On Sat, Feb 2, 2013 at 8:18 PM, Peter Maydell <peter.mayd...@linaro.org> wrote: > On 2 February 2013 18:26, Blue Swirl <blauwir...@gmail.com> wrote: >> On Sat, Feb 2, 2013 at 5:44 PM, Peter Maydell <peter.mayd...@linaro.org> >> wrote: >>> On 2 February 2013 17:37, Andreas Färber <afaer...@suse.de> wrote: >>>> Am 02.02.2013 17:49, schrieb Peter Maydell: >>>> libqtest.h has no generic endian-aware memread functions unlike Alex, >>>> you or me expected. It reads a sequence of bytes from guest memory and >>>> transmits them one-by-one over the text-based qtest protocol. >>> >>> OK, so this is just busted for accessing devices. The protocol >>> has to have some way of letting you do a 32 bit / 16 bit / 8 bit >>> access (and maybe 64 bit as well while we're here). memread >>> and memwrite are OK for RAM accesses [ie anything you'd be >>> happy to have cached or buffered in a real system] but for >>> memory mapped registers we need to have an equivalent of >>> inb/inw/inl/outb/outw/outl that guarantee to do exactly one >>> access of exactly the required width. >> >> I was also just making a patch (but not so nice as Andreas'). My >> analysis was that qtest.c and libqtest.c just pass the result of >> cpu_physical_memory_rw() as is, with no endian conversion. So the >> result needs to be converted to host CPU order just like Andreas did. >> I think the protocol is OK, my initial reaction was to put byte >> swapping there but that's not right. > > No, I think the protocol is broken, even if you ignore the > question of endianness. Consider that a device's MemoryRegion > can have different behaviour depending on whether you access it > as a byte, halfword or word size. cpu_physical_memory_rw() won't > let you make a word size access to an unaligned address.
But the width is transmitted too, so any size accesses can be performed. > > -- PMM