On 12/06/2023 19:32, Arnaldo Carvalho de Melo wrote:
Em Mon, Jun 12, 2023 at 12:13:57PM +0100, James Clark escreveu:
Changes since v2:
- Rename prev_thread -> prev_packet_thread and prev_el -> prev_packet_el
- Add a comment about tracking the previous packet's thread
Changes since v1:
- Always assume host kernel when the trace was captured at EL1 (nVHE)
- Fix EL validation to work with ETMv3
- Add a commit to make PID format accessible from struct cs_etm_auxtrace
Please take a look in my tmp.perf-tools-next branch, there were some conflicts I had to fix as those files were touched by refactorings for addr_location and thread reference counting.
Yeah I got the same result and the tests are still passing. Thanks for fixing those.
⬢[acme@toolbox perf-tools-next]$ git log --oneline -10 aa53fb2c482e70c2 (HEAD -> perf-tools-next) perf cs-etm: Add exception level consistency check 2918e9895224541f perf cs-etm: Track exception level f492a33909829a75 perf cs-etm: Make PID format accessible from struct cs_etm_auxtrace e29ec19b0751c6b2 perf cs-etm: Use previous thread for branch sample source IP e9e03e9c3ca7088c perf cs-etm: Only track threads instead of PID and TIDs 6fd34445b8c94aa7 perf map: Fix double 'struct map' reference free found with -DREFCNT_CHECKING=1 e9c0a7f63e45e76f perf srcline: Optimize comparision against SRCLINE_UNKNOWN fd87a79c7ed62804 perf hist: Fix srcline memory leak 933f9651d47cdda2 perf srcline: Change free_srcline to zfree_srcline d22cfb063bcc674e perf callchain: Use pthread keys for tls callchain_cursor ⬢[acme@toolbox perf-tools-next]$
- Arnaldo
======
Some fixes to support an issue reported by Denis Nikitin where decoding trace that contains different EL1 and EL2 kernels can crash or go into an infinite loop because the wrong kernel maps are used for the decode.
This still doesn't support distinguishing guest and host userspace, we'd still have to fix the timestamps and do a bit more work to correlate that. And I've removed PERF_RECORD_MISC_HYPERVISOR as a possible outcome of cs_etm__cpu_mode(). As far as I know this could never have been returned anyway because machine__is_host(machine) was always true due to session.machines.host being hard coded. And I'm not sure of the relevance of the difference between PERF_RECORD_MISC_KERNEL and PERF_RECORD_MISC_HYPERVISOR in this scenario.
The first commit is a tidy up, second fixes a bug that I found when comparing the exception level and thread of branch records, the third is the main fix, and the last commit is some extra error checking.
Applies to acme/perf-tools-next (42713dafc)
James Clark (5): perf cs-etm: Only track threads instead of PID and TIDs perf cs-etm: Use previous thread for branch sample source IP perf cs-etm: Make PID format accessible from struct cs_etm_auxtrace perf cs-etm: Track exception level perf cs-etm: Add exception level consistency check
.../perf/util/cs-etm-decoder/cs-etm-decoder.c | 33 +- .../perf/util/cs-etm-decoder/cs-etm-decoder.h | 4 +- tools/perf/util/cs-etm.c | 282 ++++++++++-------- tools/perf/util/cs-etm.h | 13 +- 4 files changed, 184 insertions(+), 148 deletions(-)
-- 2.34.1