'perf kvm stat report/record' generates a statistical analysis of KVM
events and can be used to analyze guest exit reasons. This patch tries
to add stat support on riscv.
Map the return value of trace_kvm_exit() to the specific cause of the
exception, and export it to userspace.
It records on two available KVM tracepoints for riscv: "kvm:kvm_entry"
and "kvm:kvm_exit", and reports statistical data which includes events
handles time, samples, and so on.
Cross compiling perf in X86 environment may encounter issues with missing
libraries and tools. Suggest compiling nativly in RISC-V environment
Simple tests go below:
# ./perf kvm record -e "kvm:kvm_entry" -e "kvm:kvm_exit"
Lowering default frequency rate from 4000 to 2500.
Please consider tweaking /proc/sys/kernel/perf_event_max_sample_rate.
[ perf record: Woken up 18 times to write data ]
[ perf record: Captured and wrote 5.433 MB perf.data.guest (62519 samples)
# ./perf kvm report
31K kvm:kvm_entry
31K kvm:kvm_exit
# ./perf kvm stat record -a
[ perf record: Woken up 3 times to write data ]
[ perf record: Captured and wrote 8.502 MB perf.data.guest (99338 samples) ]
# ./perf kvm stat report --event=vmexit
Event name Samples Sample% Time (ns) Time% Max Time (ns) Min Time (ns) Mean Time (ns)
STORE_GUEST_PAGE_FAULT 26968 54.00% 2003031800 40.00% 3361400 27600 74274
LOAD_GUEST_PAGE_FAULT 17645 35.00% 1153338100 23.00% 2513400 30800 65363
VIRTUAL_INST_FAULT 1247 2.00% 340820800 6.00% 1190800 43300 273312
INST_GUEST_PAGE_FAULT 1128 2.00% 340645800 6.00% 2123200 30200 301990
SUPERVISOR_SYSCALL 1019 2.00% 245989900 4.00% 1851500 29300 241403
LOAD_ACCESS 986 1.00% 671556200 13.00% 4180200 100700 681091
INST_ACCESS 655 1.00% 170054800 3.00% 1808300 54600 259625
HYPERVISOR_SYSCALL 21 0.00% 4276400 0.00% 716500 116000 203638
Changes from v1->v2:
- Rebased on Linux 6.9-rc3.
Changes from v2->v3:
- Add the missing assignment for 'vcpu_id_str' in patch 2.
- Remove parentheses that cause compilation errors
Shenlin Liang (2):
RISCV: KVM: add tracepoints for entry and exit events
perf kvm/riscv: Port perf kvm stat to RISC-V
arch/riscv/kvm/trace.h | 67 ++++++++++++++++
arch/riscv/kvm/vcpu.c | 7 ++
tools/perf/arch/riscv/Makefile | 1 +
tools/perf/arch/riscv/util/Build | 1 +
tools/perf/arch/riscv/util/kvm-stat.c | 79 +++++++++++++++++++
.../arch/riscv/util/riscv_exception_types.h | 35 ++++++++
6 files changed, 190 insertions(+)
create mode 100644 arch/riscv/kvm/trace.h
create mode 100644 tools/perf/arch/riscv/util/kvm-stat.c
create mode 100644 tools/perf/arch/riscv/util/riscv_exception_types.h
--
2.37.2
Gentle ping...
>
> 'perf kvm stat report/record' generates a statistical analysis of KVM
> events and can be used to analyze guest exit reasons. This patch tries
> to add stat support on riscv.
>
> Map the return value of trace_kvm_exit() to the specific cause of the
> exception, and export it to userspace.
>
> It records on two available KVM tracepoints for riscv: "kvm:kvm_entry"
> and "kvm:kvm_exit", and reports statistical data which includes events
> handles time, samples, and so on.
>
> Cross compiling perf in X86 environment may encounter issues with missing
> libraries and tools. Suggest compiling nativly in RISC-V environment
>
> Simple tests go below:
>
> # ./perf kvm record -e "kvm:kvm_entry" -e "kvm:kvm_exit"
> Lowering default frequency rate from 4000 to 2500.
> Please consider tweaking /proc/sys/kernel/perf_event_max_sample_rate.
