[dpdk-dev] [PATCH] rte_ring: fix racy dequeue/enqueue in ppc64

[Jerin Jacob]

-----Original Message-----
> Date: Tue, 17 Jul 2018 11:54:18 +0900
> From: Takeshi Yoshimura <t.yoshimura8869 at gmail.com>
> To: Jerin Jacob <jerin.jacob at caviumnetworks.com>
> Cc: dev at dpdk.org, stable at dpdk.org
> Subject: Re: [dpdk-dev] [PATCH] rte_ring: fix racy dequeue/enqueue in ppc64
> 

Cc: olivier.matz at 6wind.com
Cc: chaozhu at linux.vnet.ibm.com
Cc: konstantin.ananyev at intel.com

> 
> > Adding rte_smp_rmb() cause performance regression on non x86 platforms.
> > Having said that, load-load barrier can be expressed very  well with C11 memory
> > model. I guess ppc64 supports C11 memory model. If so,
> > Could you try CONFIG_RTE_RING_USE_C11_MEM_MODEL=y for ppc64 and check
> > original issue?
> 
> Yes, the performance regression happens on non-x86 with single
> producer/consumer.
> The average latency of an enqueue was increased from 21 nsec to 24 nsec in my
> simple experiment. But, I think it is worth it.

That varies to machine to machine. What is the burst size etc.

> 
> 
> I also tested C11 rte_ring, however, it caused the same race condition in ppc64.
> I tried to fix the C11 problem as well, but I also found the C11
> rte_ring had other potential
> incorrect choices of memory orders, which caused another race
> condition in ppc64.

Does it happens on all ppc64 machines? Or on a specific machine?
Is following tests are passing on your system without the patch?
test/test/test_ring_perf.c
test/test/test_ring.c

> 
> For example,
> __ATOMIC_ACQUIRE is passed to __atomic_compare_exchange_n(), but
> I am not sure why the load-acquire is used for the compare exchange.

It correct as per C11 acquire and release semantics.

> Also in update_tail, the pause can be called before the data copy because
> of ht->tail load without atomic_load_n.
> 
> The memory order is simply difficult, so it might take a bit longer
> time to check
> if the code is correct. I think I can fix the C11 rte_ring as another patch.
> 
> >>
> >> SPDK blobfs encountered a crash around rte_ring dequeues in ppc64.
> >> It uses a single consumer and multiple producers for a rte_ring.
> >> The problem was a load-load reorder in rte_ring_sc_dequeue_bulk().
> >
> > Adding rte_smp_rmb() cause performance regression on non x86 platforms.
> > Having said that, load-load barrier can be expressed very  well with C11 memory
> > model. I guess ppc64 supports C11 memory model. If so,
> > Could you try CONFIG_RTE_RING_USE_C11_MEM_MODEL=y for ppc64 and check
> > original issue?
> >
> >>
> >> The reordered loads happened on r->prod.tail in

There is rte_smp_rmb() just before reading r->prod.tail in 
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
_rte_ring_move_cons_head(). Would that not suffice the requirement?

Can you check adding compiler barrier and see is compiler is reordering
the stuff?

DPDK's ring implementation is based freebsd's ring implementation, I
don't see need for such barrier

https://github.com/freebsd/freebsd/blob/master/sys/sys/buf_ring.h

If it is something specific to ppc64 or a specific ppc64 machine, we 
could add a compile option as it is arch specific to avoid performance
impact on other architectures.

> >> __rte_ring_move_cons_head() (rte_ring_generic.h) and ring[idx] in
> >> DEQUEUE_PTRS() (rte_ring.h). They have a load-load control
> >> dependency, but the code does not satisfy it. Note that they are
> >> not reordered if __rte_ring_move_cons_head() with is_sc != 1 because
> >> cmpset invokes a read barrier.
> >>
> >> The paired stores on these loads are in ENQUEUE_PTRS() and
> >> update_tail(). Simplified code around the reorder is the following.
> >>
> >> Consumer             Producer
> >> load idx[ring]
> >>                      store idx[ring]
> >>                      store r->prod.tail
> >> load r->prod.tail
> >>
> >> In this case, the consumer loads old idx[ring] and confirms the load
> >> is valid with the new r->prod.tail.
> >>
> >> I added a read barrier in the case where __IS_SC is passed to
> >> __rte_ring_move_cons_head(). I also fixed __rte_ring_move_prod_head()
> >> to avoid similar problems with a single producer.
> >>
> >> Cc: stable at dpdk.org
> >>
> >> Signed-off-by: Takeshi Yoshimura <tyos at jp.ibm.com>
> >> ---
> >>  lib/librte_ring/rte_ring_generic.h | 10 ++++++----
> >>  1 file changed, 6 insertions(+), 4 deletions(-)
> >>
> >> diff --git a/lib/librte_ring/rte_ring_generic.h b/lib/librte_ring/rte_ring_generic.h
> >> index ea7dbe5b9..477326180 100644
> >> --- a/lib/librte_ring/rte_ring_generic.h
> >> +++ b/lib/librte_ring/rte_ring_generic.h
> >> @@ -90,9 +90,10 @@ __rte_ring_move_prod_head(struct rte_ring *r, unsigned int is_sp,
> >>                         return 0;
> >>
> >>                 *new_head = *old_head + n;
> >> -               if (is_sp)
> >> +               if (is_sp) {
> >> +                       rte_smp_rmb();
> >>                         r->prod.head = *new_head, success = 1;
> >> -               else
> >> +               } else
> >>                         success = rte_atomic32_cmpset(&r->prod.head,
> >>                                         *old_head, *new_head);
> >>         } while (unlikely(success == 0));
> >> @@ -158,9 +159,10 @@ __rte_ring_move_cons_head(struct rte_ring *r, unsigned int is_sc,
> >>                         return 0;
> >>
> >>                 *new_head = *old_head + n;
> >> -               if (is_sc)
> >> +               if (is_sc) {
> >> +                       rte_smp_rmb();
> >>                         r->cons.head = *new_head, success = 1;
> >> -               else
> >> +               } else
> >>                         success = rte_atomic32_cmpset(&r->cons.head, *old_head,
> >>                                         *new_head);
> >>         } while (unlikely(success == 0));
> >> --
> >> 2.17.1
> >>