[This is a fairly complex topic. Perhaps because it isn't explained well in one place.] | From: Matt Roth <mroth@xxxxxxxxxx> | 3) 32-bit hardware (such as PCI ethernet cards) can be problematic. I have a | Level 5 Networks EtherFabric EF1-21022T that required a custom driver for our | system. I believe it was a DMA bounce buffering issue. This kind of gotcha is also not well known. Too many people discover this by crashing into it. My notebook has a 802.11g interface that screws up if the DMA address is above 1G. This is buried in a proprietary driver and had to be managed by ndiswrapper (after the problem was properly diagnosed). Just last night we were chatting about fibrechannel device driver problems at the Toronto LUG. Although PCI has had 64-bit addressing for a while, a lot of controllers were too lazy to get this right. Intel's x86_64 chips did not support I/O into anything past the first 4G of physical memory (no IOMMU), at least initially. So bounce buffers are still needed. That's why early versions of Fedora for x86_64 did not work on Intel. Perhaps Intel has fixed this by now. This was one reason that those who knew what they were doing preferred AMD. (I remember a similar problem with 486 systems having more that 16M of RAM and an ISA bus.)
