On Mon, Oct 07, 2019 at 04:08:41PM +0200, Dmitry Vyukov wrote:
> On Mon, Oct 7, 2019 at 3:55 PM Christian Brauner
> <christian.brauner@xxxxxxxxxx> wrote:
> >
> > On Mon, Oct 07, 2019 at 03:50:47PM +0200, Dmitry Vyukov wrote:
> > > On Mon, Oct 7, 2019 at 3:18 PM Andrea Parri <parri.andrea@xxxxxxxxx> wrote:
> > > >
> > > > On Mon, Oct 07, 2019 at 01:01:17PM +0200, Christian Brauner wrote:
> > > > > When assiging and testing taskstats in taskstats_exit() there's a race
> > > > > when writing and reading sig->stats when a thread-group with more than
> > > > > one thread exits:
> > > > >
> > > > > cpu0:
> > > > > thread catches fatal signal and whole thread-group gets taken down
> > > > > do_exit()
> > > > > do_group_exit()
> > > > > taskstats_exit()
> > > > > taskstats_tgid_alloc()
> > > > > The tasks reads sig->stats holding sighand lock seeing garbage.
> > > >
> > > > You meant "without holding sighand lock" here, right?
> > > >
> > > >
> > > > >
> > > > > cpu1:
> > > > > task calls exit_group()
> > > > > do_exit()
> > > > > do_group_exit()
> > > > > taskstats_exit()
> > > > > taskstats_tgid_alloc()
> > > > > The task takes sighand lock and assigns new stats to sig->stats.
> > > > >
> > > > > Fix this by using READ_ONCE() and smp_store_release().
> > > > >
> > > > > Reported-by: syzbot+c5d03165a1bd1dead0c1@xxxxxxxxxxxxxxxxxxxxxxxxx
> > > > > Fixes: 34ec12349c8a ("taskstats: cleanup ->signal->stats allocation")
> > > > > Cc: stable@xxxxxxxxxxxxxxx
> > > > > Signed-off-by: Christian Brauner <christian.brauner@xxxxxxxxxx>
> > > > > Reviewed-by: Dmitry Vyukov <dvyukov@xxxxxxxxxx>
> > > > > Link: https://lore.kernel.org/r/20191006235216.7483-1-christian.brauner@xxxxxxxxxx
> > > > > ---
> > > > > /* v1 */
> > > > > Link: https://lore.kernel.org/r/20191005112806.13960-1-christian.brauner@xxxxxxxxxx
> > > > >
> > > > > /* v2 */
> > > > > - Dmitry Vyukov <dvyukov@xxxxxxxxxx>, Marco Elver <elver@xxxxxxxxxx>:
> > > > > - fix the original double-checked locking using memory barriers
> > > > >
> > > > > /* v3 */
> > > > > - Andrea Parri <parri.andrea@xxxxxxxxx>:
> > > > > - document memory barriers to make checkpatch happy
> > > > > ---
> > > > > kernel/taskstats.c | 21 ++++++++++++---------
> > > > > 1 file changed, 12 insertions(+), 9 deletions(-)
> > > > >
> > > > > diff --git a/kernel/taskstats.c b/kernel/taskstats.c
> > > > > index 13a0f2e6ebc2..978d7931fb65 100644
> > > > > --- a/kernel/taskstats.c
> > > > > +++ b/kernel/taskstats.c
> > > > > @@ -554,24 +554,27 @@ static int taskstats_user_cmd(struct sk_buff *skb, struct genl_info *info)
> > > > > static struct taskstats *taskstats_tgid_alloc(struct task_struct *tsk)
> > > > > {
> > > > > struct signal_struct *sig = tsk->signal;
> > > > > - struct taskstats *stats;
> > > > > + struct taskstats *stats_new, *stats;
> > > > >
> > > > > - if (sig->stats || thread_group_empty(tsk))
> > > > > - goto ret;
> > > > > + /* Pairs with smp_store_release() below. */
> > > > > + stats = READ_ONCE(sig->stats);
> > > >
> > > > This pairing suggests that the READ_ONCE() is heading an address
> > > > dependency, but I fail to identify it: what is the target memory
> > > > access of such a (putative) dependency?
> > >
> > > I would assume callers of this function access *stats. So the
> > > dependency is between loading stats and accessing *stats.
> >
> > Right, but why READ_ONCE() and not smp_load_acquire here?
>
> Because if all memory accesses we need to order have data dependency
> between them, READ_ONCE is enough and is cheaper on some archs (e.g.
> ARM).
> In our case there is a data dependency between loading of stats and
> accessing *stats (only Alpha could reorder that, other arches can't
> load via a pointer before loading the pointer itself (sic!)).
Right, the except-Alpha-clause is well-known...
Christian
