On Fri Aug 22, 2025 at 11:28 AM CEST, Alice Ryhl wrote:
> diff --git a/include/drm/drm_gpuvm.h b/include/drm/drm_gpuvm.h
> index
> 4a22b9d848f7b3d5710ca554f5b01556abf95985..598ba376b9430cdd4ab7bacdc15de031a887cf71
> 100644
> --- a/include/drm/drm_gpuvm.h
> +++ b/include/drm/drm_gpuvm.h
> @@ -196,10 +196,20 @@ enum drm_gpuvm_flags {
> */
> DRM_GPUVM_RESV_PROTECTED = BIT(0),
>
> + /**
> + * @DRM_GPUVM_IMMEDIATE_MODE: use the locking scheme that makes it safe
> + * to modify the GPUVM during the fence signalling path
I think this isn't entirely true yet or at least can be ambiguous for now,
because it doesn't prevent users from having DRM_GPUVM_RESV_PROTECTED set, which
requires the DMA resv lock to destroy a struct drm_gpuvm_bo, which may happen
from drm_gpuva_unlink().
So, for now, until we have the deferred drop idea in place, it also
requires DRM_GPUVM_RESV_PROTECTED to not be set.
(Eventually, we of course want both to be represented as a single flag, such
that it becomes an either or.)
I also wouldn't say "makes it safe to", but rather "makes it possible to
safely". We have no control over what the users do with the mutex. :)
NIT: missing period
> + *
> + * When set, gpuva.lock is used to protect gpuva.list in all GEM
> + * objects associated with this GPUVM. Otherwise, the GEMs dma-resv is
> + * used.
> + */
> + DRM_GPUVM_IMMEDIATE_MODE = BIT(1),
> +
> /**
> * @DRM_GPUVM_USERBITS: user defined bits
> */
> - DRM_GPUVM_USERBITS = BIT(1),
> + DRM_GPUVM_USERBITS = BIT(2),
> };
