On 21-08-2025 19:05, Danilo Krummrich wrote:
On Thu Aug 21, 2025 at 3:01 PM CEST, Boris Brezillon wrote:
On Thu, 21 Aug 2025 14:55:06 +0200
"Danilo Krummrich" <d...@kernel.org> wrote:
On Thu Aug 21, 2025 at 1:25 PM CEST, Boris Brezillon wrote:
On Thu, 21 Aug 2025 13:01:46 +0200
Boris Brezillon <boris.brezil...@collabora.com> wrote:
On a second thought, I'm now wondering why we need drm_gpuvm_map_req in
the first place. It would kinda make sense if it was containing an
bool madvise;
field, so you don't have to pass it around, but even then, I'm
wondering if we wouldn't be better off adding this field to
drm_gpuva_op_map instead and passing an drm_gpuva_op_map object to
the various map helpers (like Danilo suggested in his review of the
REPEATED mode series Caterina sent).
More on that: the very reason I introduced drm_gpuvm_map_req in the
first place is so we have a clear differentiation between an overall
map request and the sub-operations that are created to fulfill it.
Looks like this was not a concern for Danilo and he was happy with us
using _op_map for this.
The other reason we might want to add drm_gpuvm_map_req is so that
information we only need while splitting a req don't pollute
drm_gpuva_op_map. Given I was going to pass the flags to the driver's
callback anyway (meaning it's needed at the op_map level), and given
you're passing madvise as a separate bool argument to various helpers
(_map_req just contains the op, not the madvise bool), I don't think
this aspect matters.
Good catch! Indeed, when Himal picked up your struct drm_gpuvm_map_req patch,
there were additional flags included in the structure. Now that it is
essentially a transparent wrapper, I prefer to use struct drm_gpuva_op_map
directly.
However, given that you still have patches in flight that will add a flags field
to struct drm_gpuvm_map_req I think it's probably fine to introduce it right
away. Or did you drop this plan of adding those flags?
I need the flags field in the op_map too (so I can propagate it to the
drm_gpuva object), so I'd rather go with an op_map object directly and
kill drm_gpuvm_map_req now.
In this case I agree, let's use struct drm_gpuva_op_map directly.
According to the kernel documentation for the drm_gpuva_op_map
structure, it is intended to represent a single map operation generated
as the output of ops_create or the GPU VA manager. Using it as a direct
input to ops_create contradicts this definition.
For drm_gpuvm_sm_map_ops_create, the values align with those in
drm_gpuvm_map_req. However, this is not the case for
drm_gpuvm_madvise_ops_create.
If we plan to proceed with deprecating drm_gpuvm_map_req, we need to
clarify the fundamental definition of drm_gpuva_op_map:
Should it represent a user-requested map, or an operation generated by
the GPU VA manager?
