On Wed, 2008-03-12 at 23:07 +0100, Aurelien Jarno wrote: > > We finally found the problem with rtl8139 emulation. The byteswapping > depending on the target was not done for all registers. The 8139too > driver was able to cope with that, but not the 8139cp. In his patch > Hollis addressed the issue by byteswapping accesses to all registers, > but depending the host endianness (cpu_to_ and _to_cpu functions) > instead of the target endianness. This was working correctly in his > case, as he tested it on both big endian target and hosts. > > Byte swapping does not depend on host endianness nor target endianness > but on the path from the CPU to the device, which is currently and > *wrongly* implemented in Qemu as a byteswap on big endian targets. > This > has to be fixed by providing a layer modeling the endianness. It has > to > maintain a bus network/tree, so it can invert the endianness of all > the > devices behind a PCI host controller or a PCI bridge. It also has to > be > dynamic, as some chipsets (e.g. gt64xxx) have a configurable bit that > can be changed at runtime to control the endianness. > > Before it is implemented correctly, please find attached two patches > to > fix the rtl8139 and e1000 emulations on big endian hosts and big > endian > targets, using the current and *wrong* Qemu implementation with regard > to endianness. Hopefully that works as all the systems we support > happen > to do the same thing.
Thanks for pursuing this Aurelien. Although patches that replace cpu_to_* with TARGET_WORDS_BIGENDIAN usage make me cringe, these patches do indeed seem to fix BE/BE host/target behavior right now. -- Hollis Blanchard IBM Linux Technology Center
