Re: [RFC PATCH V2 4/7] x86/hw_breakpoint: Prevent data breakpoints on user_pcid_flush_mask

[Lai Jiangshan]

On Tue, May 26, 2020 at 12:39 PM Andy Lutomirski <luto@xxxxxxxxxx> wrote:
>
> On Mon, May 25, 2020 at 9:31 PM Lai Jiangshan
> <jiangshanlai+lkml@xxxxxxxxx> wrote:
> >
> > On Tue, May 26, 2020 at 12:21 PM Andy Lutomirski <luto@xxxxxxxxxx> wrote:
> > >
> > > On Mon, May 25, 2020 at 6:42 PM Lai Jiangshan <laijs@xxxxxxxxxxxxxxxxx> wrote:
> > > >
> > > > The percpu user_pcid_flush_mask is used for CPU entry
> > > > If a data breakpoint on it, it will cause an unwanted #DB.
> > > > Protect the full cpu_tlbstate structure to be sure.
> > > >
> > > > There are some other percpu data used in CPU entry, but they are
> > > > either in already-protected cpu_tss_rw or are safe to trigger #DB
> > > > (espfix_waddr, espfix_stack).
> > >
> > > How hard would it be to rework this to have DECLARE_PERCPU_NODEBUG()
> > > and DEFINE_PERCPU_NODEBUG() or similar?
> >
> >
> > I don't know, but it is an excellent idea. Although the patchset
> > protects only 2 or 3 portions of percpu data, but there is many
> > percpu data used in tracing or kprobe code. They are needed to be
> > protected too.
> >
> > Adds CC:
> > Steven Rostedt <rostedt@xxxxxxxxxxx>
> > Masami Hiramatsu <mhiramat@xxxxxxxxxx>
>
> PeterZ is moving things in the direction of more aggressively
> disabling hardware breakpoints in the nasty paths where we won't
> survive a hardware breakpoint.  Does the tracing code have portions
> that won't survive a limited amount of recursion?

Agree, after "aggressively disabling hardware breakpoints in the nasty
paths", only percpu data used by entry code needs to be protected,
even non-instrumentable percpu data used by nmi_enter() doesn't need
to be marked protected, because #DB is disabled.

Only percpu data used by entry code in ranges that #DB is not disabled
needs to be protected, there are only a small number of portions,
I don't think we need DECLARE_PERCPU_NODEBUG() or so for merely 2 or 3
portions of data. This patchset is sufficient.
(espfix_waddr, espfix_stack are not counted into, which needs more
review besides me)

>
> I'm hoping that we can keep the number of no-breakpoint-here percpu
> variables low.  Maybe we could recruit objtool to help make sure we
> got all of them, but that would be a much larger project.
>
> Would we currently survive a breakpoint on the thread stack?  I don't
> see any extremely obvious reason that we wouldn't.  Blocking such a
> breakpoint would be annoying.