On Wed, Mar 04, 2020 at 08:04:04PM +0100, Ard Biesheuvel wrote: > On Wed, 4 Mar 2020 at 19:50, Arvind Sankar <nivedita@xxxxxxxxxxxx> wrote: > > > > On Wed, Mar 04, 2020 at 07:44:50PM +0100, Ard Biesheuvel wrote: > > > > > > I've tried a couple of different ways, but I can't seem to get my > > > memory map organized in the way that will trigger the error. > > > > What does yours look like? efi_merge_regions doesn't merge everything > > that will eventually be mapped the same way, so if there are some > > non-conventional memory regions scattered over the address space, it > > might be breaking up the mappings to the point where this doesn't > > trigger. > > I have a region > > [ 0.000000] efi: mem07: [Conventional Memory| | | | | | | | > | |WB|WT|WC|UC] range=[0x0000000001400000-0x00000000b9855fff] > (2948MB) > > which gets covered correctly > > [ 0.401766] 0x0000000000a00000-0x0000000040000000 1014M > RW PSE NX pmd > [ 0.403436] 0x0000000040000000-0x0000000080000000 1G > RW PSE NX pud > [ 0.404645] 0x0000000080000000-0x00000000b9800000 920M > RW PSE NX pmd > [ 0.405844] 0x00000000b9800000-0x00000000b9a00000 2M > RW NX pte > [ 0.407436] 0x00000000b9a00000-0x00000000baa00000 16M > ro PSE x pmd > [ 0.408591] 0x00000000baa00000-0x00000000bbe00000 20M > RW PSE NX pmd > [ 0.409751] 0x00000000bbe00000-0x00000000bc000000 2M > RW NX pte > [ 0.410821] 0x00000000bc000000-0x00000000be600000 38M > RW PSE NX pmd > > However, the fact that you can provide a case where it does fail > should be sufficient justification for taking this patch. I was just > trying to give more than a regression-tested-by No, this case is exactly one that should break. But I think you're running on a processor model that _does_ support GB pages, as shown by the "pud" mapping there for the 1G-2G range. At least for my version of qemu, -cpu Haswell does not enable the pdpe1gb feature. Which cpu did you specify? Thanks.
