On 05-08-2025 15:40, Danilo Krummrich wrote:
On Tue Aug 5, 2025 at 7:24 AM CEST, Himal Prasad Ghimiray wrote:
On 05-08-2025 09:26, Matthew Brost wrote:
Also I believe Danilo's suggestion here was to define drm_gpuvm_map_req
as the argument and then embed drm_gpuva_op_map within
drm_gpuvm_map_req. So in patch [1], flags would be added to
drm_gpuvm_map_req rather than drm_gpuva_op_map.
Matt
[1] https://patchwork.freedesktop.org/patch/666211/?series=149550&rev=5
Hi Matt,
Thanks for the review. Initially, I considered using drm_gpuvm_map_req
struct instead of passing drm_gpuva_op_map directly to the gpuvm layer,
allowing it to handle split/merge decisions independently.
Generally, we should only have the flags field on struct drm_gpuva_op_map if we
need to let GPUVM pass flags for (re)map operations to drivers.
However, the upcoming patch [1] relies on this flag to determine
driver-side behavior. So at the end drm_gpuva_op_map and
drm_gpuvm_map_req might end up identical. Based on that—and Danilo’s
feedback on this patch [2] I thought It will be better to keep a single
op_map struct with the flag included.
Let's leave this to the upcoming patches, we can always adjust. For now, let's
go with what Matt summarized above please.
Sure. Thanks. will update next version to use drm_gpuvm_map_req
Boris, could you please confirm if the flag will be useful on the driver
side [1]?
[1] https://patchwork.freedesktop.org/patch/662832/?series=151264&rev=2
[2] https://patchwork.freedesktop.org/patch/662819/?series=151264&rev=2
