On Tue, Jun 10, 2014 at 07:36:32AM -0700, Paul E. McKenney wrote: > On Tue, Jun 10, 2014 at 03:01:38PM +0200, Peter Zijlstra wrote: > > On Tue, Jun 10, 2014 at 05:52:35AM -0700, Paul E. McKenney wrote: > > > On Tue, Jun 10, 2014 at 10:37:26AM +0200, Peter Zijlstra wrote: > > > > On Mon, Jun 09, 2014 at 09:26:13AM -0700, Paul E. McKenney wrote: > > > > > That would indeed be a bad thing, as it could potentially lead to > > > > > use-after-free bugs. Though one could argue that any code that > > > > > resulted > > > > > in use-after-free would be quite aggressive. But still... > > > > > > > > Let me hijack this thread for yet another issue... So I had an RCU > > > > related use-after-free the other day, and while Sasha was able to > > > > trigger it quite easily, I had a multi-day struggle to reproduce. > > > > > > > > Once I figured out what the exact problem was it was also clear to me > > > > why it was so hard for me to reproduce. > > > > > > > > So normally its easier to trigger races on bigger machines, more cpus, > > > > more concurrency, more races, all good. > > > > > > > > _However_ with RCU the grace period machinery is slower the bigger the > > > > machine, so bigger machine, slower grace period, slower RCU free, less > > > > likely to hit use-after-free. > > > > > > > > So I was thinking, and I know you all will go kick me for this because > > > > the very last thing we need is what I'm about to propose: more RCU > > > > flavours :-). > > > > > > > > How about an rcu_read_unlock() reference counted RCU variant that's > > > > ultra aggressive in doing the callbacks in order to better trigger such > > > > issues? > > > > > > If you are using synchronize_rcu() for the update side, then I suggest > > > rcutorture.gp_exp=1 to force use expediting throughout. > > > > No such luck, this was regular kfree() from call_rcu(). And the callback > > execution was typically delayed long enough to never 'see' the > > use-after-free. > > Figures. ;-) > > Well, there is always the approach of booting your big systems with most > of the CPUs turned off. Another approach would be to set HZ=10000 or > some such, assuming the kernel can actually survive that kind of abuse.
And yet another approach is to have a pair of low-priority processes per CPU that context-switch back and forth to each other if that CPU has nothing else to do. This should get rid of most of the increase in grace-period duration with increasing numbers of CPUs. Thanx, Paul -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/