On Sat, Mar 29, 2025 at 12:31 PM Rakesh J <rakeshjb...@gmail.com> wrote: > Paolo: You pointed out the size issue with .min_access_size = 1 and > .max_access_size = 4, where bswap32 was wrong for 2-byte accesses. I’ve fixed > this with size-appropriate swaps (bswap16 for 2-byte, bswap32 for 4-byte). On > the extra swap idea, I stuck with a single swap since it aligns PCI LE with > guest BE expectations without overcomplicating it—let me know if I > misunderstood.
The extra swap (compared to what the "regular" PCI data ops do) is exactly what you were doing. > I’ve sent [PATCH v2] incorporating changes: > 1.Removed gt64120_update_pci_cfgdata_mapping() and moved initialization code > to gt64120_realize() for a simpler MByteSwap check. > 2.Removed unused pci_host_data_be_ops and a misleading comment in dino.h > > 3.Size-specific swaps (bswap16 and bswap32) > I included bswap16 for 2-byte accesses in v2—should this be restricted to > 4-byte only (bswap32) per the spec, or does GT-64120 expect 2-byte config > swaps too? It’s a minor tweak, so I left it in v2 for now—happy to adjust in > a v3 if needed. Which swap to use is not really related to what the GT-64120 does, but to the interface between memory.c and the MemoryRegionOps. When access_size == 2, QEMU wants the result in bits 0..15 so you need to use bswap16. With bswap32, the result would be in bits 16..31. Paolo
