> On Apr 28, 2019, at 12:43 PM, Steven Rostedt <rostedt@xxxxxxxxxxx> wrote: > > On Sun, 28 Apr 2019 11:08:34 -0700 > Andy Lutomirski <luto@xxxxxxxxxxxxxx> wrote: > >>> >>> Perhaps adding another slot into pt_regs that gets used by int3 to >>> store a slot to emulate a call on return? >>> >>> >> >> That’s not totally nuts, although finding pt_regs isn’t entirely trivial. > > I meant on the int3 handler (which stores the pt_regs). But that’s below the stub’s RSP, so it’s toast if another interrupt happens. Or am I misunderstanding you? > >> >> I still think I prefer an approach where we just emulate the call directly. > > Then, on the return of int3, if there's anything in that slot, then we > could possibly shift the exception handler frame (that was added by the > hardware), insert the slot data into the top of the stack, and then > call iret (which the int3 handler, would add the return ip to be the > function being called), which would in essence emulate the call directly. Oh, I get it. I liked Josh’s old proposal of unconditionally shifting the #BP frame 8 bytes better. It will be interesting when kernel shadow stacks are thrown in the mix, but that’s a problem for another day.
