> From: Paolo Bonzini [mailto:paolo.bonz...@gmail.com] On Behalf Of Paolo > Bonzini > Il 28/07/2014 11:54, Pavel Dovgaluk ha scritto: > >>> > > + VMSTATE_TIMER_V(timer, RTL8139State, 5), > >> > > >> > timer need not be migrated, because it is reinstated by > >> > rtl8139_post_load. > >> > > > That's true for normal execution. > > In replay execution mode post_load can be called before cached virtual > > clock > > values are loaded. This may cause invalid setting of the timer and raising > > an IRQ which didn't happen in record mode. > > I will update this patch to fix post_load function and avoid this > > non-deterministic behavior. > > This is what worries me of this series. These invariants are not > documented anywhere, and people will break them unless you add > assertions that also hold in normal mode.
Assertions is a good idea, we added such warning message to qemu_get_timedate function to be sure, that it is used correctly with replay. Another thing, that could help for making snapshots - find a way to load replay structures before all other ones. Are there any priorities in migration states list? Priorities could also solve some other issues, because sometimes post_load function of one device uses other devices' functions. And the second ones could be not loaded yet. Pavel Dovgalyuk