It seems to me like a lot of drivers wind up getting their 
pci_set_dma_mask stuff wrong, occasionally in subtle ways.  So, I 
created a "give me 64-bit PCI DMA" helper function.

The attached patch demonstrates its use in tg3, from which the logic 
originated.  It also fixes a tiny bug in tg3, whereby it might return 
zero on error (rather than -EFOO) if the pci_set_dma_mask succeeds but 
pci_set_consistent_dma_mask fails.