GitLab

The bridge was incorrectly configured with 2 overlapping mappings (0-&FFFFFFFF and another 0-&8000000). When the VL805 wanted to bus master into main memory the bridge was confused where the transaction should go, causing a master abort to be logged on the secondary side.
Reorganise the PCI memory so that only a small window in the top 1GB is actively decoded (we can't bus master above 3GB anyway due to a chip design "feature"), meaning the rest of the secondary address space is forwarded 1:1 to the primary.
Further, because the CPU side windows can only be sized in powers of 2, reduce that to 2GB in size.