Instead of directly calling the low-level invalidate_dcache_range() and
flush_cache() functions, provide thin wrappers that take into account
alignment requirements.

While at it, fix a case where the cache was flushed but should have been
invalidated, two cases where the buffer data was flushed instead of the
descriptor and a missing cache invalidation before reading the packet
data that the NIC just wrote to memory.

Signed-off-by: Thierry Reding <treding@nvidia.com>
Patch: 276474