Sorry, I accidentally sent this reply offlist! Resending, my apologies... On 07/02/2023 20.29, Thomas Zimmermann wrote: > Hi > > Am 05.02.23 um 13:51 schrieb Asahi Lina: >> Other functions touching shmem->sgt take the pages lock, so do that here > > Really? I was just locking at the Lima driver and it apparently access > sgt without locking in [1]. Not that I disagree with the patch in general. It looks like that lima code is reimplementing a lot of helper functionality. I imagine it was written before the helpers...? I think most of that function could be replaced with a call to drm_gem_shmem_get_pages_sgt(). I don't know exactly how the lima driver works, but if there is a possibility of multiple calls to lima_heap_alloc() on the same BO without a higher-level mutex protecting it, I would say that code is racy. For the Rust abstraction (and really for a well-designed API in general) you want a coherent story on locking, so I think it makes sense to take the pages lock to manipulate the sgt, since drm_gem_shmem_get_pages_sgt() was already taking the pages lock for inner things anyway. Otherwise it's impossible to make safe without adding another discrete layer of locking around everything (I can't just take the pages lock in the wrapper since drm_gem_shmem_get_pages_sgt() would try to recursively lock it). > Best regards > Thomas > > [1] > https://elixir.bootlin.com/linux/latest/source/drivers/gpu/drm/lima/lima_gem.c#L21 ~~ Lina
