[Intel-gfx] [PATCH] i915: Drop relocation support on all new hardware (v3)

[Daniel Vetter]

On Thu, Mar 11, 2021 at 12:57:25PM -0600, Jason Ekstrand wrote:
> On Thu, Mar 11, 2021 at 12:20 PM Zbigniew Kempczyński
> <zbigniew.kempczynski at intel.com> wrote:
> >
> > On Thu, Mar 11, 2021 at 11:18:11AM -0600, Jason Ekstrand wrote:
> > > On Thu, Mar 11, 2021 at 10:51 AM Zbigniew Kempczyński
> > > <zbigniew.kempczynski at intel.com> wrote:
> > > >
> > > > On Thu, Mar 11, 2021 at 10:24:38AM -0600, Jason Ekstrand wrote:
> > > > > On Thu, Mar 11, 2021 at 9:57 AM Daniel Vetter <daniel at ffwll.ch> wrote:
> > > > > >
> > > > > > On Thu, Mar 11, 2021 at 4:50 PM Jason Ekstrand <jason at jlekstrand.net> wrote:
> > > > > > >
> > > > > > > On Thu, Mar 11, 2021 at 5:44 AM Zbigniew Kempczyński
> > > > > > > <zbigniew.kempczynski at intel.com> wrote:
> > > > > > > >
> > > > > > > > On Wed, Mar 10, 2021 at 03:50:07PM -0600, Jason Ekstrand wrote:
> > > > > > > > > The Vulkan driver in Mesa for Intel hardware never uses relocations if
> > > > > > > > > it's running on a version of i915 that supports at least softpin which
> > > > > > > > > all versions of i915 supporting Gen12 do.  On the OpenGL side, Gen12+ is
> > > > > > > > > only supported by iris which never uses relocations.  The older i965
> > > > > > > > > driver in Mesa does use relocations but it only supports Intel hardware
> > > > > > > > > through Gen11 and has been deprecated for all hardware Gen9+.  The
> > > > > > > > > compute driver also never uses relocations.  This only leaves the media
> > > > > > > > > driver which is supposed to be switching to softpin going forward.
> > > > > > > > > Making softpin a requirement for all future hardware seems reasonable.
> > > > > > > > >
> > > > > > > > > Rejecting relocations starting with Gen12 has the benefit that we don't
> > > > > > > > > have to bother supporting it on platforms with local memory.  Given how
> > > > > > > > > much CPU touching of memory is required for relocations, not having to
> > > > > > > > > do so on platforms where not all memory is directly CPU-accessible
> > > > > > > > > carries significant advantages.
> > > > > > > > >
> > > > > > > > > v2 (Jason Ekstrand):
> > > > > > > > >  - Allow TGL-LP platforms as they've already shipped
> > > > > > > > >
> > > > > > > > > v3 (Jason Ekstrand):
> > > > > > > > >  - WARN_ON platforms with LMEM support in case the check is wrong
> > > > > > > >
> > > > > > > > I was asked to review of this patch. It works along with expected
> > > > > > > > IGT check https://patchwork.freedesktop.org/patch/423361/?series=82954&rev=25
> > > > > > > >
> > > > > > > > Before I'll give you r-b - isn't i915_gem_execbuffer2_ioctl() better place
> > > > > > > > to do for loop just after copy_from_user() and check relocation_count?
> > > > > > > > We have an access to exec2_list there, we know the gen so we're able to say
> > > > > > > > relocations are not supported immediate, without entering i915_gem_do_execbuffer().
> > > > > > >
> > > > > > > I considered that but it adds an extra object list walk for a case
> > > > > > > which we expect to not happen.  I'm not sure how expensive the list
> > > > > > > walk would be if all we do is check the number of relocations on each
> > > > > > > object.  I guess, if it comes right after a copy_from_user, it's all
> > > > > > > hot in the cache so it shouldn't matter.  Ok.  I've convinced myself.
> > > > > > > I'll move it.
> > > > > >
> > > > > > I really wouldn't move it if it's another list walk. Execbuf has a lot
> > > > > > of fast-paths going on, and we have extensive tests to make sure it
> > > > > > unwinds correctly in all cases. It's not very intuitive, but execbuf
> > > > > > code isn't scoring very high on that.
> > > > >
> > > > > And here I'd just finished doing the typing to move it.  Good thing I
> > > > > hadn't closed vim yet and it was still in my undo buffer. :-)
> > > >
> > > > Before entering "slower" path from my perspective I would just check
> > > > batch object at that place. We still would have single list walkthrough
> > > > and quick check on the very beginning.
> > >
> > > Can you be more specific about what exactly you think we can check at
> > > the beginning?  Either we add a flag that we can O(1) check at the
> > > beginning or we have to check everything in exec2_list for
> > > exec2_list[n].relocation_count == 0.  That's a list walk.  I'm not
> > > seeing what up-front check you're thinking we can do without the list
> > > walk.
> >
> > I expect that last (default) or first (I915_EXEC_BATCH_FIRST) execobj
> > (batch) will likely has relocations. So we can check that single
> > object without entering i915_gem_do_execbuffer(). If that check
> > is missed (relocation_count = 0) you'll catch relocations in other
> > objects in check_relocations() as you already did. This is simple
> > optimization but we can avoid two iterations over buffer list
> > (first is in eb_lookup_vmas()).
> 
> Sure, we can do an early-exit check of the first and last objects.
> I'm just not seeing what that saves us given that we still have to do
> the full list-walk check anyway.  Especially since this is an error
> path which shouldn't be hit by real userspace drivers.

Yeah optimizing error checking sounds like the wrong thing to optimize.
Userspace is wrong, it might as well have to wait a bit until it gets that
rejection :-)
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch