[igt-dev] [PATCH i-g-t 1/7] lib/gem_submission: Add kernel exec object alignment capability

[Dixit, Ashutosh]

On Thu, 14 Oct 2021 00:14:46 -0700, Zbigniew Kempczyński wrote:
>
> On Wed, Oct 13, 2021 at 07:26:28PM -0700, Dixit, Ashutosh wrote:
> > On Tue, 12 Oct 2021 22:27:45 -0700, Zbigniew Kempczyński wrote:
> > >
> > > On Tue, Oct 12, 2021 at 09:47:18PM -0700, Dixit, Ashutosh wrote:
> > > > On Thu, 07 Oct 2021 23:54:26 -0700, Zbigniew Kempczyński wrote:
> > > > >
> > > > > +bool gem_allows_passing_alignment(int fd)
> > > > > +{
> > > > > +	struct drm_i915_gem_exec_object2 obj = {
> > > > > +		.handle = gem_create(fd, 4096),
> > > > > +	};
> > > > > +	struct drm_i915_gem_execbuffer2 execbuf = {
> > > > > +		.buffers_ptr = to_user_pointer(&obj),
> > > > > +		.buffer_count = 1,
> > > > > +	};
> > > > > +	bool ret;
> > > > > +	const uint32_t bbe = MI_BATCH_BUFFER_END;
> > > > > +
> > > > > +	gem_write(fd, obj.handle, 0, &bbe, sizeof(bbe));
> > > > > +	gem_execbuf(fd, &execbuf);
> > > > > +
> > > > > +	obj.alignment = 0x2000;
> > > >
> > > > 8K? Shouldn't this be less than 4K?
> > >
> > > I'm not sure how kernel check will look like so positive value larger
> > > than 4K is sane for me. For example if in the kernel we will have:
> > >
> > > if (entry->alignment != 0 && entry->alignment != 0x1000)
> > >	return -EINVAL;
> >
> > Sorry I am still not following why the alignment has to be > 4K. Since for
> > alignment we generally consider aligned to a 4K page, so an alignment value
> > which is multiple of 4K might not even be a correct test. I would just use
> > an alignment value of 64 here (kernel requires it to be a power of
> > 2). Wouldn't that work? Thanks.
>
> Agree, on input any value which is power of two is correct. I just choosen
> 8K because it is a) page aligned b) power of two so definitely this value
> can be used within kernel.
>
> alignment = 64 is a little bit tricky - it is correct but will be changed
> to minimum alignment before pinning so I would like to avoid such values.
> I've spent some time to go over the pinning path according what has happen
> with that 64 and I was surprised it is not used at all anywhere.
>
> Resume: any value which is power of two is correct here, I just took 8K.

Yes I saw gem_exec_alignment later and it is using similar alignment
values. So I need to look some more to understand what this alignment is.

So please change the name to either gem_allows_obj_alignment or
gem_has_obj_alignment whichever you prefer. No need to resend the series,
please go ahead and merge the R-b patches. This one is also:

Reviewed-by: Ashutosh Dixit <ashutosh.dixit at intel.com>