[dpdk-dev] [PATCH v4] vhost: use SMP barriers instead of compiler ones.
Ananyev, Konstantin
konstantin.ananyev at intel.com
Mon Mar 21 15:07:32 CET 2016
> -----Original Message-----
> From: dev [mailto:dev-bounces at dpdk.org] On Behalf Of Ilya Maximets
> Sent: Monday, March 21, 2016 4:50 AM
> To: Yuanhan Liu
> Cc: dev at dpdk.org; Xie, Huawei; Dyasly Sergey
> Subject: Re: [dpdk-dev] [PATCH v4] vhost: use SMP barriers instead of compiler ones.
>
>
>
> On 18.03.2016 15:41, Yuanhan Liu wrote:
> > On Fri, Mar 18, 2016 at 03:23:53PM +0300, Ilya Maximets wrote:
> >> Since commit 4c02e453cc62 ("eal: introduce SMP memory barriers") virtio
> >> uses architecture dependent SMP barriers. vHost should use them too.
> >>
> >> Fixes: 4c02e453cc62 ("eal: introduce SMP memory barriers")
> >>
> >> Signed-off-by: Ilya Maximets <i.maximets at samsung.com>
> >> ---
> >> lib/librte_vhost/vhost_rxtx.c | 7 ++++---
> >> 1 file changed, 4 insertions(+), 3 deletions(-)
> >>
> >> diff --git a/lib/librte_vhost/vhost_rxtx.c b/lib/librte_vhost/vhost_rxtx.c
> >> index b4da665..859c669 100644
> >> --- a/lib/librte_vhost/vhost_rxtx.c
> >> +++ b/lib/librte_vhost/vhost_rxtx.c
> >> @@ -315,7 +315,7 @@ virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
> >> rte_prefetch0(&vq->desc[desc_indexes[i+1]]);
> >> }
> >>
> >> - rte_compiler_barrier();
> >> + rte_smp_wmb();
> >>
> >> /* Wait until it's our turn to add our buffer to the used ring. */
> >> while (unlikely(vq->last_used_idx != res_start_idx))
> >> @@ -565,7 +565,7 @@ virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
> >>
> >> nr_used = copy_mbuf_to_desc_mergeable(dev, vq, start, end,
> >> pkts[pkt_idx]);
> >> - rte_compiler_barrier();
> >> + rte_smp_wmb();
> >>
> >> /*
> >> * Wait until it's our turn to add our buffer
> >> @@ -923,7 +923,8 @@ rte_vhost_dequeue_burst(struct virtio_net *dev, uint16_t queue_id,
> >> sizeof(vq->used->ring[used_idx]));
> >> }
> >>
> >> - rte_compiler_barrier();
> >> + rte_smp_wmb();
> >> + rte_smp_rmb();
> >
> > rte_smp_mb?
>
> rte_smp_mb() is a real mm_fence() on x86. And we don't need to synchronize reads with
> writes here, only reads with reads and writes with writes. It is enough because next
> increment uses read and write. Pair of barriers is better because it will not impact
> on performance on x86.
Not arguing about that particular patch, just a question:
Why do we have:
#define rte_smp_mb() rte_mb()
for x86?
Why not just:
#define rte_smp_mb() rte_compiler_barrier()
here?
I meant for situations when we do need real mfence, there is an 'rte_mb' to use.
Konstantin
>
> Best regards, Ilya Maximets.
More information about the dev
mailing list