svn commit: r199670 - head/sys/dev/bge

[John Baldwin]

On Monday 23 November 2009 10:06:08 am Scott Long wrote:
> 
> On Nov 23, 2009, at 6:38 AM, John Baldwin wrote:
> 
> > On Sunday 22 November 2009 6:48:18 pm Scott Long wrote:
> >> By definition, PCIe can't transfer across 4GB boundaries.  It's not a
> >> bug specific to Broadcom.  If you're loading dynamic buffers (i.e.
> >> mbufs), setting an appropriate boundary value in the tag will allow
> >> busdma to take care of this.  If you're allocating static buffers,
> >> busdma won't honor this.  But what you've done here is best anyways;
> >> control buffers that are going to be frequently transferred are best
> >> kept in the lower 4GB of the address space.  It simplifies PCIe
> >> handling, and it's significantly faster on PCI/PCI-X.  So I'd suggest
> >> making this the rule rather than the exception in the driver.
> >
> > Should we enforce an implicit 4GB boundary in bus_dma then?  Perhaps  
> > Host-PCI
> > bridge drivers should create a tag with a 4GB boundary that devices  
> > inherit
> > via bus_get_dma_tag().  For i386/PAE we might should always enforce  
> > a 4GB
> > boundary as well?
> 
> That was actually the point of creating bus_get_dma_tag().  I don't  
> recall how complete the back-end work of providing a inheritance tree  
> for bridges got.  Note that having this wouldn't really solve Pyun's  
> problem, because boundaries aren't honored for static allocations.   
> There's been plenty of talk about multi-segment static allocations,  
> but I don't think that that's applicable to this either.  As I said  
> before, it's best to restrict static allocations to the lower 4GB of  
> memory and not worry about the boundary at all.  But having the proper  
> inheritance would still be nice.
> 
> Another thing that I'd like to do is have an alternate for  
> bus_dma_tag_create() that takes the device_t of the device as the  
> first argument and figures out the parent tag and inheritance  
> automatically.  That's a job for the mythical /sys/kern/subr_busdma.c.

The inheritance thing will actually work as the default version of 
bus_dma_get_tag() is to just pass the request up the tree.  As long as 
devices always use bus_dma_get_tag() when creating the tag they pass down to 
their children via their own bus_dma_get_tag() implementation all the various 
restrictions should be passed down the tree.

-- 
John Baldwin