diff mbox series

[1/2] net/mlx4: fix memory barrier incorrectly placed

Message ID 1620648363-31769-2-git-send-email-fengchengwen@huawei.com (mailing list archive)
State Accepted, archived
Delegated to: Raslan Darawsheh
Headers show
Series bugfix for mlx4 and mlx5 | expand

Checks

Context Check Description
ci/checkpatch success coding style OK

Commit Message

fengchengwen May 10, 2021, 12:06 p.m. UTC
The memory barrier is used to ensure that the response is returned
only after the Tx/Rx function is set, it should place after the Rx/Tx
function is set.

Fixes: 0203d33a1059 ("net/mlx4: support secondary process")
Cc: stable@dpdk.org

Signed-off-by: Chengwen Feng <fengchengwen@huawei.com>
---
 drivers/net/mlx4/mlx4_mp.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

Comments

Viacheslav Ovsiienko May 11, 2021, 9:23 a.m. UTC | #1
> -----Original Message-----
> From: dev <dev-bounces@dpdk.org> On Behalf Of Chengwen Feng
> Sent: Monday, May 10, 2021 15:06
> To: NBU-Contact-Thomas Monjalon <thomas@monjalon.net>;
> ferruh.yigit@intel.com
> Cc: dev@dpdk.org
> Subject: [dpdk-dev] [PATCH 1/2] net/mlx4: fix memory barrier incorrectly
> placed
> 
> The memory barrier is used to ensure that the response is returned only
> after the Tx/Rx function is set, it should place after the Rx/Tx function is set.
> 
> Fixes: 0203d33a1059 ("net/mlx4: support secondary process")
> Cc: stable@dpdk.org
> 
> Signed-off-by: Chengwen Feng <fengchengwen@huawei.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>


Hi, Chengwen

Nice catch, thank you for the patches.
Just out of curiosity - did we meet the real issue with this ineffective barrier?

With best regards,
Slava

> ---
>  drivers/net/mlx4/mlx4_mp.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/drivers/net/mlx4/mlx4_mp.c b/drivers/net/mlx4/mlx4_mp.c
> index ddf7bdb..8fcfb54 100644
> --- a/drivers/net/mlx4/mlx4_mp.c
> +++ b/drivers/net/mlx4/mlx4_mp.c
> @@ -126,7 +126,6 @@ mp_secondary_handle(const struct rte_mp_msg
> *mp_msg, const void *peer)
>  	switch (param->type) {
>  	case MLX4_MP_REQ_START_RXTX:
>  		INFO("port %u starting datapath", dev->data->port_id);
> -		rte_mb();
>  		dev->tx_pkt_burst = mlx4_tx_burst;
>  		dev->rx_pkt_burst = mlx4_rx_burst;
>  #ifdef HAVE_IBV_MLX4_UAR_MMAP_OFFSET
> @@ -144,6 +143,7 @@ mp_secondary_handle(const struct rte_mp_msg
> *mp_msg, const void *peer)
>  			}
>  		}
>  #endif
> +		rte_mb();
>  		mp_init_msg(dev, &mp_res, param->type);
>  		res->result = 0;
>  		ret = rte_mp_reply(&mp_res, peer);
> --
> 2.8.1
fengchengwen May 11, 2021, 11:08 a.m. UTC | #2
On 2021/5/11 17:23, Slava Ovsiienko wrote:
>> -----Original Message-----
>> From: dev <dev-bounces@dpdk.org> On Behalf Of Chengwen Feng
>> Sent: Monday, May 10, 2021 15:06
>> To: NBU-Contact-Thomas Monjalon <thomas@monjalon.net>;
>> ferruh.yigit@intel.com
>> Cc: dev@dpdk.org
>> Subject: [dpdk-dev] [PATCH 1/2] net/mlx4: fix memory barrier incorrectly
>> placed
>>
>> The memory barrier is used to ensure that the response is returned only
>> after the Tx/Rx function is set, it should place after the Rx/Tx function is set.
>>
>> Fixes: 0203d33a1059 ("net/mlx4: support secondary process")
>> Cc: stable@dpdk.org
>>
>> Signed-off-by: Chengwen Feng <fengchengwen@huawei.com>
> Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
> 
> 
> Hi, Chengwen
> 
> Nice catch, thank you for the patches.
> Just out of curiosity - did we meet the real issue with this ineffective barrier?
> 
> With best regards,
> Slava
> 

Hi, Slava

It's just a theoretical analysis that's possible, I think it should never
happen in practice.
We found this problem when checking the memory barrier of the hns3 driver
(ps: the implementation of hns3 mp is somewhat similar to that of mlx), this
memory barrier got us a little confused, so we fix it to make it just less
confusing.

