[ofa-general] Re: mthca use of dma_sync_single is bogus

[Roland Dreier]

 > >     void
 > >     dma_sync_single_range(struct device *dev, dma_addr_t dma_handle,
 > >     		      unsigned long offset, size_t size,
 > >     		      enum dma_data_direction direction)

 > This is under Part II - Advanced dma_ usage - I don't think it's dealing with
 > non-consistent memory only (e.g. dma_declare_coherent_memory is there), and this
 > looks like a good fit.  Most functions here work for both consistent and
 > non-consistent memory...  What makes you suspicious?

I was suspicious because it is described between the main noncoherent
API stuff and dma_cache_sync().  But I think it is probably OK.

Unfortunately it is not that good a fit for our current code, since we
use pci_map_sg() to do the DMA mapping on the MTT memory instead of
dma_map_single().

 > I'm concerned that MTTs need a fair amount of memory,
 > while the amount of coherent memory might be limited.
 > Not that non-coherent memory systems are widespread ...

Yes, for example on ppc 4xx the amount of coherent memory is quite
small by default (address space for non-cached mappings is actually
what is limited, but it amounts to the same thing).

Maybe the least bad solution is to change to using dma_map_single()
instead of pci_map_sg() in mthca_memfree.c.

 - R.