On Thu, Aug 03, 2023 at 08:06:20PM +0200, Paolo Bonzini wrote: > On 8/3/23 14:06, Peter Zijlstra wrote: > > > > By marking them with STACK_FRAME_NON_STANDARD you will get no ORC data > > at all, and then you also violate the normal framepointer calling > > convention. > > > > This means that if you need to unwind here you're up a creek without no > > paddles on. > > The only weird thing that can happen is ud2 instructions that are executed > in case the vmload/vmrun/vmsave instructions causes a #GP, from the > exception handler. This code is ran with GIF disabled, so NMIs are not in the books, right? Does GIF block #MC ? > If I understand correctly those ud2 would use ORC information to show the > backtrace, but even then the frame pointer should be correct. Of these > instructions, vmrun is the only one that runs with wrong %rbp; and it is > unlikely or even impossible that a #GP happens at vmrun, because the same > operand has been used for a vmload ten instructions before. The only time I > saw that #GP it was due to a processor errata, but it happened consistently > on the vmload. > > So if frame pointer unwinding can be used in the absence of ORC, Nikunj > patch should not break anything. But framepointer unwinds rely on BP, and that is clobbered per the objtool complaint. Also, if you look at the makefile hunk that's being replaced, that was conditional on CONFIG_FRAMEPOINTS, while the annotation that's being added is not. I think objtool won't complain for ORC builds, only for framepoints builds.
