On Tue, Nov 17, 2026 at 07:32:17PM +0000, Alice Ryhl wrote:
> Tyr configures its ARM64 LPAE page tables with 4KB and 2MB block sizes,
> but arm-lpae rejects unmapping part of a 2MB block. As a result, a
> partial unmap or overlapping remap that splits a 2MB block in
> GpuVmData::sm_step_remap() fails in __arm_lpae_unmap().
> 
> This series ports the remap fix and supporting accessors from tyr-dev
> onto drm-rust-next, along with a defensive bounds check in pt_unmap():
> 
> 1. Add the GpuVa::data_ref() accessor.
> 2. Add the IoPageTable::iova_to_phys() accessor.
> 3. Expand sub-block unmaps in sm_step_remap() to 2MB boundaries when
>    the swept head/tail fragment is physically contiguous, widen the AS
>    lock over the expanded range, rebuild the surviving fragments, and
>    take the preallocated GPUVAs before mutating page tables.
> 4. Check the remaining-length subtraction in pt_unmap() against
>    underflow.

Thank you for this series. I tested it on my Rock5B with the latest Tyr ioctl
series [1] and I can no longer trigger the unmap WARNING
"Unmap of a partial large IOPTE is not allowed". 

So, for the whole series:
Tested-by: Deborah Brouwer <[email protected]>

[1] 
https://lore.kernel.org/rust-for-linux/[email protected]/

> 
> Signed-off-by: Alice Ryhl <[email protected]>
> ---
> Sami Tolvanen (4):
>       rust: drm/gpuvm: add GpuVa::data_ref accessor
>       rust: iommu: pgtable: add iova_to_phys
>       drm/tyr: vm: expand sub-block unmaps to 2MB boundaries on remap
>       drm/tyr: vm: check the remaining length in pt_unmap()
> 
>  drivers/gpu/drm/tyr/mmu.rs               |  13 +++
>  drivers/gpu/drm/tyr/mmu/address_space.rs |  29 ++++++
>  drivers/gpu/drm/tyr/vm.rs                | 152 
> ++++++++++++++++++++++++++++---
>  rust/kernel/drm/gpuvm/va.rs              |   6 ++
>  rust/kernel/iommu/pgtable.rs             |  23 +++++
>  5 files changed, 209 insertions(+), 14 deletions(-)
> ---
> base-commit: 8045d0e7d208ad389a387f1407ce0db9070805af
> change-id: 20261116-tyr-remap-2mb-7a53b8475cfc
> 
> Best regards,
> -- 
> Alice Ryhl <[email protected]>
> 

Reply via email to