> [ perf record: Woken up 18 times to write data ]
> [ perf record: Captured and wrote 5.433 MB perf.data.guest (62519 samples)
>
> # ./perf kvm report
> 31K kvm:kvm_entry
> 31K kvm:kvm_exit
>
> # ./perf kvm stat record -a
> [ perf record: Woken up 3 times to write data ]
> [ perf record: Captured and wrote 8.502 MB perf.data.guest (99338 samples) ]
>
> # ./perf kvm stat report --event=vmexit
> Event name Samples Sample% Time (ns) Time% Max Time (ns) Min Time (ns) Mean Time (ns)
> STORE_GUEST_PAGE_FAULT 26968 54.00% 2003031800 40.00% 3361400 27600 74274
> LOAD_GUEST_PAGE_FAULT 17645 35.00% 1153338100 23.00% 2513400 30800 65363
> VIRTUAL_INST_FAULT 1247 2.00% 340820800 6.00% 1190800 43300 273312
> INST_GUEST_PAGE_FAULT 1128 2.00% 340645800 6.00% 2123200 30200 301990
> SUPERVISOR_SYSCALL 1019 2.00% 245989900 4.00% 1851500 29300 241403
> LOAD_ACCESS 986 1.00% 671556200 13.00% 4180200 100700 681091
> INST_ACCESS 655 1.00% 170054800 3.00% 1808300 54600 259625
> HYPERVISOR_SYSCALL 21 0.00% 4276400 0.00% 716500 116000 203638
>
> Changes from v1->v2:
> - Rebased on Linux 6.9-rc3.
>
> Changes from v2->v3:
> - Add the missing assignment for 'vcpu_id_str' in patch 2.
> - Remove parentheses that cause compilation errors
>
> Shenlin Liang (2):
> RISCV: KVM: add tracepoints for entry and exit events
> perf kvm/riscv: Port perf kvm stat to RISC-V
>
> arch/riscv/kvm/trace.h | 67 ++++++++++++++++
> arch/riscv/kvm/vcpu.c | 7 ++
> tools/perf/arch/riscv/Makefile | 1 +
> tools/perf/arch/riscv/util/Build | 1 +
> tools/perf/arch/riscv/util/kvm-stat.c | 79 +++++++++++++++++++
> .../arch/riscv/util/riscv_exception_types.h | 35 ++++++++
> 6 files changed, 190 insertions(+)
> create mode 100644 arch/riscv/kvm/trace.h
> create mode 100644 tools/perf/arch/riscv/util/kvm-stat.c
> create mode 100644 tools/perf/arch/riscv/util/riscv_exception_types.h
>
> --
> 2.37.2
On Mon, Apr 29, 2024 at 02:13:22PM +0800, Shenlin Liang wrote:
> Gentle ping...
It would be great to have people with such boards testing your patchkit
and providing Tested-by tags.
Also the patch set has both kernel and tooling support, thus needs to
find a RiscV kernel maintainer to pick the kernel bits and then, when
that was merged, I would look at reports of tests for the tooling side
to then merge it.
Hope this clarifies the process,
- Arnaldo
> > 'perf kvm stat report/record' generates a statistical analysis of KVM
> > events and can be used to analyze guest exit reasons. This patch tries
> > to add stat support on riscv.
> >
> > Map the return value of trace_kvm_exit() to the specific cause of the
> > exception, and export it to userspace.
> >
> > It records on two available KVM tracepoints for riscv: "kvm:kvm_entry"
> > and "kvm:kvm_exit", and reports statistical data which includes events
> > handles time, samples, and so on.
> >
> > Cross compiling perf in X86 environment may encounter issues with missing
> > libraries and tools. Suggest compiling nativly in RISC-V environment
> >
> > Simple tests go below:
> >
> > # ./perf kvm record -e "kvm:kvm_entry" -e "kvm:kvm_exit"
> > Lowering default frequency rate from 4000 to 2500.
> > Please consider tweaking /proc/sys/kernel/perf_event_max_sample_rate.
> > [ perf record: Woken up 18 times to write data ]
> > [ perf record: Captured and wrote 5.433 MB perf.data.guest (62519 samples)
> >
> > # ./perf kvm report
> > 31K kvm:kvm_entry
> > 31K kvm:kvm_exit
> >
> > # ./perf kvm stat record -a
> > [ perf record: Woken up 3 times to write data ]
> > [ perf record: Captured and wrote 8.502 MB perf.data.guest (99338 samples) ]
> >
> > # ./perf kvm stat report --event=vmexit
> > Event name Samples Sample% Time (ns) Time% Max Time (ns) Min Time (ns) Mean Time (ns)
> > STORE_GUEST_PAGE_FAULT 26968 54.00% 2003031800 40.00% 3361400 27600 74274
> > LOAD_GUEST_PAGE_FAULT 17645 35.00% 1153338100 23.00% 2513400 30800 65363
> > VIRTUAL_INST_FAULT 1247 2.00% 340820800 6.00% 1190800 43300 273312
> > INST_GUEST_PAGE_FAULT 1128 2.00% 340645800 6.00% 2123200 30200 301990
> > SUPERVISOR_SYSCALL 1019 2.00% 245989900 4.00% 1851500 29300 241403
> > LOAD_ACCESS 986 1.00% 671556200 13.00% 4180200 100700 681091
> > INST_ACCESS 655 1.00% 170054800 3.00% 1808300 54600 259625
> > HYPERVISOR_SYSCALL 21 0.00% 4276400 0.00% 716500 116000 203638
> >
> > Changes from v1->v2:
> > - Rebased on Linux 6.9-rc3.
