On Tue, Nov 28, 2023 at 07:12:33PM +0000, Michael Kelley wrote: > From: kirill.shutemov@xxxxxxxxxxxxxxx <kirill.shutemov@xxxxxxxxxxxxxxx> Sent: Friday, November 24, 2023 2:06 AM > > > > On Tue, Nov 21, 2023 at 01:20:08PM -0800, mhkelley58@xxxxxxxxx wrote: > > > From: Michael Kelley <mhklinux@xxxxxxxxxxx> > > > > > > In a CoCo VM when a page transitions from encrypted to decrypted, or vice > > > versa, attributes in the PTE must be updated *and* the hypervisor must > > > be notified of the change. > > > > Strictly speaking it is not true for TDX. Conversion to shared can be > > implicit: set shared bit and touch the page will do the conversion. MapGPA > > is optional. > > Interesting. Given that, is there a reason to use the explicit > hypervisor callbacks in for private->shared transitions in > __set_mem_enc_pgtable()? It probably doesn't have direct relevance > to this patch series, but I'm just trying to understand the tradeoffs of > the implicit vs. explicit approach. And am I correct that > shared->private transitions must use the explicit approach? It must be explicit in sense, that the memory has to be accepted before use. MapGPA() is still optional. I don't like this implicit tricks. I spent a lot of time debugging an issue that was obscured by this semantics. But I think it is going to say :/ > > > Because there are two separate steps, there's > > > a window where the settings are inconsistent. Normally the code that > > > initiates the transition (via set_memory_decrypted() or > > > set_memory_encrypted()) ensures that the memory is not being accessed > > > during a transition, so the window of inconsistency is not a problem. > > > However, the load_unaligned_zeropad() function can read arbitrary memory > > > pages at arbitrary times, which could read a transitioning page during > > > the window. In such a case, CoCo VM specific exceptions are taken > > > (depending on the CoCo architecture in use). Current code in those > > > exception handlers recovers and does "fixup" on the result returned by > > > load_unaligned_zeropad(). Unfortunately, this exception handling can't > > > work in paravisor scenarios (TDX Paritioning and SEV-SNP in vTOM mode) > > > if the exceptions are routed to the paravisor. The paravisor can't > > > do load_unaligned_zeropad() fixup, so the exceptions would need to > > > be forwarded from the paravisor to the Linux guest, but there are > > > no architectural specs for how to do that. > > > > Hm. Can't we inject #PF (or #GP) into L2 if #VE/#VC handler in L1 sees > > cross-page access to shared memory while no fixup entry for the page in > > L1. It would give L2 chance to handle the situation in a transparent way. > > > > Maybe I miss something, I donno. > > I'm recounting what the Hyper-V paravisor folks say without knowing all the > details. :-( But it seems like any kind of forwarding scheme needs to be a > well-defined contract that would work for both TDX and SEV-SNP. The > paravisor in L1 might or might not be Linux-based, so the contract must be OS > independent. And the L2 guest might or might not be Linux, so there's > potential for some other kind of error to be confused with a Linux > load_unaligned_zeropad() reference. Okay, fair enough. I have hard time reasoning if it is okay for L2 which is not Linux. -- Kiryl Shutsemau / Kirill A. Shutemov
