On Tue, May 14, 2024, 03:57 Zhao Liu <zhao1....@intel.com> wrote:
> Hi Stefan,
>
> > QEMU uses clock_gettime(CLOCK_MONOTONIC) on Linux hosts. The man page
> > says:
> >
> > All CLOCK_MONOTONIC variants guarantee that the time returned by
> > consecutive calls will not go backwards, but successive calls
> > may—depending on the architecture—return identical (not-in‐
> > creased) time values.
> >
> > trace_record_start() calls clock_gettime(CLOCK_MONOTONIC) so trace events
> > should have monotonically increasing timestamps.
> >
> > I don't see a scenario where trace record A's timestamp is greater than
> > trace record B's timestamp unless the clock is non-monotonic.
> >
> > Which host CPU architecture and operating system are you running?
>
> I tested on these 2 machines:
> * CML (intel 10th) with Ubuntu 22.04 + kernel v6.5.0-28
> * MTL (intel 14th) with Ubuntu 22.04.2 + kernel v6.9.0
>
> > Please attach to the QEMU process with gdb and print out the value of
> > the use_rt_clock variable or add a printf in init_get_clock(). The value
> > should be 1.
>
> Thanks, on both above machines, use_rt_clock is 1 and there're both
> timestamp reversal issues with the following debug print:
>
> diff --git a/include/qemu/timer.h b/include/qemu/timer.h
> index 9a366e551fb3..7657785c27dc 100644
> --- a/include/qemu/timer.h
> +++ b/include/qemu/timer.h
> @@ -831,10 +831,17 @@ extern int use_rt_clock;
>
> static inline int64_t get_clock(void)
> {
> + static int64_t clock = 0;
>
Please try with a thread local variable (__thread) to check whether this
happens within a single thread.
If it only happens with a global variable then we'd need to look more
closely at race conditions in the patch below. I don't think the patch is a
reliable way to detect non-monotonic timestamps in a multi-threaded program.
if (use_rt_clock) {
> struct timespec ts;
> clock_gettime(CLOCK_MONOTONIC, &ts);
> - return ts.tv_sec * 1000000000LL + ts.tv_nsec;
> + int64_t tmp = ts.tv_sec * 1000000000LL + ts.tv_nsec;
> + if (tmp <= clock) {
> + printf("get_clock: strange, clock: %ld, tmp: %ld\n", clock,
> tmp);
> + }
> + assert(tmp > clock);
> + clock = tmp;
> + return clock;
> } else {
> /* XXX: using gettimeofday leads to problems if the date
> changes, so it should be avoided. */
> diff --git a/util/qemu-timer-common.c b/util/qemu-timer-common.c
> index cc1326f72646..3bf06eb4a4ce 100644
> --- a/util/qemu-timer-common.c
> +++ b/util/qemu-timer-common.c
> @@ -59,5 +59,6 @@ static void __attribute__((constructor))
> init_get_clock(void)
> use_rt_clock = 1;
> }
> clock_start = get_clock();
> + printf("init_get_clock: use_rt_clock: %d\n", use_rt_clock);
> }
> #endif
>
> ---
> The timestamp interval is very small, for example:
> get_clock: strange, clock: 3302130503505, tmp: 3302130503503
>
> or
>
> get_clock: strange, clock: 2761577819846455, tmp: 2761577819846395
>
> I also tried to use CLOCK_MONOTONIC_RAW, but there's still the reversal
> issue.
>
> Thanks,
> Zhao
>
>