> >
> > Changes from v2->v3:
> > - Add the missing assignment for 'vcpu_id_str' in patch 2.
> > - Remove parentheses that cause compilation errors
> >
> > Shenlin Liang (2):
> > RISCV: KVM: add tracepoints for entry and exit events
> > perf kvm/riscv: Port perf kvm stat to RISC-V
> >
> > arch/riscv/kvm/trace.h | 67 ++++++++++++++++
> > arch/riscv/kvm/vcpu.c | 7 ++
> > tools/perf/arch/riscv/Makefile | 1 +
> > tools/perf/arch/riscv/util/Build | 1 +
> > tools/perf/arch/riscv/util/kvm-stat.c | 79 +++++++++++++++++++
> > .../arch/riscv/util/riscv_exception_types.h | 35 ++++++++
> > 6 files changed, 190 insertions(+)
> > create mode 100644 arch/riscv/kvm/trace.h
> > create mode 100644 tools/perf/arch/riscv/util/kvm-stat.c
> > create mode 100644 tools/perf/arch/riscv/util/riscv_exception_types.h
> >
> > --
> > 2.37.2
On 4/22/24 01:08, Shenlin Liang wrote:
> 'perf kvm stat report/record' generates a statistical analysis of KVM
> events and can be used to analyze guest exit reasons. This patch tries
> to add stat support on riscv.
>
> Map the return value of trace_kvm_exit() to the specific cause of the
> exception, and export it to userspace.
>
> It records on two available KVM tracepoints for riscv: "kvm:kvm_entry"
> and "kvm:kvm_exit", and reports statistical data which includes events
> handles time, samples, and so on.
>
> Cross compiling perf in X86 environment may encounter issues with missing
> libraries and tools. Suggest compiling nativly in RISC-V environment
>
Indeed. I am able to test the series with native build.
For the entire series:
Tested-by: Atish Patra <[email protected]>
> Simple tests go below:
>
> # ./perf kvm record -e "kvm:kvm_entry" -e "kvm:kvm_exit"
> Lowering default frequency rate from 4000 to 2500.
> Please consider tweaking /proc/sys/kernel/perf_event_max_sample_rate.
> [ perf record: Woken up 18 times to write data ]
> [ perf record: Captured and wrote 5.433 MB perf.data.guest (62519 samples)
>
> # ./perf kvm report
> 31K kvm:kvm_entry
> 31K kvm:kvm_exit
>
> # ./perf kvm stat record -a
> [ perf record: Woken up 3 times to write data ]
> [ perf record: Captured and wrote 8.502 MB perf.data.guest (99338 samples) ]
>
> # ./perf kvm stat report --event=vmexit
> Event name Samples Sample% Time (ns) Time% Max Time (ns) Min Time (ns) Mean Time (ns)
> STORE_GUEST_PAGE_FAULT 26968 54.00% 2003031800 40.00% 3361400 27600 74274
> LOAD_GUEST_PAGE_FAULT 17645 35.00% 1153338100 23.00% 2513400 30800 65363
> VIRTUAL_INST_FAULT 1247 2.00% 340820800 6.00% 1190800 43300 273312
> INST_GUEST_PAGE_FAULT 1128 2.00% 340645800 6.00% 2123200 30200 301990
> SUPERVISOR_SYSCALL 1019 2.00% 245989900 4.00% 1851500 29300 241403
> LOAD_ACCESS 986 1.00% 671556200 13.00% 4180200 100700 681091
> INST_ACCESS 655 1.00% 170054800 3.00% 1808300 54600 259625
> HYPERVISOR_SYSCALL 21 0.00% 4276400 0.00% 716500 116000 203638
>
> Changes from v1->v2:
> - Rebased on Linux 6.9-rc3.
>
> Changes from v2->v3:
> - Add the missing assignment for 'vcpu_id_str' in patch 2.
