Re: [PATCH] uprobes: Fix handle_swbp() vs. unregister() + register() race once more

From: Ingo Molnar
Date: Fri Nov 23 2018 - 02:34:12 EST



* Oleg Nesterov <oleg@xxxxxxxxxx> wrote:

> On 11/22, Oleg Nesterov wrote:
> > On 11/22, Andrea Parri wrote:
> > >
> > > Commit 142b18ddc8143 ("uprobes: Fix handle_swbp() vs unregister() +
> > > register() race") added the UPROBE_COPY_INSN flag, and corresponding
> > > smp_wmb() and smp_rmb() memory barriers, to ensure that handle_swbp()
> > > uses fully-initialized uprobes only.
> > >
> > > However, the smp_rmb() is mis-placed: this barrier should be placed
> > > after handle_swbp() has tested for the flag, thus guaranteeing that
> > > (program-order) subsequent loads from the uprobe can see the initial
> > > stores performed by prepare_uprobe().
> > >
> > > Move the smp_rmb() accordingly. Also amend the comments associated
> > > to the two memory barriers to indicate their actual locations.
> > >
> > > Signed-off-by: Andrea Parri <andrea.parri@xxxxxxxxxxxxxxxxxxxx>
> > > Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> > > Cc: Ingo Molnar <mingo@xxxxxxxxxx>
> > > Cc: Arnaldo Carvalho de Melo <acme@xxxxxxxxxx>
> > > Cc: Alexander Shishkin <alexander.shishkin@xxxxxxxxxxxxxxx>
> > > Cc: Jiri Olsa <jolsa@xxxxxxxxxx>
> > > Cc: Namhyung Kim <namhyung@xxxxxxxxxx>
> > > Cc: Oleg Nesterov <oleg@xxxxxxxxxx>
> > > Cc: stable@xxxxxxxxxx
> > > Fixes: 142b18ddc8143 ("uprobes: Fix handle_swbp() vs unregister() + register() race")
> >
> > Thanks,
> >
> > Acked-by: Oleg Nesterov <oleg@xxxxxxxxxx>
>
> Yes, but I am not sure this is the -stable material...

So I left the Cc: stable tag intact, because this is a really low-risk
fix (it just moves barriers around), and clearly fixes a bug that people
might or might not have observed.

Even if they observed it the race is probably very hard to reproduce and
almost impossible to report - so we are better off propagating this fix
to -stable, as there's no realistic actionable way for users to actually
complain about the bug if it affects them.

That's the general backporting policy for race fixes, unless they are
really, really intrusive - which this one isn't really.

Thanks,

Ingo