10:52 <agx_> lynxeye: do i remember correctly you had a simple reproducer to https://gitlab.freedesktop.org/mesa/mesa/-/issues/3854

10:53 <agx_> i got a chance to look at that yesterday and would like to test the fix^hack i have before cleaning up

10:53 <marex> agx_: piglit maybe ?

10:53 <marex> that one runs in parallel

10:54 <austriancoder> agx_: https://freenode.irclog.whitequark.org/etnaviv/search?q=softpin

10:55 <lynxeye> agx_: Yes, I've noticed some kernel messages looking like address space collisions while running piglit with the default parallelism of one job per core.

10:56 <marex> qtwebengine probably triggers the same thing , with its parallel rendering

10:58 <lynxeye> marex: parallel rendering is not really a good trigger here. What you need is to have BOs still being active on the GPU while the userspace thinks they are already freed. Most workloads don't cycle through enough BOs with long running GPU jobs to expose the issue.

11:01 <marex> lynxeye: and you;re sure thats not what happens with qtwebengine ?

11:03 <lynxeye> marex: Most BO allocations are done from the cache, which will not expose the address space collision bug. Any workload actually triggering BO large scale BO cycling through the kernel will have their performance tank. Application developers are pretty good at keeping buffers around as they know about the performance implications. ;)

13:01 <agx_> lynxeye: thanks!