svn commit: r244469 - head/sys/arm/arm

[Ian Lepore]

On Tue, 2012-12-25 at 08:19 +1100, Peter Jeremy wrote:
> On 2012-Dec-20 00:35:27 +0000, Olivier Houchard <cognet at FreeBSD.org> wrote:
> >Author: cognet
> >Date: Thu Dec 20 00:35:26 2012
> >New Revision: 244469
> >URL: http://svnweb.freebsd.org/changeset/base/244469
> >
> >Log:
> >  Use the new allocator in bus_dmamem_alloc().
> >
> >Modified:
> >  head/sys/arm/arm/busdma_machdep-v6.c
> 
> Do you have any plans to make similar changes to the other busdma_machdep.c
> files?
> 

I wrote that allocator code to address some specific needs of
architectures that use software-assisted cache coherency rather than
hardware bus snooping.  On such platforms, the bus_dmamem_alloc()
implementations treat BUS_DMA_COHERENT as a request for uncacheable
memory, and some drivers have come to rely on that for correct
operation.  (The concept is valid, overloading BUS_DMA_COHERENT to have
different meanings on different platforms, maybe not so much so.)

The old code on arm had the side effect of making an entire page
uncacheable if any small dma buffer got allocated within that page. That
was bad for performance, but it worked.  When armv6 came along that
caused problems with ldrex/strex instructions, which will fault if they
operate on uncacheable memory.

So the main thrust of the allocation changes was to make it easy to
efficiently allocate small(-ish) dma buffers from a pool of pages that
have caching turned off, without accidentally affecting memory that
isn't being used for dma buffers.  A nice secondary benefit is that they
help eliminate partial cacheline flushes, because the allocator ensures
that all dma buffers are aligned and padded to cacheline boundaries.

So any busdma implementation that has similar needs can benefit from the
new allocator code.  I was thinking MIPS would be a good candidate, but
I've only glanced at the mips code and noted its superficial similarity
to the arm code.

-- Ian