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

[Feng Tang]

On Thu, Aug 01, 2019 at 02:19:53PM +0800, Rong Chen wrote:
> >>>>>>>>>
> >>>>>>>>>commit: 90f479ae51afa45efab97afdde9b94b9660dd3e4 ("drm/mgag200: Replace struct mga_fbdev with generic framebuffer emulation")
> >>>>>>>>>https://kernel.googlesource.com/pub/scm/linux/kernel/git/next/linux-next.git master
> >>>>>>>>Daniel, Noralf, we may have to revert this patch.
> >>>>>>>>
> >>>>>>>>I expected some change in display performance, but not in VM. Since it's
> >>>>>>>>a server chipset, probably no one cares much about display performance.
> >>>>>>>>So that seemed like a good trade-off for re-using shared code.
> >>>>>>>>
> >>>>>>>>Part of the patch set is that the generic fb emulation now maps and
> >>>>>>>>unmaps the fbdev BO when updating the screen. I guess that's the cause
> >>>>>>>>of the performance regression. And it should be visible with other
> >>>>>>>>drivers as well if they use a shadow FB for fbdev emulation.
> >>>>>>>For fbcon we should need to do any maps/unamps at all, this is for the
> >>>>>>>fbdev mmap support only. If the testcase mentioned here tests fbdev
> >>>>>>>mmap handling it's pretty badly misnamed :-) And as long as you don't
> >>>>>>>have an fbdev mmap there shouldn't be any impact at all.
> >>>>>>The ast and mgag200 have only a few MiB of VRAM, so we have to get the
> >>>>>>fbdev BO out if it's not being displayed. If not being mapped, it can be
> >>>>>>evicted and make room for X, etc.
> >>>>>>
> >>>>>>To make this work, the BO's memory is mapped and unmapped in
> >>>>>>drm_fb_helper_dirty_work() before being updated from the shadow FB. [1]
> >>>>>>That fbdev mapping is established on each screen update, more or less.
> >>>>>> From my (yet unverified) understanding, this causes the performance
> >>>>>>regression in the VM code.
> >>>>>>
> >>>>>>The original code in mgag200 used to kmap the fbdev BO while it's being
> >>>>>>displayed; [2] and the drawing code only mapped it when necessary (i.e.,
> >>>>>>not being display). [3]
> >>>>>Hm yeah, this vmap/vunmap is going to be pretty bad. We indeed should
> >>>>>cache this.
> >>>>>
> >>>>>>I think this could be added for VRAM helpers as well, but it's still a
> >>>>>>workaround and non-VRAM drivers might also run into such a performance
> >>>>>>regression if they use the fbdev's shadow fb.
> >>>>>Yeah agreed, fbdev emulation should try to cache the vmap.
> >>>>>
> >>>>>>Noralf mentioned that there are plans for other DRM clients besides the
> >>>>>>console. They would as well run into similar problems.
> >>>>>>
> >>>>>>>>The thing is that we'd need another generic fbdev emulation for ast and
> >>>>>>>>mgag200 that handles this issue properly.
> >>>>>>>Yeah I dont think we want to jump the gun here.  If you can try to
> >>>>>>>repro locally and profile where we're wasting cpu time I hope that
> >>>>>>>should sched a light what's going wrong here.
> >>>>>>I don't have much time ATM and I'm not even officially at work until
> >>>>>>late Aug. I'd send you the revert and investigate later. I agree that
> >>>>>>using generic fbdev emulation would be preferable.
> >>>>>Still not sure that's the right thing to do really. Yes it's a
> >>>>>regression, but vm testcases shouldn run a single line of fbcon or drm
> >>>>>code. So why this is impacted so heavily by a silly drm change is very
> >>>>>confusing to me. We might be papering over a deeper and much more
> >>>>>serious issue ...
> >>>>It's a regression, the right thing is to revert first and then work
> >>>>out the right thing to do.
> >>>Sure, but I have no idea whether the testcase is doing something
> >>>reasonable. If it's accidentally testing vm scalability of fbdev and
> >>>there's no one else doing something this pointless, then it's not a
> >>>real bug. Plus I think we're shooting the messenger here.
> >>>
> >>>>It's likely the test runs on the console and printfs stuff out while running.
> >>>But why did we not regress the world if a few prints on the console
> >>>have such a huge impact? We didn't get an entire stream of mails about
> >>>breaking stuff ...
> >>The regression seems not related to the commit.  But we have retested
> >>and confirmed the regression.  Hard to understand what happens.
> >Does the regressed test cause any output on console while it's
> >measuring? If so, it's probably accidentally measuring fbcon/DRM code in
> >addition to the workload it's trying to measure.
> >
> 
> Sorry, I'm not familiar with DRM, we enabled the console to output logs, and
> attached please find the log file.
> 
> "Command line: ... console=tty0 earlyprintk=ttyS0,115200
> console=ttyS0,115200 vga=normal rw"

We did more check, and found this test machine does use the
mgag200 driver. 

And we are suspecting the regression is caused by 

commit cf1ca9aeb930df074bb5bbcde55f935fec04e529
Author: Thomas Zimmermann <tzimmermann at suse.de>
Date:   Wed Jul 3 09:58:24 2019 +0200

    drm/fb-helper: Map DRM client buffer only when required
    
    This patch changes DRM clients to not map the buffer by default. The
    buffer, like any buffer object, should be mapped and unmapped when
    needed.
    
    An unmapped buffer object can be evicted to system memory and does
    not consume video ram until displayed. This allows to use generic fbdev
    emulation with drivers for low-memory devices, such as ast and mgag200.
    
    This change affects the generic framebuffer console. HW-based consoles
    map their console buffer once and keep it mapped. Userspace can mmap this
    buffer into its address space. The shadow-buffered framebuffer console
    only needs the buffer object to be mapped during updates. While not being
    updated from the shadow buffer, the buffer object can remain unmapped.
    Userspace will always mmap the shadow buffer.
 
which may add more load when fbcon is busy printing out messages.

We are doing more test inside 0day to confirm.

Thanks,
Feng