mbox series

[v3,0/5] Correct perf sampling with Guest VMs

Message ID 20240912205133.4171576-1-coltonlewis@google.com (mailing list archive)
Headers show
Series Correct perf sampling with Guest VMs | expand

Message

Colton Lewis Sept. 12, 2024, 8:51 p.m. UTC 
v3:
  * Clarify final commit message further
  * Remove an unused variable in perf_arch_misc_flags()

v2:
https://lore.kernel.org/kvm/20240911222433.3415301-1-coltonlewis@google.com/

v1:
https://lore.kernel.org/kvm/20240904204133.1442132-1-coltonlewis@google.com/

This series cleans up perf recording around guest events and improves
the accuracy of the resulting perf reports.

Perf was incorrectly counting any PMU overflow interrupt that occurred
while a VCPU was loaded as a guest event even when the events were not
truely guest events. This lead to much less accurate and useful perf
recordings.

See as an example the below reports of `perf record
dirty_log_perf_test -m 2 -v 4` before and after the series on ARM64.

Without series:

Samples: 15K of event 'instructions', Event count (approx.): 31830580924
Overhead  Command          Shared Object        Symbol
  54.54%  dirty_log_perf_  dirty_log_perf_test  [.] run_test
   5.39%  dirty_log_perf_  dirty_log_perf_test  [.] vcpu_worker
   0.89%  dirty_log_perf_  [kernel.vmlinux]     [k] release_pages
   0.70%  dirty_log_perf_  [kernel.vmlinux]     [k] free_pcppages_bulk
   0.62%  dirty_log_perf_  dirty_log_perf_test  [.] userspace_mem_region_find
   0.49%  dirty_log_perf_  dirty_log_perf_test  [.] sparsebit_is_set
   0.46%  dirty_log_perf_  dirty_log_perf_test  [.] _virt_pg_map
   0.46%  dirty_log_perf_  dirty_log_perf_test  [.] node_add
   0.37%  dirty_log_perf_  dirty_log_perf_test  [.] node_reduce
   0.35%  dirty_log_perf_  [kernel.vmlinux]     [k] free_unref_page_commit
   0.33%  dirty_log_perf_  [kernel.vmlinux]     [k] __kvm_pgtable_walk
   0.31%  dirty_log_perf_  [kernel.vmlinux]     [k] stage2_attr_walker
   0.29%  dirty_log_perf_  [kernel.vmlinux]     [k] unmap_page_range
   0.29%  dirty_log_perf_  dirty_log_perf_test  [.] test_assert
   0.26%  dirty_log_perf_  [kernel.vmlinux]     [k] __mod_memcg_lruvec_state
   0.24%  dirty_log_perf_  [kernel.vmlinux]     [k] kvm_s2_put_page

With series:

Samples: 15K of event 'instructions', Event count (approx.): 31830580924
Samples: 15K of event 'instructions', Event count (approx.): 30898031385
Overhead  Command          Shared Object        Symbol
  54.05%  dirty_log_perf_  dirty_log_perf_test  [.] run_test
   5.48%  dirty_log_perf_  [kernel.kallsyms]    [k] kvm_arch_vcpu_ioctl_run
   4.70%  dirty_log_perf_  dirty_log_perf_test  [.] vcpu_worker
   3.11%  dirty_log_perf_  [kernel.kallsyms]    [k] kvm_handle_guest_abort
   2.24%  dirty_log_perf_  [kernel.kallsyms]    [k] up_read
   1.98%  dirty_log_perf_  [kernel.kallsyms]    [k] __kvm_tlb_flush_vmid_ipa_nsh
   1.97%  dirty_log_perf_  [kernel.kallsyms]    [k] __pi_clear_page
   1.30%  dirty_log_perf_  [kernel.kallsyms]    [k] down_read
   1.13%  dirty_log_perf_  [kernel.kallsyms]    [k] release_pages
   1.12%  dirty_log_perf_  [kernel.kallsyms]    [k] __kvm_pgtable_walk
   1.08%  dirty_log_perf_  [kernel.kallsyms]    [k] folio_batch_move_lru
   1.06%  dirty_log_perf_  [kernel.kallsyms]    [k] __srcu_read_lock
   1.03%  dirty_log_perf_  [kernel.kallsyms]    [k] get_page_from_freelist
   1.01%  dirty_log_perf_  [kernel.kallsyms]    [k] __pte_offset_map_lock
   0.82%  dirty_log_perf_  [kernel.kallsyms]    [k] handle_mm_fault
   0.74%  dirty_log_perf_  [kernel.kallsyms]    [k] mas_state_walk

Colton Lewis (5):
  arm: perf: Drop unused functions
  perf: Hoist perf_instruction_pointer() and perf_misc_flags()
  powerpc: perf: Use perf_arch_instruction_pointer()
  x86: perf: Refactor misc flag assignments
  perf: Correct perf sampling with guest VMs

 arch/arm/include/asm/perf_event.h            |  7 ---
 arch/arm/kernel/perf_callchain.c             | 17 -------
 arch/arm64/include/asm/perf_event.h          |  4 --
 arch/arm64/kernel/perf_callchain.c           | 28 ------------
 arch/powerpc/include/asm/perf_event_server.h |  6 +--
 arch/powerpc/perf/callchain.c                |  2 +-
 arch/powerpc/perf/callchain_32.c             |  2 +-
 arch/powerpc/perf/callchain_64.c             |  2 +-
 arch/powerpc/perf/core-book3s.c              |  4 +-
 arch/s390/include/asm/perf_event.h           |  6 +--
 arch/s390/kernel/perf_event.c                |  4 +-
 arch/x86/events/core.c                       | 47 +++++++++++---------
 arch/x86/include/asm/perf_event.h            | 12 ++---
 include/linux/perf_event.h                   | 26 +++++++++--
 kernel/events/core.c                         | 27 ++++++++++-
 15 files changed, 95 insertions(+), 99 deletions(-)


base-commit: da3ea35007d0af457a0afc87e84fddaebc4e0b63
--
2.46.0.662.g92d0881bb0-goog

Comments

Oliver Upton Sept. 19, 2024, 10:17 a.m. UTC | #1 
On Thu, Sep 12, 2024 at 08:51:28PM +0000, Colton Lewis wrote:
> v3:
>   * Clarify final commit message further
>   * Remove an unused variable in perf_arch_misc_flags()
> 
> v2:
> https://lore.kernel.org/kvm/20240911222433.3415301-1-coltonlewis@google.com/
> 
> v1:
> https://lore.kernel.org/kvm/20240904204133.1442132-1-coltonlewis@google.com/
> 
> This series cleans up perf recording around guest events and improves
> the accuracy of the resulting perf reports.

Please fix the intermediate build issue, and also test that each patch
in the series compiles. With that corrected, for the series:

Reviewed-by: Oliver Upton <oliver.upton@linux.dev>

A nice follow-up on the arm64 side would be to further constrain
kvm_arch_pmi_in_guest() to return true iff we exited the guest due to an
IRQ.