Anthony Liguori <anth...@codemonkey.ws> writes: > Rusty Russell <ru...@rustcorp.com.au> writes: > >> Anthony Liguori <aligu...@us.ibm.com> writes: >>> We were memcpy()'ing a structure to the wire :-/ Since savevm really >>> only works on x86 today, lets just declare that this element is sent >>> over the wire as a little endian value in order to fix the bitness. >>> >>> Unfortunately, we also send raw pointers and size_t which are going >>> to be different values on a 32-bit vs. 64-bit QEMU so we need to also >>> deal with that case. >>> >>> A lot of values that should have been previously ignored are now sent >>> as 0 and ignored on the receive side too. >> >> Don't we want to transition to vmstate anyway? Can we just do that, and >> relegate the existing slightly broken code, to legacy? > > What worries me is if someone changes VirtQueueElement, then all the > sudden migration breaks. By transitioning to what I've sent, we at > least have a better documented protocol that isn't prone to subtle > breakage anymore. Plus, we resolve the endian issue before it becomes a > bigger problem when David actually gets live migration working reliably > on PPC...
My transition was to copy that structure to VirtQueueSavedElement, and it's only used for loading old versions. With the new code we only need the head from that structure. > I'm certainly in favor of cleaning up the savevm format and probably > leaving the existing load/save functions as-is for legacy purposes. > I'll leave that as an exercise for someone else though :-) What is the rule about new versions? Can we introduce a new save version at any time, or only at major qemu version changes? Thanks, Rusty.