[dpdk-dev] [PATCH] test/pmd_perf: change the way to drain the port

[dpdk-dev] [PATCH] test/pmd_perf: change the way to drain the port

[Julien Meunier]

If the port has received less than ``pkt_per_port`` packets (for
example, the port has missed some packets), the test is in an infinite
loop.

Instead of expecting a number of packet to receive, let the port to be
drained by itself. If no more packets are received, the test can
continue.

Fixes: 002ade70e933 ("app/test: measure cycles per packet in Rx/Tx")
Cc: stable@dpdk.org

Signed-off-by: Julien Meunier <julien.meunier@nokia.com>
---
 test/test/test_pmd_perf.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/test/test/test_pmd_perf.c b/test/test/test_pmd_perf.c
index f5095c8..286e09d 100644
--- a/test/test/test_pmd_perf.c
+++ b/test/test/test_pmd_perf.c
@@ -493,15 +493,15 @@ main_loop(__rte_unused void *args)
 
 	for (i = 0; i < conf->nb_ports; i++) {
 		portid = conf->portlist[i];
-		int nb_free = pkt_per_port;
+		int nb_free = 0;
 		do { /* dry out */
 			nb_rx = rte_eth_rx_burst(portid, 0,
 						 pkts_burst, MAX_PKT_BURST);
 			nb_tx = 0;
 			while (nb_tx < nb_rx)
 				rte_pktmbuf_free(pkts_burst[nb_tx++]);
-			nb_free -= nb_rx;
-		} while (nb_free != 0);
+			nb_free += nb_rx;
+		} while (nb_rx != 0);
 		printf("free %d mbuf left in port %u\n", pkt_per_port, portid);
 	}
 
-- 
2.10.2

Re: [dpdk-dev] [dpdk-stable] [PATCH] test/pmd_perf: change the way to drain the port

[Ferruh Yigit]

On 1/2/2019 3:55 PM, Julien Meunier wrote:
> If the port has received less than ``pkt_per_port`` packets (for
> example, the port has missed some packets), the test is in an infinite
> loop.
> 
> Instead of expecting a number of packet to receive, let the port to be
> drained by itself. If no more packets are received, the test can
> continue.

This looks like fixing the test_pmd_perf test case, which can stuck into endless
loop without this patch, and since there will be already a new version for below
comment, can you please update the patch title to describe the fix, like

test/pmd_perf: fix ....


> 
> Fixes: 002ade70e933 ("app/test: measure cycles per packet in Rx/Tx")
> Cc: stable@dpdk.org
> 
> Signed-off-by: Julien Meunier <julien.meunier@nokia.com>
> ---
>  test/test/test_pmd_perf.c | 6 +++---
>  1 file changed, 3 insertions(+), 3 deletions(-)
> 
> diff --git a/test/test/test_pmd_perf.c b/test/test/test_pmd_perf.c
> index f5095c8..286e09d 100644
> --- a/test/test/test_pmd_perf.c
> +++ b/test/test/test_pmd_perf.c
> @@ -493,15 +493,15 @@ main_loop(__rte_unused void *args)
>  
>  	for (i = 0; i < conf->nb_ports; i++) {
>  		portid = conf->portlist[i];
> -		int nb_free = pkt_per_port;
> +		int nb_free = 0;

'nb_free' is not more used or required, it can be removed completely I think.

>  		do { /* dry out */
>  			nb_rx = rte_eth_rx_burst(portid, 0,
>  						 pkts_burst, MAX_PKT_BURST);
>  			nb_tx = 0;
>  			while (nb_tx < nb_rx)
>  				rte_pktmbuf_free(pkts_burst[nb_tx++]);
> -			nb_free -= nb_rx;
> -		} while (nb_free != 0);
> +			nb_free += nb_rx;
> +		} while (nb_rx != 0);

Isn't there already something wrong with this logic? It assumes after test done
device still has 'pkt_per_port' packets in its queues, it tries to receive and
free them, but:

nb_free = pkt_per_port = MAX_TRAFFIC_BURST = 2048
RTE_TEST_RX_DESC_DEFAULT = RTE_TEST_TX_DESC_DEFAULT = 1024

When device queue length is 1024, how it can be holding 2048 packets? So it
can't exit from this loop. Since this should be working, what am I missing?

But overall, this stage is after the test done and for cleanup, I think your
suggestion is reasonable, only please check above 'nb_free' comment.

