Hi David, On Tue, Mar 19, 2024 at 11:26:05AM +0100, David Hildenbrand wrote: > On 19.03.24 01:10, Sean Christopherson wrote: > > On Mon, Mar 18, 2024, Vishal Annapurve wrote: > > > On Mon, Mar 18, 2024 at 3:02 PM David Hildenbrand <david@xxxxxxxxxx> wrote: > > > > Second, we should find better ways to let an IOMMU map these pages, > > > > *not* using GUP. There were already discussions on providing a similar > > > > fd+offset-style interface instead. GUP really sounds like the wrong > > > > approach here. Maybe we should look into passing not only guest_memfd, > > > > but also "ordinary" memfds. > > > > +1. I am not completely opposed to letting SNP and TDX effectively convert > > pages between private and shared, but I also completely agree that letting > > anything gup() guest_memfd memory is likely to end in tears. > > Yes. Avoid it right from the start, if possible. > > People wanted guest_memfd to *not* have to mmap guest memory ("even for > ordinary VMs"). Now people are saying we have to be able to mmap it in order > to GUP it. It's getting tiring, really. >From the pKVM side, we're working on guest_memfd primarily to avoid diverging from what other CoCo solutions end up using, but if it gets de-featured (e.g. no huge pages, no GUP, no mmap) compared to what we do today with anonymous memory, then it's a really hard sell to switch over from what we have in production. We're also hoping that, over time, guest_memfd will become more closely integrated with the mm subsystem to enable things like hypervisor-assisted page migration, which we would love to have. Today, we use the existing KVM interfaces (i.e. based on anonymous memory) and it mostly works with the one significant exception that accessing private memory via a GUP pin will crash the host kernel. If all guest_memfd() can offer to solve that problem is preventing GUP altogether, then I'd sooner just add that same restriction to what we currently have instead of overhauling the user ABI in favour of something which offers us very little in return. On the mmap() side of things for guest_memfd, a simpler option for us than what has currently been proposed might be to enforce that the VMM has unmapped all private pages on vCPU run, failing the ioctl if that's not the case. It needs a little more tracking in guest_memfd but I think GUP will then fall out in the wash because only shared pages will be mapped by userspace and so GUP will fail by construction for private pages. We're happy to pursue alternative approaches using anonymous memory if you'd prefer to keep guest_memfd limited in functionality (e.g. preventing GUP of private pages by extending mapping_flags as per [1]), but we're equally willing to contribute to guest_memfd if extensions are welcome. What do you prefer? Cheers, Will [1] https://lore.kernel.org/r/4b0fd46a-cc4f-4cb7-9f6f-ce19a2d3064e@xxxxxxxxxx