On Sun, 2025-11-30 at 16:47 +0000, Alex Bennée wrote:
> Ilya Leoshkevich <[email protected]> writes:
>
> > On Fri, 2025-11-28 at 18:25 +0100, Ilya Leoshkevich wrote:
> > > On Fri, 2025-11-28 at 14:39 +0100, Thomas Huth wrote:
> > > > From: Thomas Huth <[email protected]>
> > > >
> > > > We just have to make sure that we can set the endianness to big
> > > > endian,
> > > > then we can also run this test on s390x.
> > > >
> > > > Signed-off-by: Thomas Huth <[email protected]>
> > > > ---
> > > > Marked as RFC since it depends on the fix for this bug (so it
> > > > cannot
> > > > be merged yet):
> > > >
> > > > https://lore.kernel.org/qemu-devel/[email protected]
> > > > /
> > > >
> > > > tests/functional/reverse_debugging.py | 4 +++-
> > > > tests/functional/s390x/meson.build | 1 +
> > > > tests/functional/s390x/test_reverse_debug.py | 21
> > > > ++++++++++++++++++++
> > > > 3 files changed, 25 insertions(+), 1 deletion(-)
> > > > create mode 100755
> > > > tests/functional/s390x/test_reverse_debug.py
> > >
> > > Reviewed-by: Ilya Leoshkevich <[email protected]>
> > >
> > >
> > > I have a simple fix which helps with your original report, but
> > > not
> > > with this test. I'm still investigating.
> > >
> > > --- a/target/s390x/machine.c
> > > +++ b/target/s390x/machine.c
> > > @@ -52,6 +52,14 @@ static int cpu_pre_save(void *opaque)
> > > kvm_s390_vcpu_interrupt_pre_save(cpu);
> > > }
> > >
> > > + if (tcg_enabled()) {
> > > + /*
> > > + * Ensure symmetry with cpu_post_load() with respect to
> > > + * CHECKPOINT_CLOCK_VIRTUAL.
> > > + */
> > > + tcg_s390_tod_updated(CPU(cpu), RUN_ON_CPU_NULL);
> > > + }
> > > +
> > > return 0;
> > > }
> >
> > Interestingly enough, this patch fails only under load, e.g., if I
> > run
> > make check -j"$(nproc)" or if I run your test in isolation, but
> > with
> > stress-ng cpu in background. The culprit appears to be:
> >
> > s390_tod_load()
> > qemu_s390_tod_set()
> > async_run_on_cpu(tcg_s390_tod_updated)
> >
> > Depending on the system load, this additional
> > tcg_s390_tod_updated()
> > may or may not end up being called during handle_backward(). If it
> > does, we get an infinite loop again, because now we need two
> > checkpoints.
> >
> > I have a feeling that this code may be violating some record-replay
> > requirement, but I can't quite put my finger on it. For example,
> > async_run_on_cpu() does not sound like something deterministic, but
> > then again it just queues work for rr_cpu_thread_fn(), which is
> > supposed to be deterministic.
>
> The the async_run_on_cpu is called from the vcpu thread in response
> to a
> deterministic event at a known point in time it should be fine. If it
> came from another thread that is not synchronised via replay_lock
> then
> things will go wrong.
>
> But this is a VM load save helper?
Yes, and it's called from the main thread. Either during
initialization, or as a reaction to GDB packets.
Here is the call stack:
qemu_loadvm_state()
qemu_loadvm_state_main()
qemu_loadvm_section_start_full()
vmstate_load()
vmstate_load_state()
cpu_post_load()
tcg_s390_tod_updated()
update_ckc_timer()
timer_mod()
s390_tod_load()
qemu_s390_tod_set() # via tdc->set()
async_run_on_cpu(tcg_s390_tod_updated)
So you think we may have to take the replay lock around
load_snapshot()? So that all async_run_on_cpu() calls it makes end up
being handled by the vCPU thread deterministically.
