Re: [PATCH v2 5/7] x86/mm, tracing: Fix CR2 corruption

[Linus Torvalds]

On Fri, Jul 5, 2019 at 12:16 PM Andy Lutomirski <luto@xxxxxxxxxxxxxx> wrote:
>
> If nothing else, MOV to CR2 is architecturally serializing, so, unless thereâs some fancy unwinding involved, this will be quite slow.

That's why the NMI code does this:

        if (unlikely(this_cpu_read(nmi_cr2) != read_cr2()))
                write_cr2(this_cpu_read(nmi_cr2));

so that it normally only does a read. Only if you actually took a page
fault will it restore cr2 to the old value (and if you took a page
fault the performance issues will be _there_, not in the "restore cr2"
part)

                Linus