On Fri, Oct 20, 2023, Alexey Kardashevskiy wrote: > > On 20/10/23 11:13, Sean Christopherson wrote: > > On Fri, Oct 20, 2023, Alexey Kardashevskiy wrote: > > > Plus, GHCB now has to go via the userspace before talking to the PSP which > > > was not the case so far (though I cannot think of immediate implication > > > right now). > > > > Any argument along the lines of "because that's how we've always done it" is going > > to fall on deaf ears. If there's a real performance bottleneck with kicking out > > to userspace, then I'll happily work to figure out a solution. If. > > No, not performance, I was trying to imagine what can go wrong if multiple > vcpus are making this call, all exiting to QEMU, in a loop, racing, > something like this. I am not at all concerned about userspace being able to handle parallel requests to get a certificate. Per-vCPU exits that access global/shared resources might not be super common, but they're certainly not rare. E.g. a guest access to an option ROM can trigger memslot updates in QEMU, which requires at least taking a mutex to guard KVM_SET_USER_MEMORY_REGION, and IIRC QEMU also uses RCU to protect QEMU accesses to address spaces. Given that we know there will be scenarios where certificates are changed/updated, I wouldn't be at all surprised if handling this in userspace is actually easier as it will give userspace more control and options, and make it easier to reason about the resulting behavior. E.g. userspace could choose between a lockless scheme and a r/w lock if there's a need to ensure per-VM and global certs are updated atomically from the guest's perspective.
