Re: [Patch] dma_sync_to_device

["David S. Miller" <davem@redhat.com>]

On Wed, 11 Feb 2004 20:46:11 -0700
Deepak Saxena <dsaxena@plexity.net> wrote:

> /me groks now. I am assuming this is a 2.7 thing as it is
> reinterpreting/redefining the API.

No, it is adding an operation that previously was impossible to
make occur.

> In ARM, pci_dma_sync_single() does
> a cache flush, which is why I was confused asked about two cache flushes.
> What you are proposing is that by definition pci_dma_sync_* syncs 
> bridges caches with system memory, while pci_dma_sync_device_* syncs 
> the CPU's cache with system memory.
> 
> This will definetely confuse a lot of driver writers. 

What I am proposing, is to keep pci_dma_sync_*() with it's current definition
which is to transfer control of the buffer from DEVICE to CPU.  It has always
been defined this way.

I am also adding a new operation, pci_dma_sync_device_*() which transfers control
of the buffer from CPU back to DEVICE.  This operation was not possible previously.

Therefore we will merge this into both 2.6.x and 2.4.x as soon as a suitable full
and reviewed patch exists.

If you flush the cache on ARM for pci_dma_sync_*(), so what?  If the cpu writes to
the buffer then asks the device to DMA from the area it's going to see garbage.

To be honest there aren't many pci_dma_sync_*() users, and thus it is relatively
easy to audit them all for problems in this area.  Yes, updates and clarifications
in the docs will be necessary as well.