[PATCH RESEND v13 00/10] Guest Last Branch Recording Enabling

From: Like Xu
Date: Thu Oct 29 2020 - 23:56:39 EST


Hi All,

PMU features matter. Please help review this rebased version for the
next kernel release and with this patch set, the following error will be
gone forever and cloud developers can better understand their
programs with less profiling overhead:

$ perf record -b lbr ${WORKLOAD}
or $ perf record --call-graph lbr ${WORKLOAD}
Error:
cycles: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'

We already have host perf support to implement guest LBR, please
check more details in each commit and feel free to test and comment.

v12->v13 Changelog:
- remove perf patches since they're merged already;
- add a minor patch to refactor MSR_IA32_DEBUGCTLMSR set/get handler;
- add a minor patch to expose vmx_set_intercept_for_msr();
- add a minor patch to adjust features visibility via IA32_PERF_CAPABILITIES;
- spilt the big patch to three pieces (0004-0006) for better understanding and review;
- make the LBR_FMT exposure patch as the last step to enable guest LBR;

Previous:
https://lore.kernel.org/kvm/20200613080958.132489-1-like.xu@xxxxxxxxxxxxxxx/

---

The last branch recording (LBR) is a performance monitor unit (PMU)
feature on Intel processors that records a running trace of the most
recent branches taken by the processor in the LBR stack. This patch
series is going to enable this feature for plenty of KVM guests.

The user space could configure whether it's enabled or not for each
guest via MSR_IA32_PERF_CAPABILITIES msr. As a first step, a guest
could only enable LBR feature if its cpu model is the same as the
host since the LBR feature is still one of model specific features.

If it's enabled on the guest, the guest LBR driver would accesses the
LBR MSR (including IA32_DEBUGCTLMSR and records MSRs) as host does.
The first guest access on the LBR related MSRs is always interceptible.
The KVM trap would create a special LBR event (called guest LBR event)
which enables the callstack mode and none of hardware counter is assigned.
The host perf would enable and schedule this event as usual.

Guest's first access to a LBR registers gets trapped to KVM, which
creates a guest LBR perf event. It's a regular LBR perf event which gets
the LBR facility assigned from the perf subsystem. Once that succeeds,
the LBR stack msrs are passed through to the guest for efficient accesses.
However, if another host LBR event comes in and takes over the LBR
facility, the LBR msrs will be made interceptible, and guest following
accesses to the LBR msrs will be trapped and meaningless.

Because saving/restoring tens of LBR MSRs (e.g. 32 LBR stack entries) in
VMX transition brings too excessive overhead to frequent vmx transition
itself, the guest LBR event would help save/restore the LBR stack msrs
during the context switching with the help of native LBR event callstack
mechanism, including LBR_SELECT msr.

If the guest no longer accesses the LBR-related MSRs within a scheduling
time slice and the LBR enable bit is unset, vPMU would release its guest
LBR event as a normal event of a unused vPMC and the pass-through
state of the LBR stack msrs would be canceled.

---

LBR testcase:
echo 1 > /proc/sys/kernel/watchdog
echo 25 > /proc/sys/kernel/perf_cpu_time_max_percent
echo 5000 > /proc/sys/kernel/perf_event_max_sample_rate
echo 0 > /proc/sys/kernel/perf_cpu_time_max_percent
./perf record -b ./br_instr a

- Perf report on the host:
Samples: 72K of event 'cycles', Event count (approx.): 72512
Overhead Command Source Shared Object Source Symbol Target Symbol Basic Block Cycles
12.12% br_instr br_instr [.] cmp_end [.] lfsr_cond 1
11.05% br_instr br_instr [.] lfsr_cond [.] cmp_end 5
8.81% br_instr br_instr [.] lfsr_cond [.] cmp_end 4
5.04% br_instr br_instr [.] cmp_end [.] lfsr_cond 20
4.92% br_instr br_instr [.] lfsr_cond [.] cmp_end 6
4.88% br_instr br_instr [.] cmp_end [.] lfsr_cond 6
4.58% br_instr br_instr [.] cmp_end [.] lfsr_cond 5

- Perf report on the guest:
Samples: 92K of event 'cycles', Event count (approx.): 92544
Overhead Command Source Shared Object Source Symbol Target Symbol Basic Block Cycles
12.03% br_instr br_instr [.] cmp_end [.] lfsr_cond 1
11.09% br_instr br_instr [.] lfsr_cond [.] cmp_end 5
8.57% br_instr br_instr [.] lfsr_cond [.] cmp_end 4
5.08% br_instr br_instr [.] lfsr_cond [.] cmp_end 6
5.06% br_instr br_instr [.] cmp_end [.] lfsr_cond 20
4.87% br_instr br_instr [.] cmp_end [.] lfsr_cond 6
4.70% br_instr br_instr [.] cmp_end [.] lfsr_cond 5

Conclusion: the profiling results on the guest are similar to that on the host.

Like Xu (10):
KVM: x86: Move common set/get handler of MSR_IA32_DEBUGCTLMSR to VMX
KVM: x86/vmx: Make vmx_set_intercept_for_msr() non-static and expose it
KVM: x86/pmu: Use IA32_PERF_CAPABILITIES to adjust features visibility
KVM: vmx/pmu: Clear PMU_CAP_LBR_FMT when guest LBR is disabled
KVM: vmx/pmu: Create a guest LBR event when vcpu sets DEBUGCTLMSR_LBR
KVM: vmx/pmu: Pass-through LBR msrs to when the guest LBR event is ACTIVE
KVM: vmx/pmu: Reduce the overhead of LBR pass-through or cancellation
KVM: vmx/pmu: Emulate legacy freezing LBRs on virtual PMI
KVM: vmx/pmu: Expose LBR_FMT in the MSR_IA32_PERF_CAPABILITIES
KVM: vmx/pmu: Release guest LBR event via lazy release mechanism

arch/x86/kvm/pmu.c | 12 +-
arch/x86/kvm/pmu.h | 5 +
arch/x86/kvm/vmx/capabilities.h | 22 ++-
arch/x86/kvm/vmx/pmu_intel.c | 292 +++++++++++++++++++++++++++++++-
arch/x86/kvm/vmx/vmx.c | 52 +++++-
arch/x86/kvm/vmx/vmx.h | 28 +++
arch/x86/kvm/x86.c | 15 +-
7 files changed, 400 insertions(+), 26 deletions(-)

--
2.21.3