Re: [Mesa3d-dev] dynamic code generation with gcc

[Gareth H. <ga...@va...>]

Josh Vanderhoof wrote:
> 
> I was too vague.  You could mprotect() the page after the DMA buffer
> to PROT_NONE and install a SIGSEGV handler that flushes the buffer if
> the SIGSEGV was on the end of the DMA buffer.  Then you can just let
> Vertex3f segfault when it uses the buffer up.  You would save yourself
> a test and a predicted jump per call.  Not much, but it sounds like
> you really want to optimize this.

Sorry, I get you.  You might even be able to do fancy things and catch
this in the kernel module -- that would be neat :-)  Save us an ioctl...

However, the way we manage DMA space will change soon and this will no
longer be possible.  The drivers will be using a dynamic allocation
scheme to get better utilization of the DMA space.  Fixed size buffers
suck.  I'll play around with similar ideas -- mprotect()ing the DMA
buffer and catching the segfault in the kernel sure would be neat... 
Don't know if we can do that, but it's worth looking at.

-- Gareth