should do multiplexing if the guest requires four and the
host requires two. But I doubt this patch set implements the
multiplexing, because the multiplexing should be part of the slow path,
which will be supported in the next step.
Could you please share more details regarding your environment?
How do you handle the case that guest programs two counters and the host
programs four counters?
If counter 0 is available, guests can use counter 0. That's the fast
path. I think the fast path should be more common even both host and
guest are profiling. Because except for some specific events, we may
move the host event to the counters which are not required by guest if
we have enough resources.
And if you don't have enough resources?
As my understanding, multiplexing should be the only choice if we don't
have enough resources.
Thanks,
Kan
