Re: [PATCH -v5 00/17] Rewrite x86/ftrace to use text_poke (and more)

From: Masami Hiramatsu
Date: Tue Nov 26 2019 - 03:58:20 EST


On Tue, 26 Nov 2019 09:11:04 +0900
Masami Hiramatsu <mhiramat@xxxxxxxxxx> wrote:

> On Mon, 25 Nov 2019 12:32:45 -0500
> Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:
>
> > On Mon, 25 Nov 2019 12:55:34 +0900
> > Masami Hiramatsu <mhiramat@xxxxxxxxxx> wrote:
> >
> > > void text_poke_bp_batch(struct text_poke_loc *tp, unsigned int nr_entries)
> > > ...
> > > on_each_cpu(do_sync_core, NULL, 1);
> > > /*
> > > * sync_core() implies an smp_mb() and orders this store against
> > > * the writing of the new instruction.
> > > */
> > > bp_patching.vec = NULL;
> > > bp_patching.nr_entries = 0;
> > > }
> > > -----
> > >
> > > I think the "on_each_cpu(do_sync_core, NULL, 1);" can sync the pipeline
> > > but doesn't ensure all ongoing int3 handling is done. Thus, we may need a
> >
> > How does it not ensure all ongoing int3 handling is done? int3 is done
> > with interrupts disabled, and the on_each_cpu() requires all CPUs to
> > have had their interrupts enabled, thus int3 handling should be
> > completed. Perhaps we need another sync core?
> >
> > on_each_cpu(do_sync_core, NULL, 1);
> > bp_patching.nr_entries = 0;
> > on_each_cpu(do_sync_core, NULL, 1);
> > bp_patching.vec = NULL;
>
> OK, let me check.
>
> The 1st sync_core will ensure the poking "int3" is removed. Thus any
> int3-hit address (ip) should NOT match the bp_patching.vec[*].addr after
> that. At this point, "if (likely(!bp_patching.nr_entries))" check does not
> work.
>
> And the 2nd sync_core will ensure all poke_int3_handler() will see the
> bp_patching.nr_entries = 0.
> After this point, "if (likely(!bp_patching.nr_entries))" works and
> poke_int3_handler() will exit soon. (before touching bp_patching.vec)
>

Oops, I've tested a bit older kernel (with above change it seems to be fixed).
I'll check the latest -tip.

Thank you,

--
Masami Hiramatsu <mhiramat@xxxxxxxxxx>