On 10/2/2022 12:06 pm, Ravi Bangoria wrote:
Hi Jim,
On 10-Feb-22 3:10 AM, Jim Mattson wrote:
On Wed, Feb 9, 2022 at 2:19 AM Like Xu <like.xu.linux@xxxxxxxxx> wrote:
On 4/2/2022 9:01 pm, Jim Mattson wrote:
On Fri, Feb 4, 2022 at 1:33 AM Ravi Bangoria <ravi.bangoria@xxxxxxx> wrote:
On 03-Feb-22 11:25 PM, Jim Mattson wrote:
On Wed, Feb 2, 2022 at 9:18 PM Ravi Bangoria <ravi.bangoria@xxxxxxx> wrote:
Hi Jim,
On 03-Feb-22 9:39 AM, Jim Mattson wrote:
On Wed, Feb 2, 2022 at 2:52 AM Ravi Bangoria <ravi.bangoria@xxxxxxx> wrote:
Perf counter may overcount for a list of Retire Based Events. Implement
workaround for Zen3 Family 19 Model 00-0F processors as suggested in
Revision Guide[1]:
To count the non-FP affected PMC events correctly:
o Use Core::X86::Msr::PERF_CTL2 to count the events, and
o Program Core::X86::Msr::PERF_CTL2[43] to 1b, and
o Program Core::X86::Msr::PERF_CTL2[20] to 0b.
Note that the specified workaround applies only to counting events and
not to sampling events. Thus sampling event will continue functioning
as is.
Although the issue exists on all previous Zen revisions, the workaround
is different and thus not included in this patch.
This patch needs Like's patch[2] to make it work on kvm guest.
IIUC, this patch along with Like's patch actually breaks PMU
virtualization for a kvm guest.
Suppose I have some code which counts event 0xC2 [Retired Branch
Instructions] on PMC0 and event 0xC4 [Retired Taken Branch
Instructions] on PMC1. I then divide PMC1 by PMC0 to see what
percentage of my branch instructions are taken. On hardware that
suffers from erratum 1292, both counters may overcount, but if the
inaccuracy is small, then my final result may still be fairly close to
reality.
With these patches, if I run that same code in a kvm guest, it looks
like one of those events will be counted on PMC2 and the other won't
be counted at all. So, when I calculate the percentage of branch
instructions taken, I either get 0 or infinity.
Events get multiplexed internally. See below quick test I ran inside
guest. My host is running with my+Like's patch and guest is running
with only my patch.
Your guest may be multiplexing the counters. The guest I posited does not.
It would be helpful if you can provide an example.
Perf on any current Linux distro (i.e. without your fix).
The patch for errata #1292 (like most hw issues or vulnerabilities) should be
applied to both the host and guest.
As I'm sure you are aware, guests are often not patched. For example,
It's true. What a real world.
we have a lot of Debian-9 guests running on Milan, despite the fact
that it has to be booted with "nopcid" due to a bug introduced on
4.9-stable. We submitted the fix and notified Debian about a year ago,
but they have not seen fit to cut a new kernel. Do you think they will
cut a new kernel for this patch?
Indeed, thanks for your user stories.
For non-patched guests on a patched host, the KVM-created perf_events
will be true for is_sampling_event() due to get_sample_period().
I think we (KVM) have a congenital defect in distinguishing whether guest
counters are used in counting mode or sampling mode, which is just
a different use of pure software.
I have no idea what you are saying. However, when kvm sees a guest
counter used in sampling mode, it will just request a PERF_TYPE_RAW
perf event with the INT bit set in 'config.' If it sees a guest
The counters work very simply: increments until it overflows.
The use of INT bit is not related to counting or sampling mode.
A pmu driver can set the INT bit, but set a very small ctr value and not
expect it to overflow, and it can be used for counting mode as well, right?
We don't know under what circumstances the overcount will occur, maybe
it's related to the INT bit and maybe bot, but absolutely it's nothing to do with
the software check is_sampling_event().
counter used in counting mode, it will either request a PERF_TYPE_RAW
perf event or a PERF_TYPE_HARDWARE perf event, depending on whether or
not it finds the requested event in amd_event_mapping[].
I hope that you are not saying that kvm's *thread-pinned* perf events
are not being multiplexed at the host level, because that completely
breaks PMU virtualization.
IIUC, multiplexing happens inside the guest.
I'm not sure that multiplexing is the answer. Extrapolation may
introduce greater imprecision than the erratum.
If you run the same test on the patched host, the PMC2 will be
used in a multiplexing way. This is no different.
If you count something like "instructions retired" three ways:
1) Unfixed counter
2) PMC2 with the fix
3) Multiplexed on PMC2 with the fix
Is (3) always more accurate than (1)?
Since Ravi has gone dark, I will answer my own question.
Sorry about the delay. I was discussing this internally with hw folks.
For better reproducibility, I simplified his program to:
int main() { return 0;}
On an unpatched Milan host, I get instructions retired between 21911
and 21915. I get branch instructions retired between 5565 and 5566. It
does not matter if I count them separately or at the same time.
After applying v3 of Ravi's patch, if I try to count these events at
the same time, I get 36869 instructions retired and 4962 branch
instructions on the first run. On subsequent runs, perf refuses to
count both at the same time. I get branch instructions retired between
5565 and 5567, but no instructions retired. Instead, perf tells me:
Some events weren't counted. Try disabling the NMI watchdog:
echo 0 > /proc/sys/kernel/nmi_watchdog
perf stat ...
echo 1 > /proc/sys/kernel/nmi_watchdog
If I just count one thing at a time (on the patched kernel), I get
between 21911 and 21916 instructions retired, and I get between 5565
and 5566 branch instructions retired.
I don't know under what circumstances the unfixed counters overcount
or by how much. However, for this simple test case, the fixed PMC2
yields the same results as any unfixed counter. Ravi's patch, however
makes counting two of these events simultaneously either (a)
impossible, or (b) highly inaccurate (from 10% under to 68% over).
In further discussions with our hardware team, I am given to understand
that the conditions under which the overcounting can happen, is quite
rare. In my tests, I've found that the patched vs. unpatched cases are
not significantly different to warrant the restriction introduced by
That's cute and thank you both.
I hope we can come to this conclusion before the code is committed.
But the kvm's patch may have made those PMU driver developers who read the
erratum #1292 a little happier, wouldn't it ?
this fix. I have requested Peter to hold off pushing this fix.
The loss of accuracy is due to a reduction in the number of trustworthy counters,
not to these two workaround patches. Any multiplexing (whatever on the host or
the guest) will result in a loss of accuracy. Right ?
Yes, that's my point. Fixing one inaccuracy by using a mechanism that
introduces another inaccuracy only makes sense if the inaccuracy you
are fixing is worse than the inaccuracy you are introducing. That does
Couldn't agree more, and in response to similar issues,
we may adopt a quantitative-first strategy in the future.
not appear to be the case here, but I am not privy to all of the
details of this erratum.
Thanks,
Ravi
