On Wed, 2022-01-05 at 19:51 +0800, Xi Ruoyao wrote: > On Wed, 2022-01-05 at 17:40 +0800, Huacai Chen wrote: > > Hi, Ruoyao, > > > > The problem still exists in 5.16-rc8, can you try to change > > cpu_relax() definition to smp_mb()? It seems can fix the problem. > > Is there any workload which can triggers the panic? I can't trigger it > by building and testing GCC, or building the kernel anymore. > > And is your "stable" issue the same one I'd encountered? To me changing > barrier() to smp_mb() may fix some deadlock, but not a panic. (I'm not > an expert on CPU architecture or kernel programming, so maybe I'm wrong > here.) > > I'll put my 3A5000 machine into a loop building kernel and see if I can > trigger the panic again... I can't reproduce the issue on 5.16-rc8. But I can reproduce it on 5.16-rc5 and the s/barrier/smp_mb/ change fixes the issue. I'm still puzzled: if there some workload which can reproduce the issue more deterministic? -- Xi Ruoyao <xry111@xxxxxxxxxxxxxxxx> School of Aerospace Science and Technology, Xidian University
