mbox series

[0/3] Add perf event support for guest VM

Message ID 20250224095618.1436016-1-maobibo@loongson.cn (mailing list archive)
Headers show
Series Add perf event support for guest VM | expand

Message

bibo mao Feb. 24, 2025, 9:56 a.m. UTC 
From perf pmu interrupt is normal IRQ rather than NMI, so code cannot be
profiled if interrupt is disabled. However it is possible to profile
guest kernel in this situation from host side, profile result is more
accurate from host than that from guest.

Perf event support for guest VM is added here, and the below is the
example:
perf kvm --host --guest --guestkallsyms=guest-kallsyms
     --guestmodules=guest-modules  top

Overhead  Shared Object               Symbol
  20.02%  [guest.kernel]              [g] __arch_cpu_idle
  16.74%  [guest.kernel]              [g] queued_spin_lock_slowpath
  10.05%  [kernel]                    [k] __arch_cpu_idle
   2.00%  [guest.kernel]              [g] clear_page
   1.62%  [guest.kernel]              [g] copy_page
   1.50%  [guest.kernel]              [g] next_uptodate_folio
   1.41%  [guest.kernel]              [g] queued_write_lock_slowpath
   1.41%  [guest.kernel]              [g] unmap_page_range
   1.36%  [guest.kernel]              [g] mod_objcg_state
   1.30%  [guest.kernel]              [g] osq_lock
   1.28%  [guest.kernel]              [g] __slab_free
   0.98%  [guest.kernel]              [g] copy_page_range

Bibo Mao (3):
  LoongArch: KVM: Add stub for kvm_arch_vcpu_preempted_in_kernel
  LoongArch: KVM: Implement arch specified functions for guest perf
  LoongArch: KVM: Register perf callback for guest

 arch/loongarch/include/asm/kvm_host.h |  2 ++
 arch/loongarch/kvm/Kconfig            |  1 +
 arch/loongarch/kvm/main.c             |  2 ++
 arch/loongarch/kvm/vcpu.c             | 31 +++++++++++++++++++++++++++
 4 files changed, 36 insertions(+)


base-commit: 2408a807bfc3f738850ef5ad5e3fd59d66168996

Comments

Huacai Chen March 1, 2025, 1:41 p.m. UTC | #1 
Series applied, thanks.

Huacai

On Mon, Feb 24, 2025 at 5:56 PM Bibo Mao <maobibo@loongson.cn> wrote:
>
> From perf pmu interrupt is normal IRQ rather than NMI, so code cannot be
> profiled if interrupt is disabled. However it is possible to profile
> guest kernel in this situation from host side, profile result is more
> accurate from host than that from guest.
>
> Perf event support for guest VM is added here, and the below is the
> example:
> perf kvm --host --guest --guestkallsyms=guest-kallsyms
>      --guestmodules=guest-modules  top
>
> Overhead  Shared Object               Symbol
>   20.02%  [guest.kernel]              [g] __arch_cpu_idle
>   16.74%  [guest.kernel]              [g] queued_spin_lock_slowpath
>   10.05%  [kernel]                    [k] __arch_cpu_idle
>    2.00%  [guest.kernel]              [g] clear_page
>    1.62%  [guest.kernel]              [g] copy_page
>    1.50%  [guest.kernel]              [g] next_uptodate_folio
>    1.41%  [guest.kernel]              [g] queued_write_lock_slowpath
>    1.41%  [guest.kernel]              [g] unmap_page_range
>    1.36%  [guest.kernel]              [g] mod_objcg_state
>    1.30%  [guest.kernel]              [g] osq_lock
>    1.28%  [guest.kernel]              [g] __slab_free
>    0.98%  [guest.kernel]              [g] copy_page_range
>
> Bibo Mao (3):
>   LoongArch: KVM: Add stub for kvm_arch_vcpu_preempted_in_kernel
>   LoongArch: KVM: Implement arch specified functions for guest perf
>   LoongArch: KVM: Register perf callback for guest
>
>  arch/loongarch/include/asm/kvm_host.h |  2 ++
>  arch/loongarch/kvm/Kconfig            |  1 +
>  arch/loongarch/kvm/main.c             |  2 ++
>  arch/loongarch/kvm/vcpu.c             | 31 +++++++++++++++++++++++++++
>  4 files changed, 36 insertions(+)
>
>
> base-commit: 2408a807bfc3f738850ef5ad5e3fd59d66168996
> --
> 2.39.3
>
>