Best Regards

>> ---
>>  drivers/net/mlx4/mlx4_mp.c | 2 +-
>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/drivers/net/mlx4/mlx4_mp.c b/drivers/net/mlx4/mlx4_mp.c
>> index ddf7bdb..8fcfb54 100644
>> --- a/drivers/net/mlx4/mlx4_mp.c
>> +++ b/drivers/net/mlx4/mlx4_mp.c
>> @@ -126,7 +126,6 @@ mp_secondary_handle(const struct rte_mp_msg
>> *mp_msg, const void *peer)
>>  	switch (param->type) {
>>  	case MLX4_MP_REQ_START_RXTX:
>>  		INFO("port %u starting datapath", dev->data->port_id);
>> -		rte_mb();
>>  		dev->tx_pkt_burst = mlx4_tx_burst;
>>  		dev->rx_pkt_burst = mlx4_rx_burst;
>>  #ifdef HAVE_IBV_MLX4_UAR_MMAP_OFFSET
>> @@ -144,6 +143,7 @@ mp_secondary_handle(const struct rte_mp_msg
>> *mp_msg, const void *peer)
>>  			}
>>  		}
>>  #endif
>> +		rte_mb();
>>  		mp_init_msg(dev, &mp_res, param->type);
>>  		res->result = 0;
>>  		ret = rte_mp_reply(&mp_res, peer);
>> --
>> 2.8.1
> 
> 
> .
>
Viacheslav Ovsiienko May 11, 2021, 12:02 p.m. UTC | #3
> -----Original Message-----
> From: fengchengwen <fengchengwen@huawei.com>
> Sent: Tuesday, May 11, 2021 14:09
> To: Slava Ovsiienko <viacheslavo@nvidia.com>; NBU-Contact-Thomas
> Monjalon <thomas@monjalon.net>; ferruh.yigit@intel.com
> Cc: dev@dpdk.org
> Subject: Re: [dpdk-dev] [PATCH 1/2] net/mlx4: fix memory barrier incorrectly
> placed
> 
> 
> 
> On 2021/5/11 17:23, Slava Ovsiienko wrote:
> >> -----Original Message-----
> >> From: dev <dev-bounces@dpdk.org> On Behalf Of Chengwen Feng
> >> Sent: Monday, May 10, 2021 15:06
> >> To: NBU-Contact-Thomas Monjalon <thomas@monjalon.net>;
> >> ferruh.yigit@intel.com
> >> Cc: dev@dpdk.org
> >> Subject: [dpdk-dev] [PATCH 1/2] net/mlx4: fix memory barrier
> >> incorrectly placed
> >>
> >> The memory barrier is used to ensure that the response is returned
> >> only after the Tx/Rx function is set, it should place after the Rx/Tx
> function is set.
> >>
> >> Fixes: 0203d33a1059 ("net/mlx4: support secondary process")
> >> Cc: stable@dpdk.org
> >>
> >> Signed-off-by: Chengwen Feng <fengchengwen@huawei.com>
> > Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
> >
> >
> > Hi, Chengwen
> >
> > Nice catch, thank you for the patches.
> > Just out of curiosity - did we meet the real issue with this ineffective
> barrier?
> >
> > With best regards,
> > Slava
> >
> 
> Hi, Slava
> 
> It's just a theoretical analysis that's possible, I think it should never happen in
> practice.
> We found this problem when checking the memory barrier of the hns3 driver
> (ps: the implementation of hns3 mp is somewhat similar to that of mlx), this
> memory barrier got us a little confused, so we fix it to make it just less
> confusing.

OK, thank you for the answer.

With best regards,
Slava


> 
> Best Regards
> 
> >> ---
> >>  drivers/net/mlx4/mlx4_mp.c | 2 +-
> >>  1 file changed, 1 insertion(+), 1 deletion(-)
> >>
> >> diff --git a/drivers/net/mlx4/mlx4_mp.c b/drivers/net/mlx4/mlx4_mp.c
> >> index ddf7bdb..8fcfb54 100644
> >> --- a/drivers/net/mlx4/mlx4_mp.c
> >> +++ b/drivers/net/mlx4/mlx4_mp.c
> >> @@ -126,7 +126,6 @@ mp_secondary_handle(const struct rte_mp_msg
> >> *mp_msg, const void *peer)
> >>  	switch (param->type) {
> >>  	case MLX4_MP_REQ_START_RXTX:
> >>  		INFO("port %u starting datapath", dev->data->port_id);
> >> -		rte_mb();
> >>  		dev->tx_pkt_burst = mlx4_tx_burst;
> >>  		dev->rx_pkt_burst = mlx4_rx_burst;  #ifdef
> >> HAVE_IBV_MLX4_UAR_MMAP_OFFSET @@ -144,6 +143,7 @@
> >> mp_secondary_handle(const struct rte_mp_msg *mp_msg, const void
> >> *peer)
> >>  			}
> >>  		}
> >>  #endif
> >> +		rte_mb();
> >>  		mp_init_msg(dev, &mp_res, param->type);
> >>  		res->result = 0;
> >>  		ret = rte_mp_reply(&mp_res, peer);
> >> --
> >> 2.8.1
> >
> >
> > .
> >
diff mbox series

Patch

diff --git a/drivers/net/mlx4/mlx4_mp.c b/drivers/net/mlx4/mlx4_mp.c
index ddf7bdb..8fcfb54 100644
--- a/drivers/net/mlx4/mlx4_mp.c
+++ b/drivers/net/mlx4/mlx4_mp.c
@@ -126,7 +126,6 @@  mp_secondary_handle(const struct rte_mp_msg *mp_msg, const void *peer)
 	switch (param->type) {
 	case MLX4_MP_REQ_START_RXTX:
 		INFO("port %u starting datapath", dev->data->port_id);
-		rte_mb();
 		dev->tx_pkt_burst = mlx4_tx_burst;
 		dev->rx_pkt_burst = mlx4_rx_burst;
 #ifdef HAVE_IBV_MLX4_UAR_MMAP_OFFSET
@@ -144,6 +143,7 @@  mp_secondary_handle(const struct rte_mp_msg *mp_msg, const void *peer)
 			}
 		}
 #endif
+		rte_mb();
 		mp_init_msg(dev, &mp_res, param->type);
 		res->result = 0;
 		ret = rte_mp_reply(&mp_res, peer);