From: Anton Blanchard <anton@samba.org>

Introduce dma_error() and pci_dma_error() which are used to detect failures
in pci_map_single.

From: "David S. Miller" <davem@redhat.com>

Currently, for an existing DMA mapping, there is a way to transfer buffer
ownership back to the cpu, yet there is no way to give it back to the device
again explicitly.  The latter really is needed on platforms where the PCI
subsystem does not snoop the cpu caches, MIPS is one example.

Many drivers were expecting the existing DMA sync interface to handle both
directions, which was wrong.

Now, with this change, we have explicit interfaces for DMA syncing to/from
the device and the cpu.

dma_alloc_[non]coherent now takes the GFP_ flags as the last argument.
The flags passed in may not interfere with the memory zone.

The compat API doesn't cope correctly if struct pci_dev is NULL, as it
can be for the x86 fake pci_dev handling.

use a #include mechanism for generic implementations of the pci_
API in terms of the dma_ one