On Mon, Aug 21, 2017 at 06:56:01AM -0700, Andy Lutomirski wrote: > > > > On Aug 21, 2017, at 3:33 AM, Peter Zijlstra <pet...@infradead.org> wrote:
> >> > >> Using a kernel thread solves the problem for real. Anything that > >> blindly accesses user memory in kernel thread context is terminally > >> broken no matter what. > > > > So perf-callchain doesn't do it 'blindly', it wants either: > > > > - user_mode(regs) true, or > > - task_pt_regs() set. > > > > However I'm thinking that if the kernel thread has ->mm == &efi_mm, the > > EFI code running could very well have user_mode(regs) being true. > > > > intel_pmu_pebs_fixup() OTOH 'blindly' assumes that the LBR addresses are > > accessible. It bails on error though. So while its careful, it does > > attempt to access the 'user' mapping directly. Which should also trigger > > with the EFI code. > > > > And I'm not seeing anything particularly broken with either. The PEBS > > fixup relies on the CPU having just executed the code, and if it could > > fetch and execute the code, why shouldn't it be able to fetch and read? > > There are two ways this could be a problem. One is that u privileged > user apps shouldn't be able to read from EFI memory. Ah, but only root can create per-cpu events or attach events to kernel threads (with sensible paranoia levels). > The other is that, if EFI were to have IO memory mapped at a "user" > address, perf could end up reading it. Ah, but in neither mode does perf assume much, the LBR follows branches the CPU took and thus we _know_ there was code there, not MMIO. And the stack unwind simply follows the stack up, although I suppose it could be 'tricked' into probing MMIO. We can certainly add an "->mm != ->active_mm" escape clause to the unwind code. Although I don't see how we're currently avoiding the same problem with existing userspace unwinds, userspace can equally have MMIO mapped. But neither will use pre-existing user addresses in the efi_mm I think.
