Re: [PATCH] perf, x86: Optimize intel_pmu_pebs_fixup_ip()

[Linus Torvalds <torvalds@linux-foundation.org>]

On Thu, Oct 17, 2013 at 3:01 PM, Peter Zijlstra <peterz@infradead.org> wrote:
>
> Oh wait,.. now that Steven fixed being able to take faults from NMI
> context; we could actually try copy_from_user_inatomic(). Being able to
> directly access userspace would make the whole deal a lot easier again.

Careful! There is one magic piece of state that you need to
save-and-restore if you do this, namely %cr2. Taking a page fault
always writes to %cr2, and we must *not* corrupt it in the NMI
handler.

Also, right now, it looks like we call notify_page_fault() in the
atomic page fault case, and that would be deadly from within an NMI.

But if you move the "in_atomic()" check earlier in __do_page_fault(),
you can *try* to do something like this:

  unsigned long
  copy_from_user_nmi(void *to, const void __user *from, unsigned long n)
  {
        unsigned long cr2, flags,ret;

        if (__range_not_ok(from, n, TASK_SIZE))
                return 0;
        local_irq_save(flags);
        cr2 = read_cr2();
        ret = __copy_from_user_inatomic(to, from, n);
        /* Reading cr2 is likely much faster than writing it - but go
check this.. */
        if (cr2 != read_cr2())
                write_cr2(cr2);
        local_irq_restore(flags);
        return n - ret;
  }

or something close to that. But you absolutely *have* to save/restore
%cr2 (the above tries to avoid writing it if it didn't change,
somebody should check the timings on that to see whether it makes
sense or not).

             Linus