Before the dma transfer, pci_map_sg will map the memory space covered by sg_table, and it will flush and invalidate cache indexed by that memory space. After the dma transfer, pci_unmap_sg also needs to be called to flush and invalidate the same cache. But I do not know why the second flush will be demanded. I think that in the interval between the two flush, there would be nothing to access the memory covered by the dma. But it is not the case. I want to know what can access that memory and I need some help.
In 2.4, memcpy's prefetch may (and, in practice, do, no smiles, it cost me a lot of time to realize) access that memory. I though it has been fixed in 2.6 someday.
Regards, Gleb.
