Re: [Qemu-devel] [PATCH 1/5] Add target memory mapping API

[Avi Kivity <avi@redhat.com>]

Jamie Lokier wrote:
> Linux itself had some issues with _its_ DMA API recently: people have
> been writing drivers using the Linux DMA API making broken assumptions
> that happen to work on x86, and work some of the time on other
> architectures.  These things don't show up during driver tests, and
> are very difficult to track down later.  I suspect "overwrites the
> remaining buffer with randomness/zeros but only under bounce-buffer
> conditions" will be similarly unlikely to trigger, and very difficult
> to track down if it causes a problem anywhere.
>   

This reminds me -- Gleb pointed out that Linux itself bounces some DMA 
operations, and will copy the entire range.  So Linux itself doesn't 
satisfy these requirements.

> The recent solution in Linux is to add some debugging options which
> check it's used correctly, even on x86.
>
> Here's a final thought, to do with performance:
>
> e1000, configured for jumbo frames, receives logs of small packets,
> and the DMA subsystem has to bounce-copy the data for some reason (Ian
> suggested maybe always doing that with Xen for DMA to guest RAM?)
>
> Without a length passed to unmap, won't it copy 65536 bytes per packet
> (most from cold cache) because that's the amount set up for DMA to
> receive from /dev/tap, instead of 256 or 1514 bytes per packet which
> is their actual size?
>   

It will.  If we bounce.  We'll never bounce in real life.

-- 
Do not meddle in the internals of kernels, for they are subtle and quick to panic.