On Tue, Jun 23, 2020 at 05:23:26PM +0200, Peter Zijlstra wrote:

On Tue, Jun 23, 2020 at 04:59:14PM +0200, Joerg Roedel wrote:

quoted

On Tue, Jun 23, 2020 at 04:53:44PM +0200, Peter Zijlstra wrote:

quoted

+noinstr void idtentry_validate_ist(struct pt_regs *regs)
+{
+	if ((regs->sp & ~(EXCEPTION_STKSZ-1)) ==
+	    (_RET_IP_ & ~(EXCEPTION_STKSZ-1)))
+		die("IST stack recursion", regs, 0);
+}

Yes, this is a start, it doesn't cover the case where the NMI stack is
in-between, so I think you need to walk down regs->sp too.

That shouldn't be possible with the current code, I think.

Not with the current code, but possibly with SNP #VC exceptions:

      ->  First #VC
	  -> NMI before VC handler switched off its IST stack
	     (now on NMI IST stack)
	      -> Second SNP #VC exception before the NMI handler did the
		 #VC stack check (because HV messed around with some pages
		 touched there).

In the second #VC you use the same IST stack as in the first #VC, but
the the NMI-stack in-between.

Reliability of that depends on the unwinder, I wouldn't want the guess
uwinder to OOPS me by accident.

It doesn't use the full unwinder, it just assumes that there is a
pt_regs struct at the top of every kernel stack and walks through them
until SP points to a user-space stack.

As long as the assumption that there is a pt_regs struct on top of every
stack holds, this should be safe. The assumption might be wrong when an
exception happens during SYSCALL/SYSENTER entry, when the return frame
is not written by hardware.


	Joerg