> - Remove parentheses that cause compilation errors
>
> Shenlin Liang (2):
> RISCV: KVM: add tracepoints for entry and exit events
> perf kvm/riscv: Port perf kvm stat to RISC-V
>
> arch/riscv/kvm/trace.h | 67 ++++++++++++++++
> arch/riscv/kvm/vcpu.c | 7 ++
> tools/perf/arch/riscv/Makefile | 1 +
> tools/perf/arch/riscv/util/Build | 1 +
> tools/perf/arch/riscv/util/kvm-stat.c | 79 +++++++++++++++++++
> .../arch/riscv/util/riscv_exception_types.h | 35 ++++++++
> 6 files changed, 190 insertions(+)
> create mode 100644 arch/riscv/kvm/trace.h
> create mode 100644 tools/perf/arch/riscv/util/kvm-stat.c
> create mode 100644 tools/perf/arch/riscv/util/riscv_exception_types.h
>
On Mon, Apr 22, 2024 at 1:42 PM Shenlin Liang
<[email protected]> wrote:
>
> 'perf kvm stat report/record' generates a statistical analysis of KVM
> events and can be used to analyze guest exit reasons. This patch tries
> to add stat support on riscv.
>
> Map the return value of trace_kvm_exit() to the specific cause of the
> exception, and export it to userspace.
>
> It records on two available KVM tracepoints for riscv: "kvm:kvm_entry"
> and "kvm:kvm_exit", and reports statistical data which includes events
> handles time, samples, and so on.
>
> Cross compiling perf in X86 environment may encounter issues with missing
> libraries and tools. Suggest compiling nativly in RISC-V environment
>
> Simple tests go below:
>
> # ./perf kvm record -e "kvm:kvm_entry" -e "kvm:kvm_exit"
> Lowering default frequency rate from 4000 to 2500.
> Please consider tweaking /proc/sys/kernel/perf_event_max_sample_rate.
> [ perf record: Woken up 18 times to write data ]
> [ perf record: Captured and wrote 5.433 MB perf.data.guest (62519 samples)
>
> # ./perf kvm report
> 31K kvm:kvm_entry
> 31K kvm:kvm_exit
>
> # ./perf kvm stat record -a
> [ perf record: Woken up 3 times to write data ]
> [ perf record: Captured and wrote 8.502 MB perf.data.guest (99338 samples) ]
>
> # ./perf kvm stat report --event=vmexit
> Event name Samples Sample% Time (ns) Time% Max Time (ns) Min Time (ns) Mean Time (ns)
> STORE_GUEST_PAGE_FAULT 26968 54.00% 2003031800 40.00% 3361400 27600 74274
> LOAD_GUEST_PAGE_FAULT 17645 35.00% 1153338100 23.00% 2513400 30800 65363
> VIRTUAL_INST_FAULT 1247 2.00% 340820800 6.00% 1190800 43300 273312
> INST_GUEST_PAGE_FAULT 1128 2.00% 340645800 6.00% 2123200 30200 301990
> SUPERVISOR_SYSCALL 1019 2.00% 245989900 4.00% 1851500 29300 241403
> LOAD_ACCESS 986 1.00% 671556200 13.00% 4180200 100700 681091
> INST_ACCESS 655 1.00% 170054800 3.00% 1808300 54600 259625
> HYPERVISOR_SYSCALL 21 0.00% 4276400 0.00% 716500 116000 203638
>
> Changes from v1->v2:
> - Rebased on Linux 6.9-rc3.
>
> Changes from v2->v3:
> - Add the missing assignment for 'vcpu_id_str' in patch 2.
> - Remove parentheses that cause compilation errors
>
> Shenlin Liang (2):
> RISCV: KVM: add tracepoints for entry and exit events
> perf kvm/riscv: Port perf kvm stat to RISC-V
>
> arch/riscv/kvm/trace.h | 67 ++++++++++++++++
> arch/riscv/kvm/vcpu.c | 7 ++
> tools/perf/arch/riscv/Makefile | 1 +
> tools/perf/arch/riscv/util/Build | 1 +
> tools/perf/arch/riscv/util/kvm-stat.c | 79 +++++++++++++++++++
> .../arch/riscv/util/riscv_exception_types.h | 35 ++++++++
> 6 files changed, 190 insertions(+)
> create mode 100644 arch/riscv/kvm/trace.h
> create mode 100644 tools/perf/arch/riscv/util/kvm-stat.c
> create mode 100644 tools/perf/arch/riscv/util/riscv_exception_types.h
Queued this series for Linux-6.11
Thanks,
Anup