Thanks,
ferruh

Re: [dpdk-dev] [dpdk-stable] [PATCH] test/pmd_perf: change the way to drain the port

[Meunier, Julien (Nokia - FR/Paris-Saclay)]

Hi,

Inline reply,

On 08/01/2019 18:33, Ferruh Yigit wrote:
> On 1/2/2019 3:55 PM, Julien Meunier wrote:
>> If the port has received less than ``pkt_per_port`` packets (for
>> example, the port has missed some packets), the test is in an infinite
>> loop.
>>
>> Instead of expecting a number of packet to receive, let the port to be
>> drained by itself. If no more packets are received, the test can
>> continue.
> 
> This looks like fixing the test_pmd_perf test case, which can stuck into endless
> loop without this patch, and since there will be already a new version for below
> comment, can you please update the patch title to describe the fix, like
> 
> test/pmd_perf: fix ....

Sure ! I will be more careful next time with the title of my patches

>>
>> Fixes: 002ade70e933 ("app/test: measure cycles per packet in Rx/Tx")
>> Cc: stable@dpdk.org
>>
>> Signed-off-by: Julien Meunier <julien.meunier@nokia.com>
>> ---
>>   test/test/test_pmd_perf.c | 6 +++---
>>   1 file changed, 3 insertions(+), 3 deletions(-)
>>
>> diff --git a/test/test/test_pmd_perf.c b/test/test/test_pmd_perf.c
>> index f5095c8..286e09d 100644
>> --- a/test/test/test_pmd_perf.c
>> +++ b/test/test/test_pmd_perf.c
>> @@ -493,15 +493,15 @@ main_loop(__rte_unused void *args)
>>   
>>   	for (i = 0; i < conf->nb_ports; i++) {
>>   		portid = conf->portlist[i];
>> -		int nb_free = pkt_per_port;
>> +		int nb_free = 0;
> 
> 'nb_free' is not more used or required, it can be removed completely I think.

Damn.. Missing one little correction: nb_free should be used in the printf.

- printf("free %d mbuf left in port %u\n", pkt_per_port, portid);
+ printf("free %d mbuf left in port %u\n", nb_free, portid);

> 
>>   		do { /* dry out */
>>   			nb_rx = rte_eth_rx_burst(portid, 0,
>>   						 pkts_burst, MAX_PKT_BURST);
>>   			nb_tx = 0;
>>   			while (nb_tx < nb_rx)
>>   				rte_pktmbuf_free(pkts_burst[nb_tx++]);
>> -			nb_free -= nb_rx;
>> -		} while (nb_free != 0);
>> +			nb_free += nb_rx;
>> +		} while (nb_rx != 0);
> 
> Isn't there already something wrong with this logic? It assumes after test done
> device still has 'pkt_per_port' packets in its queues, it tries to receive and
> free them, but:
> 
> nb_free = pkt_per_port = MAX_TRAFFIC_BURST = 2048
> RTE_TEST_RX_DESC_DEFAULT = RTE_TEST_TX_DESC_DEFAULT = 1024

All ports are configured with the following number of descriptors:
nb_rxd = MAX_TRAFFIC_BURST;
nb_txd = MAX_TRAFFIC_BURST;

In this case, all is OK.

But, for the test SC_CONTINUOUS (which is, by the way, the default one), 
this number is reduced
nb_rxd = RTE_TEST_RX_DESC_DEFAULT ;
nb_txd = RTE_TEST_TX_DESC_DEFAULT ;

> When device queue length is 1024, how it can be holding 2048 packets? So it
> can't exit from this loop. Since this should be working, what am I missing?

In the main_loop, the test xmits 2048 pkts per port. Then, do_measure 
-measure_rxtx for example- receives all incoming packets and resends to 
the port.

However... Without my patch, on a previous DPDK version (17.08), which 
configures less RX and TX descriptors (RXD=128 TXD=512), I didn't notice 
this issue on ixgbe PMD (but not on fm10k).

So... You're right, how it can be holding 2048 packets with this 
configuration... I will check on my side.

> But overall, this stage is after the test done and for cleanup, I think your
> suggestion is reasonable, only please check above 'nb_free' comment.

I will submit a new patch.

Thanks for your comments !

