On Wed, May 04, 2022 at 01:07:26PM +0300, Andy Shevchenko wrote: > On Wed, May 04, 2022 at 11:57:29AM +0200, Jason A. Donenfeld wrote: > > On Wed, May 04, 2022 at 11:42:27AM +0200, Jason A. Donenfeld wrote: > > > So more likely is that this patch just helps unmask a real issue > > > elsewhere -- linker, compiler, or register restoration after preemption. > > > I don't think there's anything to do with regards to the patch of this > > > thread, as it's clearly fine. > > > > The problem even goes away if I just add a nop... > > Alignment? Compiler bug? HW issue? Probably one of those, yea. Removing the instruction addresses, the only difference between the two compiles is: https://xn--4db.cc/Rrn8usaX/diff#line-440 So either there's some alignment going on here, a compiler thing I haven't spotted yet, or some very fragile interrupt/preemption behavior that's interacting with this, either on the kernel side or the QEMU side. (I've never touched real HW for this; I just got nerd sniped when wondering why my wireguard CI was failing...) Jason
