[Intel-gfx] Merging TTM branches through the Intel tree?

[Daniel Vetter]

On Fri, Jun 04, 2021 at 11:01:40AM +0200, Thomas Hellström wrote:
> 
> On 6/4/21 9:51 AM, Christian König wrote:
> > Am 03.06.21 um 09:36 schrieb Daniel Vetter:
> > > On Thu, Jun 3, 2021 at 8:50 AM Thomas Hellström
> > > <thomas.hellstrom at linux.intel.com> wrote:
> > > > 
> > > > On 6/2/21 8:40 PM, Daniel Vetter wrote:
> > > > > On Wed, Jun 02, 2021 at 11:48:41AM +0200, Christian König wrote:
> > > > > > Am 02.06.21 um 11:16 schrieb Thomas Hellström (Intel):
> > > > > > > On 6/2/21 10:32 AM, Christian König wrote:
> > > > > > > > Uff I'm just waiting for feedback from Philip to
> > > > > > > > merge a large patch
> > > > > > > > set for TTM through drm-misc-next.
> > > > > > > > 
> > > > > > > > I'm pretty sure we will run into merge conflicts if you try to push
> > > > > > > > your changes through the Intel tree.
> > > > > > > > 
> > > > > > > > Christian.
> > > > > > > OK, so what would be the best approach here?, Adding
> > > > > > > the TTM patches to
> > > > > > > drm-misc-next when your set has landed?
> > > > > > I think I will send out out my set to Matthew once more
> > > > > > for review, then
> > > > > > push the common TTM stuff to drm-misc-next as much as possible.
> > > > > > 
> > > > > > Then you should be able to land your stuff to
> > > > > > drm-misc-next and rebase on
> > > > > > the end result.
> > > > > > 
> > > > > > Just need to note to David that drm-misc-next should be
> > > > > > merged to drm-next
> > > > > > before the Intel patches depending on that stuff land as well.
> > > > > Other option (because the backmerges tend to be slow) is a
> > > > > topic branch,
> > > > > and we just eat/resolve the conflicts in both drm-misc-next and
> > > > > drm-intel-gt-next in the merge commit. If it's not too bad (I haven't
> > > > > looked at what exactly we need for the i915 side from ttm in detail).
> > > > > 
> > > > > But also often figuring out the topic branch logistics takes
> > > > > longer than
> > > > > just merging to drm-misc-next as the patches get ready.
> > > > > -Daniel
> > > > Daniel: So the thing we need to get into TTM is the iterator-based
> > > > move_memcpy which is more adaptable than the current one and needed to
> > > > support non-linear lmem buffers, some bug-fixes and minor changes to be
> > > > able to keep our short-term-pinning while on the LRU. A necessary evil.
> > > > 
> > > > Christian: it looks like you have landed some TTM changes already, in
> > > > particular the &bo->mem -> bo->resource change which is the main
> > > > conflict I think.
> > 
> > Yes, I thought that pushing this with Matthew rb should solve at least a
> > bit of the conflict.
> > 
> > > > Is the 10 patches self-allocation series the main
> > > > remaining part?
> > 
> > Yes, exactly. I only need Matthew's, Daniel's or your ok and I'm good to
> > go as well
> > 
> > > > That will probably cause some conflicts with already
> > > > pushed i915 TTM setup code, but otherwise will not conflict with the
> > > > rest of the TTM code I think, which should make it possible to bring in
> > > > our TTM changes after conflict resolution with what you've already
> > > > pushed. The memcpy code is pretty self-contained.
> > > I think in that case topic branch on top of drm-next (once the ttm
> > > bits we conflict with are there) is probably best, and then pull that
> > > into drm-misc-next and drm-intel-gt-next. Merge window freeze is also
> > > approach, so without topic branch we'd be stuck until like -rc2 when
> > > drm-next reopens. I guess Maarten can do the topic branch logistics in
> > > drm-misc.git for this.
> > 
> > That approach sounds good to me as well.
> > 
> > The amdgpu branch had some merge conflicts as well, but nothing we
> > couldn't fix.
> 
> OK, so this is going to be a little tricky, I guess.
> 
> From what I can tell, the memcpy TTM stuff is resolved locally and can be
> merged to drm-misc-next immediately. It might have a very minor conflict
> with your 10 patches I think, if any.
> 
> Your 10 patches will conflict slightly with current drm-intel-gt-next I
> think.
> 
> Remaining intel patches will conflict only with current drm-misc-next.
> 
> So We could have pull order
> 
> - drm-misc-next up to bot not including your 10 patches,
> - drm-intel-gt-next
> - drm-misc-next from your 10 paches and onwards,
> - Intel's ttm enablement topic branch.

If it's just slight conflicts then I wouldn't bother with careful merge
order. Because if we do this we can get around to the i915 ttm topic
branch only when we're back to -rc2.

We can also validate any conflicts in drm-tip easily before they get baked
in in drm-next.

So I'd just go with
- drm-misc-next gets those 10 patches from Christian and the memcpy prep
  stuff from you, gets send to drm-next (that's probably the last feature
  pull for 5.14 anyway, maybe another one)
- drm-intel-gt-next gets send to drm-next
- topic branch with remaining i915 ttm work that's in flight on top of
  drm-next and we pull that into drm-misc-next and drm-intel-gt-next as
  needed

Only thing we need for this is a few days of testing to make sure any
conflicts between -misc-next and -gt-next are fully validated.

Adding Dave for that so he knows too.

> Whether I push the ttm memcpy stuff before your 10 patches or after
> shouldn't really matter except it might take some time to resolve the 10
> patches - drm-intel-gt-next conflict in drm-tip.
> 
> So OK to merge the memcpy stuff to drm-misc-next now or do you want me to
> hold on?
> 
> I'll take a look at what's remaining to review in your series. I guess it's
> in our interest that both these series get merged asap.

Yeah that part I think makes sense.
-Daniel

> 
> /Thomas
> 
> 
> 
> > 
> > Christian.
> > 
> > > -Daniel
> > 

-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch