On Thu, Mar 19, 2020 at 9:24 AM Joerg Roedel <joro@xxxxxxxxxx> wrote: > > On Thu, Mar 19, 2020 at 08:44:03AM -0700, Andy Lutomirski wrote: > > On Thu, Mar 19, 2020 at 2:14 AM Joerg Roedel <joro@xxxxxxxxxx> wrote: > > > > > > From: Tom Lendacky <thomas.lendacky@xxxxxxx> > > > > > > Add the handler for #VC exceptions invoked at runtime. > > > > If I read this correctly, this does not use IST. If that's true, I > > don't see how this can possibly work. There at least two nasty cases > > that come to mind: > > > > 1. SYSCALL followed by NMI. The NMI IRET hack gets to #VC and we > > explode. This is fixable by getting rid of the NMI EFLAGS.TF hack. > > Not an issue in this patch-set, the confusion comes from the fact that I > left some parts of the single-step-over-iret code in the patch. But it > is not used. The NMI handling in this patch-set sends the NMI-complete > message before the IRET, when the kernel is still in a safe environment > (kernel stack, kernel cr3). Got it! > > > 2. tools/testing/selftests/x86/mov_ss_trap_64. User code does MOV > > (addr), SS; SYSCALL, where addr has a data breakpoint. We get #DB > > promoted to #VC with no stack. > > Also not an issue, as debugging is not supported at the moment in SEV-ES > guests (hardware has no way yet to save/restore the debug registers > across #VMEXITs). But this will change with future hardware. If you look > at the implementation for dr7 read/write events, you see that the dr7 > value is cached and returned, but does not make it to the hardware dr7. Eek. This would probably benefit from some ptrace / perf logic to prevent the kernel or userspace from thinking that debugging works. I guess this means that #DB only happens due to TF or INT01. I suppose this is probably okay. > > I though about using IST for the #VC handler, but the implications for > nesting #VC handlers made me decide against it. But for future hardware > that supports debugging inside SEV-ES guests it will be an issue. I'll > think about how to fix the problem, it probably has to be IST :( Or future generations could have enough hardware support for debugging that #DB doesn't need to be intercepted or can be re-injected correctly with the #DB vector. > > Regards, > > Joerg