> Thanks,
> ferruh
> 

Best regards,
Julien Meunier

[dpdk-dev] [PATCH v2] test/pmd_perf: fix the way to drain the port

[Julien Meunier]

If the port has received less than ``pkt_per_port`` packets (for
example, the port has missed some packets), the test is in an infinite
loop.

Instead of expecting a number of packet to receive, let the port to be
drained by itself. If no more packets are received, the test can
continue.

Fixes: 002ade70e933 ("app/test: measure cycles per packet in Rx/Tx")
Cc: stable@dpdk.org

Signed-off-by: Julien Meunier <julien.meunier@nokia.com>
---
v2:
* rename commit title
* fix nb_free display
---
 test/test/test_pmd_perf.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/test/test/test_pmd_perf.c b/test/test/test_pmd_perf.c
index f5095c8..c7e2df3 100644
--- a/test/test/test_pmd_perf.c
+++ b/test/test/test_pmd_perf.c
@@ -493,16 +493,16 @@ main_loop(__rte_unused void *args)
 
 	for (i = 0; i < conf->nb_ports; i++) {
 		portid = conf->portlist[i];
-		int nb_free = pkt_per_port;
+		int nb_free = 0;
 		do { /* dry out */
 			nb_rx = rte_eth_rx_burst(portid, 0,
 						 pkts_burst, MAX_PKT_BURST);
 			nb_tx = 0;
 			while (nb_tx < nb_rx)
 				rte_pktmbuf_free(pkts_burst[nb_tx++]);
-			nb_free -= nb_rx;
-		} while (nb_free != 0);
-		printf("free %d mbuf left in port %u\n", pkt_per_port, portid);
+			nb_free += nb_rx;
+		} while (nb_rx != 0);
+		printf("free %d mbuf left in port %u\n", nb_free, portid);
 	}
 
 	if (count == 0)
-- 
2.10.2

Re: [dpdk-dev] [PATCH v2] test/pmd_perf: fix the way to drain the port

[Ferruh Yigit]

On 2/3/2019 7:42 PM, Julien Meunier wrote:
> If the port has received less than ``pkt_per_port`` packets (for
> example, the port has missed some packets), the test is in an infinite
> loop.
> 
> Instead of expecting a number of packet to receive, let the port to be
> drained by itself. If no more packets are received, the test can
> continue.
> 
> Fixes: 002ade70e933 ("app/test: measure cycles per packet in Rx/Tx")
> Cc: stable@dpdk.org
> 
> Signed-off-by: Julien Meunier <julien.meunier@nokia.com>
> ---
> v2:
> * rename commit title
> * fix nb_free display
> ---
>  test/test/test_pmd_perf.c | 8 ++++----
>  1 file changed, 4 insertions(+), 4 deletions(-)
> 
> diff --git a/test/test/test_pmd_perf.c b/test/test/test_pmd_perf.c
> index f5095c8..c7e2df3 100644
> --- a/test/test/test_pmd_perf.c
> +++ b/test/test/test_pmd_perf.c
> @@ -493,16 +493,16 @@ main_loop(__rte_unused void *args)
>  
>  	for (i = 0; i < conf->nb_ports; i++) {
>  		portid = conf->portlist[i];
> -		int nb_free = pkt_per_port;
> +		int nb_free = 0;
>  		do { /* dry out */
>  			nb_rx = rte_eth_rx_burst(portid, 0,
>  						 pkts_burst, MAX_PKT_BURST);
>  			nb_tx = 0;
>  			while (nb_tx < nb_rx)
>  				rte_pktmbuf_free(pkts_burst[nb_tx++]);
> -			nb_free -= nb_rx;
> -		} while (nb_free != 0);
> -		printf("free %d mbuf left in port %u\n", pkt_per_port, portid);
> +			nb_free += nb_rx;
> +		} while (nb_rx != 0);
> +		printf("free %d mbuf left in port %u\n", nb_free, portid);


In the test logic there is an expectation that 'pkt_per_port' packets will be
received.
We are losing that intention here with this update. What do you think updating
the log to include it, like:
"free %d (expected %d) mbuf left in port %u\n", nb_free, pkt_per_port, portid

Re: [dpdk-dev] [PATCH v2] test/pmd_perf: fix the way to drain the port

[Meunier, Julien (Nokia - FR/Paris-Saclay)]

Hi,

Sorry for the delay. Inline reply.

