Re: [Qemu-devel] [PATCH 2/6] PCI DMA API

[Paul Brook <paul@codesourcery.com>]

On Sunday 30 March 2008, Anthony Liguori wrote:
> Paul Brook wrote:
> > On Saturday 29 March 2008, Anthony Liguori wrote:
> >> This patch introduces a PCI DMA API and some generic code to support
> >> other DMA APIs.  Two types are introduced: PhysIOVector and IOVector.  A
> >> DMA API maps a PhysIOVector, which is composed of target_phys_addr_t,
> >> into an IOVector, which is composed of void *.
> >
> > Devices should not be using IOVector. They should either use the DMA copy
> > routines to copy from a PhysIOVector into a local buffer, or they should
> > pass a PhysIOVector to a block/network read/write routine. The DMA API
> > should allow devices to be agnostic about how DMA is implemented. They
> > should not be trying to manually implement zero copy.
>
> Someone has to do the translation of PhysIOVector => IOVector.  It
> doesn't seem logical to me to do it in the IO backend level because the
> block subsystem doesn't know how to do that translation.  You would have
> to pass the PhysIOVector although with a translation function and an
> opaque pointer.

The entity processing the data shouldn't need to know or care how the 
translation is done. PhysIOVector should describe everything it need to know.

> What could work is if the DMA API functions mapped PhysIOVector =>
> PhysIOVector and then the network and block subsystems could operate on
> a PhysIOVector.  I have patches that implement vector IO for net and
> block but didn't want to include them in this series to keep things simple.

IMHO this is the only sane way to implement zero-copy.

> >> This enables zero-copy IO to be preformed without introducing
> >> assumptions of phys_ram_base.  This API is at the PCI device level to
> >> enable support of per-device IOMMU remapping.
> >
> > By my reading it *requires* bridges be zero-copy.  For big-endian targets
> > we need to ability to byteswap accesses.
>
> You mean via ld/st_phys?  

By whatever means the bridge deems necessary. The whole point of the DMA API 
is that you're transferring a block of data. The API allows intermediate 
busses to transform that data (and address) without the block handler needing 
to know or care.

With your current scheme a byteswapping bus has to allocate a single large 
buffer for the whole vector, even if the device then ends up copying unto a 
local buffer in small chunks.

Paul