Re: [Patch] dma_sync_to_device

[Matt Porter <mporter@kernel.crashing.org>]

On Wed, Feb 11, 2004 at 10:30:56AM -0800, David S. Miller wrote:
> On Wed, 11 Feb 2004 11:18:00 -0700
> Matt Porter <mporter@kernel.crashing.org> wrote:
> 
> > Sure, other non cache coherent arch's that I'm aware of (PPC, ARM, etc.)
> > already implement the least expensive cache operations based on the
> > direction parameter in pci_dma_sync_single(). On PPC, we do the right
> > thing based on each of three valid directions, I don't yet see what
> > additional information pci_dma_sync_to_device_single() provides. 
> 
> There are two points in time where you want to sync:
> 
> 1) Right after the device has done a DMA transaction, and the cpu
>    wishes to read/write the datum.
> 
> 2) Right after the cpu has read/write the datum, and we like to let the
>    device DMA to/from the thing again.
>
> That is the distinction provided by the two interfaces.
 
Ok, this distinction helps a bit.

> Consider something like MIPS, cache flushes needed for both of the above
> operations:
> 
> 1) pci_map_single(), device DMA's from the buffer.
> 
> 2) pci_dma_sync_single().  Cpu writes some new command or
>    status flag into the buffer.
> 
> 3) pci_dma_sync_to_device_single(), now device is asked to DMA from the buffer
>    again.
> 
> Cache flushes are needed on MIPS for both step #2 and #3, and different kinds of
> flushes in fact.
> 
> Do you understand the need for this now?

Actually, not yet.  Is it not possible for MIPS to determine the correct
cache operation to use if step #3 used a pci_dma_sync_single() with a
TO_DEVICE direction?  

I'm guessing that MIPS must have some kind of bridge cache in order to
require the pci_dma_sync_to_device_single() if I'm starting to follow
this.

-Matt