On 07/02/2019 13:28, Ferruh Yigit wrote:
> On 2/3/2019 7:42 PM, Julien Meunier wrote:
>> If the port has received less than ``pkt_per_port`` packets (for
>> example, the port has missed some packets), the test is in an infinite
>> loop.
>>
>> Instead of expecting a number of packet to receive, let the port to be
>> drained by itself. If no more packets are received, the test can
>> continue.
>>
>> Fixes: 002ade70e933 ("app/test: measure cycles per packet in Rx/Tx")
>> Cc: stable@dpdk.org
>>
>> Signed-off-by: Julien Meunier <julien.meunier@nokia.com>
>> ---
>> v2:
>> * rename commit title
>> * fix nb_free display
>> ---
>>   test/test/test_pmd_perf.c | 8 ++++----
>>   1 file changed, 4 insertions(+), 4 deletions(-)
>>
>> diff --git a/test/test/test_pmd_perf.c b/test/test/test_pmd_perf.c
>> index f5095c8..c7e2df3 100644
>> --- a/test/test/test_pmd_perf.c
>> +++ b/test/test/test_pmd_perf.c
>> @@ -493,16 +493,16 @@ main_loop(__rte_unused void *args)
>>   
>>   	for (i = 0; i < conf->nb_ports; i++) {
>>   		portid = conf->portlist[i];
>> -		int nb_free = pkt_per_port;
>> +		int nb_free = 0;
>>   		do { /* dry out */
>>   			nb_rx = rte_eth_rx_burst(portid, 0,
>>   						 pkts_burst, MAX_PKT_BURST);
>>   			nb_tx = 0;
>>   			while (nb_tx < nb_rx)
>>   				rte_pktmbuf_free(pkts_burst[nb_tx++]);
>> -			nb_free -= nb_rx;
>> -		} while (nb_free != 0);
>> -		printf("free %d mbuf left in port %u\n", pkt_per_port, portid);
>> +			nb_free += nb_rx;
>> +		} while (nb_rx != 0);
>> +		printf("free %d mbuf left in port %u\n", nb_free, portid);
> 
> 
> In the test logic there is an expectation that 'pkt_per_port' packets will be
> received.
> We are losing that intention here with this update. What do you think updating
> the log to include it, like:
> "free %d (expected %d) mbuf left in port %u\n", nb_free, pkt_per_port, portid
> 

OK. But, after thinking, I should add a little timeout in order to drain 
the port during N cycles (like it was already done in the function 
poll_burst - timeout), just to be sure that all packets are dequeued.

I will upload a new patch today.

Best regards,
Julien Meunier

Re: [dpdk-dev] [PATCH v2] test/pmd_perf: fix the way to drain the port

[Ferruh Yigit]

On 2/18/2019 11:25 AM, Meunier, Julien (Nokia - FR/Paris-Saclay) wrote:
> Hi,
> 
> Sorry for the delay. Inline reply.
> 
> On 07/02/2019 13:28, Ferruh Yigit wrote:
>> On 2/3/2019 7:42 PM, Julien Meunier wrote:
>>> If the port has received less than ``pkt_per_port`` packets (for
>>> example, the port has missed some packets), the test is in an infinite
>>> loop.
>>>
>>> Instead of expecting a number of packet to receive, let the port to be
>>> drained by itself. If no more packets are received, the test can
>>> continue.
>>>
>>> Fixes: 002ade70e933 ("app/test: measure cycles per packet in Rx/Tx")
>>> Cc: stable@dpdk.org
>>>
>>> Signed-off-by: Julien Meunier <julien.meunier@nokia.com>
>>> ---
>>> v2:
>>> * rename commit title
>>> * fix nb_free display
>>> ---
>>>   test/test/test_pmd_perf.c | 8 ++++----
>>>   1 file changed, 4 insertions(+), 4 deletions(-)
>>>
>>> diff --git a/test/test/test_pmd_perf.c b/test/test/test_pmd_perf.c
>>> index f5095c8..c7e2df3 100644
>>> --- a/test/test/test_pmd_perf.c
>>> +++ b/test/test/test_pmd_perf.c
>>> @@ -493,16 +493,16 @@ main_loop(__rte_unused void *args)
>>>   
>>>   	for (i = 0; i < conf->nb_ports; i++) {
>>>   		portid = conf->portlist[i];
>>> -		int nb_free = pkt_per_port;
>>> +		int nb_free = 0;
>>>   		do { /* dry out */
>>>   			nb_rx = rte_eth_rx_burst(portid, 0,
>>>   						 pkts_burst, MAX_PKT_BURST);
>>>   			nb_tx = 0;
>>>   			while (nb_tx < nb_rx)
>>>   				rte_pktmbuf_free(pkts_burst[nb_tx++]);
>>> -			nb_free -= nb_rx;
>>> -		} while (nb_free != 0);
>>> -		printf("free %d mbuf left in port %u\n", pkt_per_port, portid);
>>> +			nb_free += nb_rx;
>>> +		} while (nb_rx != 0);
>>> +		printf("free %d mbuf left in port %u\n", nb_free, portid);
>>
>>
>> In the test logic there is an expectation that 'pkt_per_port' packets will be
>> received.
>> We are losing that intention here with this update. What do you think updating
>> the log to include it, like:
>> "free %d (expected %d) mbuf left in port %u\n", nb_free, pkt_per_port, portid
>>
> 
> OK. But, after thinking, I should add a little timeout in order to drain 
> the port during N cycles (like it was already done in the function 
> poll_burst - timeout), just to be sure that all packets are dequeued.

