Re: [Intel-gfx] [PULL] drm-intel-next-fixes

[Rodrigo Vivi <rodrigo.vivi@intel.com>]

On Fri, Aug 05, 2022 at 05:25:43PM +0200, Mauro Carvalho Chehab wrote:
> On Fri, 5 Aug 2022 10:39:44 -0400
> Rodrigo Vivi <rodrigo.vivi@intel.com> wrote:
> 
> > On Fri, Aug 05, 2022 at 10:46:57AM +0200, Mauro Carvalho Chehab wrote:
> > > Hi Rodrigo,
> > > 
> > > On Thu, 4 Aug 2022 13:33:06 -0400
> > > Rodrigo Vivi <rodrigo.vivi@intel.com> wrote:
> > >   
> > > > Hi Dave and Daniel,
> > > > 
> > > > Here goes drm-intel-next-fixes-2022-08-04:
> > > > 
> > > > - disable pci resize on 32-bit systems (Nirmoy)
> > > > - don't leak the ccs state (Matt)
> > > > - TLB invalidation fixes (Chris)
> > > > 
> > > > Thanks,
> > > > Rodrigo.
> > > > 
> > > > The following changes since commit 2bc7ea71a73747a77e7f83bc085b0d2393235410:
> > > > 
> > > >   Merge tag 'topic/nouveau-misc-2022-07-27' of git://anongit.freedesktop.org/drm/drm into drm-next (2022-07-27 11:34:07 +1000)
> > > > 
> > > > are available in the Git repository at:
> > > > 
> > > >   git://anongit.freedesktop.org/drm/drm-intel tags/drm-intel-next-fixes-2022-08-04
> > > > 
> > > > for you to fetch changes up to e57b9369e0c6f60155027e120fafd4b57e385b71:
> > > > 
> > > >   drm/i915/gt: Batch TLB invalidations (2022-08-01 09:48:06 -0400)
> > > > 
> > > > ----------------------------------------------------------------
> > > > - disable pci resize on 32-bit systems (Nirmoy)
> > > > - don't leak the ccs state (Matt)
> > > > - TLB invalidation fixes (Chris)
> > > > 
> > > > ----------------------------------------------------------------
> > > > Chris Wilson (4):
> > > >       drm/i915/gt: Ignore TLB invalidations on idle engines
> > > >       drm/i915/gt: Invalidate TLB of the OA unit at TLB invalidations
> > > >       drm/i915/gt: Skip TLB invalidations once wedged  
> > >   
> > > >       drm/i915/gt: Batch TLB invalidations  
> > > This patch actually adds a regression due to a silly mistake. 
> > > The fix is here:
> > > 
> > > 	https://patchwork.freedesktop.org/patch/496249/?series=106805&rev=4
> > >  
> > 
> > Thank you for the heads up.
> > 
> > Since that patch is not merged yet, what are your recommendations here?
> > Should I remove this from drm-intel-next-fixes now?
> 
> Those patches are OK to merge:
> 
>        drm/i915/gt: Ignore TLB invalidations on idle engines
>        drm/i915/gt: Invalidate TLB of the OA unit at TLB invalidations
>        drm/i915/gt: Skip TLB invalidations once wedged  
> 
> And helps reduce performance regressions due to TLB cache invalidation.
> So, I would keep them.
> 
> With regards to the 4th patch, please don't merge:
> 
> 	drm/i915/gt: Batch TLB invalidations
> 
> or merge it together with:
> 
> 	drm/i915: pass a pointer for tlb seqno at vma_invalidate_tlb()

Since this one has not landed yet I won't rush it in. I'm removing the
Batch TLB invalidations and sending a new pull request without that
patch.

Thank you for the info and great work there.

> 
> > Which regression is worst?
> 
> The regression caused by the batch patch is a lot worse, as it
> effectively disables TLB cache invalidation. The fix was meant to
> be merged at the patch, but somehow between several rebases and 3
> machines involved on tests, internal development and upstream, the
> branch actually sent upstream was not the right one. It got only
> noticed after reviewing a newer patch. Yeah, that sucks.
> 
> > Also dim has trouble with fixes for fixes in the same round.
> > Please ping me when you get that patch merged so I can pull that.
> 
> OK.
> 
> Regards,
> Mauro