perf/x86/intel/ds: Fix the conversion from TSC to perf time
authorKan Liang <kan.liang@linux.intel.com>
Wed, 25 Jan 2023 20:49:25 +0000 (12:49 -0800)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Fri, 10 Mar 2023 08:32:45 +0000 (09:32 +0100)
commit28057ce72204f009db75550dc916cce33cdd2ee7
tree3b02a1bd0b67e7956c446617b5e305de3becf208
parent6b4fcc4e8a3016e85766c161daf0732fca16c3a3
perf/x86/intel/ds: Fix the conversion from TSC to perf time

[ Upstream commit 89e97eb8cec0f1af5ebf2380308913256ca7915a ]

The time order is incorrect when the TSC in a PEBS record is used.

 $perf record -e cycles:upp dd if=/dev/zero of=/dev/null
  count=10000
 $ perf script --show-task-events
       perf-exec     0     0.000000: PERF_RECORD_COMM: perf-exec:915/915
              dd   915   106.479872: PERF_RECORD_COMM exec: dd:915/915
              dd   915   106.483270: PERF_RECORD_EXIT(915:915):(914:914)
              dd   915   106.512429:          1 cycles:upp:
 ffffffff96c011b7 [unknown] ([unknown])
 ... ...

The perf time is from sched_clock_cpu(). The current PEBS code
unconditionally convert the TSC to native_sched_clock(). There is a
shift between the two clocks. If the TSC is stable, the shift is
consistent, __sched_clock_offset. If the TSC is unstable, the shift has
to be calculated at runtime.

This patch doesn't support the conversion when the TSC is unstable. The
TSC unstable case is a corner case and very unlikely to happen. If it
happens, the TSC in a PEBS record will be dropped and fall back to
perf_event_clock().

Fixes: 47a3aeb39e8d ("perf/x86/intel/pebs: Fix PEBS timestamps overwritten")
Reported-by: Namhyung Kim <namhyung@kernel.org>
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/all/CAM9d7cgWDVAq8-11RbJ2uGfwkKD6fA-OMwOKDrNUrU_=8MgEjg@mail.gmail.com/
Signed-off-by: Sasha Levin <sashal@kernel.org>
arch/x86/events/intel/ds.c