On 7/31/2020 1:48 PM, Jason Gunthorpe wrote: > On Fri, Jul 31, 2020 at 01:15:34PM -0400, Steven Sistare wrote: >> On 7/31/2020 12:56 PM, Jason Gunthorpe wrote: >>> On Fri, Jul 31, 2020 at 12:11:52PM -0400, Steven Sistare wrote: >>>>> Your preservation-across-exec use-case might or might not need the >>>>> VMA to be mapped at the same address. >>>> >>>> It does. qemu registers memory with vfio which remembers the va's in kernel >>>> metadata for the device. >>> >>> Once the memory is registered with vfio the VA doesn't matter, vfio >>> will keep the iommu pointing at the same physical pages no matter >>> where they are mapped. >> >> Yes, but there are other code paths that compute and use offsets between va and the >> base va. Mapping at a different va in the new process breaks vfio; I have tried it. > > Maybe you could fix vfio instead of having this adventure, if vfio is > the only motivation. Maybe. We still need to preserve an anonymous segment, though. MADV_DOEXEC, or mshare, or something else. And I think the ability to preserve memory containing pointers to itself is an interesting use case, though not ours. - Steve
