On 27/01/2023 10:55 am, Juergen Gross wrote: > > For xenstore-stubdom it is easier, as there is no reason to prefer > the PV mode over PVH (in fact I'm still working on Xenstore LU for > the PVH case, making the decision even easier).
I fully support having both PV and PVH options available. It helps maintain architecturally "clean" interfaces, and not decisions are about performance. But a PVH stub xenstored cannot outperform a equivalent PV stub xenstored, for x86 architectural reasons. PV stub xenstored doesn't do any of the usual things that causes PV guests to be generally slow (different address spaces, user/kernel context switching), and event/grants to PV guests is architecturally more efficient than the HVM/PVH equivalents. PV will never go away fully die, because we've got enough fastpaths where this matters, but I do expect that we'll eventually get to a point where we don't have any general purpose OSes running as PV - only single-task workloads such as stub xenstored. The great thing about having both PV and PVH available is that people will be able to see the concrete data confirming that PV is better, without having to simply trust me on the matter. ~Andrew
