On 9 November 2014 11:43, Ard Biesheuvel <ard.biesheuvel@xxxxxxxxxx> wrote: > On 7 November 2014 20:13, Christoffer Dall <christoffer.dall@xxxxxxxxxx> wrote: >> Hi Magnus, >> >> I think the check in is_mmio_pfn() should be modified if it is common to >> have coherent memory which is marked as reserved to Linux. However, I'm >> not entirely sure which other solutions would work with all other >> architectures in KVM using this check. >> >> I'm cc'ing Paolo and Ard here (Ard recently looked into this area as >> well), but it may be worth trying to summarize the problem a little more >> consicely and ask the question at a braoder scope, for example on >> kvm@xxxxxxxxxxxxxxx. >> > > What is ultimately biting us here (and the fact that I had to add > !is_zero_pfn() to kvm_is_mmio_pfn() was already a strong hint) is that > kvm_is_mmio_pfn() is the wrong test to perform against a pfn to decide > how it should be mapped. > > kvm_is_mmio_pfn() is used on the other archs to distinguish 'special' > pfns from the page frames that take part in ordinary demand paging by > the host kernel, and that can be reassigned once the refcount drops to > zero. kvm_is_mmio_pfn() == true is a necessary but not sufficient > condition to decide that pfn points to device memory. > > I think the proper way to deal with this would be to rename > kvm_is_mmio_pfn() to kvm_is_reserved_pfn() and drop the is_zero_pfn() > test from it, effectively reverting my change for all archs except > ARM. Then, we implement a kvm_is_mmio_pfn() that lives up to its name, > and only returns false for regions that are in fact MMIO and not I meant 'returns true', of course. > reserved RAM. > -- Ard. _______________________________________________ kvmarm mailing list kvmarm@xxxxxxxxxxxxxxxxxxxxx https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
