On Thu, Jun 20, 2013 at 07:38:18PM +0100, Will Deacon wrote: > On Thu, Jun 20, 2013 at 07:28:47PM +0100, Christoffer Dall wrote: > > On Thu, Jun 20, 2013 at 07:15:25PM +0100, Will Deacon wrote: > > > On Thu, Jun 20, 2013 at 06:14:09PM +0100, Christoffer Dall wrote: > > > > ok, I was trying to think about how it would break, and if a guest needs > > > > a TLB invalidation to be visisble by other CPUs it would have to have a > > > > dsb/isb itself after the operation, and that would eventually be > > > > executed once the VCPU was rescheduled, but potentially on another CPU, > > > > but then I wonder if the PCPU migration on the host wouldn't take care > > > > of it? > > > > > > Actually, it's worse than both of you think :) > > > > > > The dsb *must* be executed on the same physical CPU as the TLB invalidation. > > > The same virtual CPU isn't enough, which is all that is guaranteed by the > > > guest. If you don't have a dsb on your vcpu migration path, then you need > > > something here. > > > > > > The same thing applies to cache maintenance operations. > > > > > But are we not sure that a dsb will happen anywhere in the kernel if a > > process is migrated to a different core? > > Yes, we have a dsb when we unlock the runqueue for a CPU. That's why Linux > doesn't crash and burn usually. If vcpu migration always goes through the > usual scheduling paths, then you don't have a problem. > Right, a vcpu is simply a thread, a process, so it gets migrated on the host as any other process. I gather this means we don't need these, except maybe for the VMID rollover case, which I honestly didn't fully understand, but maybe it can be added for that specific case instead? -Christoffer _______________________________________________ kvmarm mailing list kvmarm@xxxxxxxxxxxxxxxxxxxxx https://lists.cs.columbia.edu/cucslists/listinfo/kvmarm
