On Thu, Feb 23, 2017 at 03:29:53PM +0000, Peter Maydell wrote: > On 23 February 2017 at 15:21, Paolo Bonzini <pbonz...@redhat.com> wrote: > > > > > > On 23/02/2017 15:35, Peter Maydell wrote: > >> On 23 February 2017 at 12:53, Paolo Bonzini <pbonz...@redhat.com> wrote: > >>> > >>> > >>> On 23/02/2017 13:26, Peter Maydell wrote: > >>>> On 23 February 2017 at 11:43, Paolo Bonzini <pbonz...@redhat.com> wrote: > >>>>> On 23/02/2017 12:34, Peter Maydell wrote: > >>>>>> We should probably update the doc comment to note that the > >>>>>> pointer is to host-endianness memory (and that this is not > >>>>>> like normal RAM which is target-endian)... > >>>>> > >>>>> I wouldn't call it host-endianness memory, and I disagree that normal > >>>>> RAM is target-endian---in both cases it's just a bunch of bytes. > >>>>> > >>>>> However, the access done by the MemoryRegionOps callbacks needs to match > >>>>> the endianness declared by the MemoryRegionOps themselves. > >>>> > >>>> Well, if the guest stores a bunch of integers to the memory, which > >>>> way round do you see them when you look at the bunch of bytes? > >>> > >>> You see them in whatever endianness the guest used. > >> > >> I'm confused. I said "normal RAM and this ramdevice memory are > >> different", and you seem to be saying they're the same. I don't > >> think they are (in particular I think with a BE guest on an > >> LE host they'll look different). > > > > No, they look entirely the same. The only difference is that they go > > through MemoryRegionOps instead of memcpy. > > Then we have a different problem, because the thing this patch > is claiming to fix is that the memory the device is backed by > (from vfio) is little-endian and we're not accessing it right. > > RAM of the usual sort is target-endian (by which I mean "when the guest > does a write of 32-bits 0x12345678, and you look at the memory byte > by byte then the order of bytes is either 0x12 0x34 0x56 0x78 if > TARGET_LITTLE_ENDIAN or 0x78 0x56 0x34 0x12 if TARGET_BIG_ENDIAN"). > > AIUI what we want for this VFIO case is "when the guest does > a 32-bit write of 0x12345678 then the bytes are 0x12 0x34 0x56 0x78 > regardless of whether TARGET_BIG_ENDIAN or not".
At least in the case of KVM and MMIO emulation (which is the case here), run->mmio.data should be considered as a byte stream without endianness, and what is needed is for QEMU to transfer data between run->mmio.data and the device (or whatever is backing the MMIO region) without any net byte swap. So if QEMU is doing a 32-bit host-endian load from run->mmio.data (for an MMIO store), then it needs to do a 32-bit host-endian store into the device memory. In other words, the RAM memory region needs to be considered host endian to match run->mmio.data being considered host endian. Paul.
