Re: [PATCH v3 0/2] drm/panthor: Fix panthor+FEX-Emu

[Boris Brezillon <boris.brezillon@collabora.com>]

On Fri,  6 Jun 2025 10:09:30 +0200
Boris Brezillon <boris.brezillon@collabora.com> wrote:

> Hello,
> 
> This is an attempt a couple bugs exposed by FEX-Emu. The first one
> is pretty trivial and should be uncontroversial, since it's just
> a missing padding field in one of our uAPI structs. We are getting
> away with it on arm32 because of the alignment rules provided by
> the Arm ABI, but x86 has relaxed constraints for u64 fields, and
> this bug is definitely hit when running a 32-bit x86 mesa binary
> under FEX Emu.
> 
> The second fix is addressing a problem we have because FEX-Emu is
> an aarch64 process executing 32-bit x86 code, meaning the check
> we do on the is-32bit-task check we do to figure out the MMIO
> offset seen by the user won't work. In order to fix that, we add
> an ioctl to let the user explicitly set this offset. The offset
> can only be set early on, if no MMIO range has been mapped before.
> 
> With those, and the mesa MR at [1], I managed to run a 32-bit x86
> glmark2 through FEX without using the host mesa (if we were to use
> the thunked mesa lib, both the kernel and mesa would use
> MMIO_OFFSET_64BIT, and the problem doesn't exist anymore).
> 
> Regards,
> 
> Boris
> 
> Changes in v2:
> - Simplify the logic in patch2 to have a lockless solution that's
>   still safe for what we need
> 
> Changes in v3:
> - Fix a comment
> 
> [1]https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34573
> 
> Boris Brezillon (2):
>   drm/panthor: Add missing explicit padding in drm_panthor_gpu_info
>   drm/panthor: Fix the user MMIO offset logic for emulators

Queued to drm-misc-next.

> 
>  drivers/gpu/drm/panthor/panthor_device.h | 18 ++++++++
>  drivers/gpu/drm/panthor/panthor_drv.c    | 56 +++++++++++++++++-------
>  include/uapi/drm/panthor_drm.h           | 41 +++++++++++++++++
>  3 files changed, 99 insertions(+), 16 deletions(-)
>