Messages in this thread | | | Subject | Re: [PATCH 0/3] RFC: addition to DMA API | From | Mark Salter <> | Date | Thu, 01 Sep 2011 15:14:18 -0400 |
| |
On Thu, 2011-09-01 at 18:31 +0100, Will Deacon wrote: > I don't think what we're seeing in this case is caused by mismatched memory > attributes, especially as passing `nosmp' on the command-line makes the > performance issue disappear.
I'm coming to think we are dealing with two different problems.
We have the original problem where adding the write buffer flush to EHCI gives a 4x performance boost to USB. Also adding nosmp to the cmdline gives pretty much the same boost. This is looking like something other than just data getting held up in a write buffer.
On the other hand, on a nosmp kernel, I get about 3-4% performance boost for hdparm -t using the write buffer flush patch vs. without it.
So, regardless of what turns out to be the actual cause of the 4x problem, it may still be worthwhile to have the explicit write buffer sync API if we can't avoid using buffered mappings for DMA.
--Mark
| |