On Mon Jul 7, 2025 at 7:04 PM CEST, Caterina Shablia wrote:
> From: Asahi Lina <lina+ker...@asahilina.net>
>
> To be able to support "fake sparse" mappings without relying on GPU page
> fault handling, drivers may need to create large (e.g. 4GiB) mappings of
> the same page repeatedly (or same range of pages). Doing this through
> individual mappings would be very wasteful. This can be handled better
> by using a flag on map creation, but to do it safely, drm_gpuvm needs to
> be aware of this special case.
>
> Add a flag that signals that a given mapping is a page mapping, which is
> repeated all over the entire requested VA range. This tweaks the
> sm_map() logic to treat the GEM offsets differently when mappings are
> a repeated ones so they are not incremented as they would be with regular
> mappings.
>
> The size of the GEM portion to repeat is passed through
> drm_gpuva::gem::range. Most of the time it will be a page size, but
> it can be bigger as long as it's less that drm_gpuva::va::range, and
> drm_gpuva::gem::range is a multiple of drm_gpuva::va::range.
Should be "as long as it's less that drm_gpuva::va::range, and
drm_gpuva::va::range is a multiple of drm_gpuva::gem::range".
I also think this feature deserves its own section in the global GPUVM
documentation -- please add a corresponding paragraph.
> +static int check_map_req(struct drm_gpuvm *gpuvm,
Let's call this validate_map_request().
> + const struct drm_gpuvm_map_req *req)
> +{
> + if (unlikely(!drm_gpuvm_range_valid(gpuvm, req->va.addr,
> req->va.range)))
> + return -EINVAL;
> +
> + if (req->flags & DRM_GPUVA_REPEAT) {
> + u64 va_range = req->va.range;
> +
> + /* For a repeated mapping, GEM range must be > 0
> + * and a multiple of the VA range.
> + */
> + if (unlikely(!req->gem.range ||
> + va_range < req->gem.range ||
> + do_div(va_range, req->gem.range)))
> + return -EINVAL;
> + }
> +
> + return 0;
> +}
> +
> static int
> __drm_gpuvm_sm_map(struct drm_gpuvm *gpuvm,
> const struct drm_gpuvm_ops *ops, void *priv,
> @@ -2137,6 +2179,7 @@ __drm_gpuvm_sm_map(struct drm_gpuvm *gpuvm,
> struct drm_gpuva reqva = {
> .va.addr = req->va.addr,
> .va.range = req->va.range,
> + .gem.range = req->gem.range,
> .gem.offset = req->gem.offset,
> .gem.obj = req->gem.obj,
> .flags = req->flags,
> @@ -2144,7 +2187,8 @@ __drm_gpuvm_sm_map(struct drm_gpuvm *gpuvm,
> u64 req_end = req->va.addr + req->va.range;
> int ret;
>
> - if (unlikely(!drm_gpuvm_range_valid(gpuvm, req->va.addr,
> req->va.range)))
> + ret = check_map_req(gpuvm, req);
> + if (unlikely(ret))
> return -EINVAL;
>
> drm_gpuvm_for_each_va_range_safe(va, next, gpuvm, req->va.addr,
> req_end) {
> @@ -2175,7 +2219,8 @@ __drm_gpuvm_sm_map(struct drm_gpuvm *gpuvm,
> .va.addr = req_end,
> .va.range = range - req->va.range,
> .gem.obj = obj,
> - .gem.offset = offset + req->va.range,
> + .gem.range = va->gem.range,
> + .gem.offset = offset,
Why change this from offset + req->va.range to just offset?
Same for similar other changes below.
Also it seems that we need to update the documentation which shows all potential
cases when calling __drm_gpuvm_sm_map() [1].
[1] https://docs.kernel.org/gpu/drm-mm.html#split-and-merge
> .flags = va->flags,
> };
> struct drm_gpuva_op_unmap u = {
> @@ -2183,6 +2228,9 @@ __drm_gpuvm_sm_map(struct drm_gpuvm *gpuvm,
> .keep = merge,
> };
>
> + if (!(va->flags & DRM_GPUVA_REPEAT))
> + n.gem.offset += req->va.range;
> +
> ret = op_remap_cb(ops, priv, NULL, &n, &u);
> if (ret)
> return ret;
> @@ -2194,6 +2242,7 @@ __drm_gpuvm_sm_map(struct drm_gpuvm *gpuvm,
> .va.addr = addr,
> .va.range = ls_range,
> .gem.obj = obj,
> + .gem.range = va->gem.range,
> .gem.offset = offset,
> .flags = va->flags,
> };
> @@ -2220,11 +2269,14 @@ __drm_gpuvm_sm_map(struct drm_gpuvm *gpuvm,
> .va.addr = req_end,
> .va.range = end - req_end,
> .gem.obj = obj,
> - .gem.offset = offset + ls_range +
> - req->va.range,
> + .gem.range = va->gem.range,
> + .gem.offset = offset,
> .flags = va->flags,
> };
>
> + if (!(va->flags & DRM_GPUVA_REPEAT))
> + n.gem.offset += ls_range +
> req->va.range;
> +
> ret = op_remap_cb(ops, priv, &p, &n, &u);
> if (ret)
> return ret;
> @@ -2250,7 +2302,8 @@ __drm_gpuvm_sm_map(struct drm_gpuvm *gpuvm,
> .va.addr = req_end,
> .va.range = end - req_end,
> .gem.obj = obj,
> - .gem.offset = offset + req_end - addr,
> + .gem.range = va->gem.range,
> + .gem.offset = offset,
> .flags = va->flags,
> };
> struct drm_gpuva_op_unmap u = {
> @@ -2258,6 +2311,9 @@ __drm_gpuvm_sm_map(struct drm_gpuvm *gpuvm,
> .keep = merge,
> };
>
> + if (!(va->flags & DRM_GPUVA_REPEAT))
> + n.gem.offset += req_end - addr;
> +
> ret = op_remap_cb(ops, priv, NULL, &n, &u);
> if (ret)
> return ret;
> @@ -2294,6 +2350,7 @@ __drm_gpuvm_sm_unmap(struct drm_gpuvm *gpuvm,
> prev.va.addr = addr;
> prev.va.range = req_addr - addr;
> prev.gem.obj = obj;
> + prev.gem.range = va->gem.range;
> prev.gem.offset = offset;
> prev.flags = va->flags;
>
> @@ -2304,7 +2361,10 @@ __drm_gpuvm_sm_unmap(struct drm_gpuvm *gpuvm,
> next.va.addr = req_end;
> next.va.range = end - req_end;
> next.gem.obj = obj;
> - next.gem.offset = offset + (req_end - addr);
> + prev.gem.range = va->gem.range;
> + next.gem.offset = offset;
> + if (!(va->flags & DRM_GPUVA_REPEAT))
> + next.gem.offset += req_end - addr;
> next.flags = va->flags;
>
> next_split = true;
> diff --git a/include/drm/drm_gpuvm.h b/include/drm/drm_gpuvm.h
> index f77a89e791f1..629e8508f99f 100644
> --- a/include/drm/drm_gpuvm.h
> +++ b/include/drm/drm_gpuvm.h
> @@ -56,10 +56,19 @@ enum drm_gpuva_flags {
> */
> DRM_GPUVA_SPARSE = (1 << 1),
>
> + /**
> + * @DRM_GPUVA_REPEAT:
> + *
> + * Flag indicating that the &drm_gpuva is a mapping of a GEM
> + * portion repeated multiple times to fill the virtual address
"of a GEM object with a certain range that is repeated multiple times to ..."
