Re: [LKP] [drm/mgag200] 90f479ae51: vm-scalability.median -18.8% regression

[Daniel Vetter]

On Wed, Sep 4, 2019 at 1:15 PM Dave Airlie <airlied@xxxxxxxxx> wrote:
>
> On Wed, 4 Sep 2019 at 19:17, Daniel Vetter <daniel@xxxxxxxx> wrote:
> >
> > On Wed, Sep 4, 2019 at 10:35 AM Feng Tang <feng.tang@xxxxxxxxx> wrote:
> > >
> > > Hi Daniel,
> > >
> > > On Wed, Sep 04, 2019 at 10:11:11AM +0200, Daniel Vetter wrote:
> > > > On Wed, Sep 4, 2019 at 8:53 AM Thomas Zimmermann <tzimmermann@xxxxxxx> wrote:
> > > > >
> > > > > Hi
> > > > >
> > > > > Am 04.09.19 um 08:27 schrieb Feng Tang:
> > > > > >> Thank you for testing. But don't get too excited, because the patch
> > > > > >> simulates a bug that was present in the original mgag200 code. A
> > > > > >> significant number of frames are simply skipped. That is apparently the
> > > > > >> reason why it's faster.
> > > > > >
> > > > > > Thanks for the detailed info, so the original code skips time-consuming
> > > > > > work inside atomic context on purpose. Is there any space to optmise it?
> > > > > > If 2 scheduled update worker are handled at almost same time, can one be
> > > > > > skipped?
> > > > >
> > > > > To my knowledge, there's only one instance of the worker. Re-scheduling
> > > > > the worker before a previous instance started, will not create a second
> > > > > instance. The worker's instance will complete all pending updates. So in
> > > > > some way, skipping workers already happens.
> > > >
> > > > So I think that the most often fbcon update from atomic context is the
> > > > blinking cursor. If you disable that one you should be back to the old
> > > > performance level I think, since just writing to dmesg is from process
> > > > context, so shouldn't change.
> > >
> > > Hmm, then for the old driver, it should also do the most update in
> > > non-atomic context?
> > >
> > > One other thing is, I profiled that updating a 3MB shadow buffer needs
> > > 20 ms, which transfer to 150 MB/s bandwidth. Could it be related with
> > > the cache setting of DRM shadow buffer? say the orginal code use a
> > > cachable buffer?
> >
> > Hm, that would indicate the write-combining got broken somewhere. This
> > should definitely be faster. Also we shouldn't transfer the hole
> > thing, except when scrolling ...
>
> First rule of fbcon usage, you are always effectively scrolling.
>
> Also these devices might be on a PCIE 1x piece of wet string, not sure
> if the numbers reflect that.

pcie 1x 1.0 is 250MB/s, so yeah with a bit of inefficiency and
overhead not entirely out of the question that 150MB/s is actually the
hw limit. If it's really pcie 1x 1.0, no idea where to check that.
Also might be worth to double-check that the gpu pci bar is listed as
wc in debugfs/x86/pat_memtype_list.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch