Use the virtual instead of the physical timer for measuring the time
taken to execute the microbenchmark.
Internal testing discovered a performance regression on this test
starting with Linux commit 680232a94c12 "KVM: arm64: timers: Allow
save/restoring of the physical timer". Oliver Upton speculates QEMU is
changing the guest physical counter to have a nonzero offset since it
gained the ability as of that commit. As a consequence KVM is
trap-and-emulating here on architectures without FEAT_ECV.
While this isn't a correctness issue, the trap-and-emulate overhead of
physical counter emulation on systems without ECV leads to surprising
microbenchmark results.
Signed-off-by: Colton Lewis <coltonlewis@xxxxxxxxxx>
---
arm/micro-bench.c | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/arm/micro-bench.c b/arm/micro-bench.c
index bfd181dc..fbe59d03 100644
--- a/arm/micro-bench.c
+++ b/arm/micro-bench.c
@@ -348,10 +348,10 @@ static void loop_test(struct exit_test *test)
while (ntimes < test->times && total_ns.ns < NS_5_SECONDS) {
isb();
- start = read_sysreg(cntpct_el0);
+ start = read_sysreg(cntvct_el0);
test->exec();
isb();
- end = read_sysreg(cntpct_el0);
+ end = read_sysreg(cntvct_el0);
ntimes++;
total_ticks += (end - start);
--
2.42.0.rc1.204.g551eb34607-goog
