On Tue, Oct 14, 2025 at 10:12 AM Alex Deucher <[email protected]> wrote:
> On Tue, Oct 14, 2025 at 2:49 AM Marek Olšák <[email protected]> wrote: > > > > On Mon, Oct 13, 2025 at 3:11 PM Alex Deucher <[email protected]> > wrote: > >> > >> On Mon, Oct 13, 2025 at 10:21 AM Liang, Prike <[email protected]> > wrote: > >> > > >> > [Public] > >> > > >> > Regards, > >> > Prike > >> > > >> > > -----Original Message----- > >> > > From: Alex Deucher <[email protected]> > >> > > Sent: Monday, October 13, 2025 9:44 PM > >> > > To: Liang, Prike <[email protected]> > >> > > Cc: Deucher, Alexander <[email protected]>; amd- > >> > > [email protected] > >> > > Subject: Re: [PATCH 3/7] drm/amdgpu/gfx: add eop size and alignment > to shadow > >> > > info > >> > > > >> > > On Mon, Oct 13, 2025 at 4:54 AM Liang, Prike <[email protected]> > wrote: > >> > > > > >> > > > [Public] > >> > > > > >> > > > We may need to update the userspace EOP buffer request; > otherwise, the EOP > >> > > buffer validation may fail. > >> > > > >> > > Existing userspace should be ok. It currently uses PAGE_SIZE which > is larger than > >> > > 2048. > >> > The mesa uses the EOP size as max_t(u32, PAGE_SIZE, > AMDGPU_GPU_PAGE_SIZE) which is sees larger than 2048, so the kernel > validates the eop buffer can be successful at this point. > >> > > >> > But the mesa may need to use the shadow_info->eop_size as well in the > future? > >> > >> Ideally mesa would query the kernel to get the proper minimum size. > >> Yogesh will be looking at this. > >> > >> Alex > > > > > > Does the EOP buffer store privileged information? What is its content? > > It stores end of pipe events for the compute queue generated from > things like RELEASE_MEM or AQL packets. They are specific to each > user queue. In theory corrupting or messing with the data in the > buffer should only affect that queue. > RELEASE_MEM has a hidden implicit VMID parameter. That's why it's important to know whether it's stored in the EOP buffer that can be overwritten by userspace. Marek
