Using asm-generic/dma-mapping.h does not work because pushing
the call down to pci_alloc_coherent() causes the gfp_t argument
of dma_alloc_coherent() to be ignored.

Fix this by implementing things directly, and adding a gfp_t
argument we can use in the internal call down to the PCI DMA
implementation of pci_alloc_coherent().

This fixes massive memory corruption when using the sound driver
layer, which passes things like __GFP_COMP down into these
routines and (correctly) expects that to work.
Signed-off-by: David S. Miller <davem@davemloft.net>