Re: Excessive kernel time inside ttm when scrolling on a r700 in a multimonitor setup

[Andres Freund <andres@anarazel.de>]

Hi,

On 2012-09-17 13:15:14 +0200, Andres Freund wrote:
> On Monday, September 17, 2012 11:59:34 AM Jerome Glisse wrote:
> > On Mon, Sep 17, 2012 at 5:13 AM, Andres Freund <andres@anarazel.de> wrote:
> > > On Monday, September 17, 2012 10:47:10 AM Jerome Glisse wrote:
> > >> On Sun, Sep 16, 2012 at 7:18 PM, Andres Freund <andres@anarazel.de> wrote:
> > >> > Hi everyone,
> > >> > 
> > >> > In several gui programs, most notably and reproducably, I can
> > >> > trivially make Xorg spend the whole cpu time in the kernel. It might
> > >> > be an Xorg bug, but from my untrained gut feeling it doesn't look
> > >> > that way.
> > >> > 
> > >> > Anything more you need than whats already in the email?
> > >> 
> > >> > Here are some details:
> > >> Is it an AGP card ? It looks like it's the PAT update that kills perf
> > >> for you, not much we can do, the pool allocator is probably starve.
> > > 
> > > If you mean AGP as in the old strange graphics bus, no. Its a pcie card.
> > > 
> > > If that helps you, more detailedly the card is a: SAPPHIRE Radeon HD6870
> > > 
> > > Unfortunately the manufacturers page is remarkably devoid of details...:
> > > http://www.sapphiretech.com/presentation/product/?cid=1&gid=3&sgid=1037&l
> > > id=1&pid=1270&leg=0
> > > 
> > > Greetings,
> > > 
> > > Andres
> > 
> > Well, then the issue is that it tries to allocate write combined pages
> > on PCIE which should not happen. I don't see how in the code you endup
> > in that path unless drm_pci_device_is_agp return true for some
> > reasons.

I've recently searched for the same problem on slightly different
hardware, and found my own old post. A bit of further digging and code
reading lead to the discovery that reconfiguring the kernel to not use
SWIOTLB but rather VT-D solved the problem.
i.e once
Jun 10 21:08:19 awork2 kernel: PCI-DMA: Intel(R) Virtualization Technology for Directed I/O
instead of
Jun 08 21:19:56 awork2 kernel: PCI-DMA: Using software bounce buffering for IO (SWIOTLB)
redrawing wasn't slow anymore.

If I see correctly it is the following codepath that doesn't work well
on multi socket systems:
static int radeon_ttm_tt_populate(struct ttm_tt *ttm)
{
...
#ifdef CONFIG_SWIOTLB
	if (swiotlb_nr_tbl()) {
		return ttm_dma_populate(&gtt->ttm, rdev->dev);
	}
#endif
...


I have no idea whether this is an expected problem, a bug, a scalability
limitation or anything. I primarily brought it up because somebody else
(or I myself) might later end up finding this thread again.

If there's something that could possibly be done about this, I'm happy
to help researching.

Greetings,

Andres Freund