Not sure if we need this, at this stage all packets should be in device Rx
queue, can rte_eth_rx_burst() return 0 when there are packets waiting in the queue?

Anyway, this is after measurement done, and to free to the packets, so adding a
timeout (retry) mechanism won't hurt if you prefer to add this.

Thanks,
ferruh

> 
> I will upload a new patch today.
> 
> Best regards,
> Julien Meunier
> 

[dpdk-dev] [PATCH v3] test/pmd_perf: fix the way to drain the port

[Julien Meunier]

If the port has received less than ``pkt_per_port`` packets (for
example, the port has missed some packets), the test is in an infinite
loop.

Instead of expecting a number of packet to receive, let the port to be
drained by itself. If no more packets are received, the test can
continue.

Fixes: 002ade70e933 ("app/test: measure cycles per packet in Rx/Tx")
Cc: stable@dpdk.org

Signed-off-by: Julien Meunier <julien.meunier@nokia.com>
---
v3:
* add timeout on stop
* add log details
v2:
* rename commit title
* fix nb_free display
---
 test/test/test_pmd_perf.c | 13 +++++++++----
 1 file changed, 9 insertions(+), 4 deletions(-)

diff --git a/test/test/test_pmd_perf.c b/test/test/test_pmd_perf.c
index f5095c8..ed8524a 100644
--- a/test/test/test_pmd_perf.c
+++ b/test/test/test_pmd_perf.c
@@ -493,16 +493,21 @@ main_loop(__rte_unused void *args)
 
 	for (i = 0; i < conf->nb_ports; i++) {
 		portid = conf->portlist[i];
-		int nb_free = pkt_per_port;
+		int nb_free = 0;
+		uint64_t timeout = 10000;
 		do { /* dry out */
 			nb_rx = rte_eth_rx_burst(portid, 0,
 						 pkts_burst, MAX_PKT_BURST);
 			nb_tx = 0;
 			while (nb_tx < nb_rx)
 				rte_pktmbuf_free(pkts_burst[nb_tx++]);
-			nb_free -= nb_rx;
-		} while (nb_free != 0);
-		printf("free %d mbuf left in port %u\n", pkt_per_port, portid);
+			nb_free += nb_rx;
+
+			if (unlikely(nb_rx == 0))
+				timeout--;
+		} while (nb_free != pkt_per_port && timeout != 0);
+		printf("free %d (expected %d) mbuf left in port %u\n", nb_free,
+		       pkt_per_port, portid);
 	}
 
 	if (count == 0)
-- 
2.10.2

Re: [dpdk-dev] [PATCH v3] test/pmd_perf: fix the way to drain the port

[Ferruh Yigit]

On 2/20/2019 9:06 PM, Julien Meunier wrote:
> If the port has received less than ``pkt_per_port`` packets (for
> example, the port has missed some packets), the test is in an infinite
> loop.
> 
> Instead of expecting a number of packet to receive, let the port to be
> drained by itself. If no more packets are received, the test can
> continue.
> 
> Fixes: 002ade70e933 ("app/test: measure cycles per packet in Rx/Tx")
> Cc: stable@dpdk.org
> 
> Signed-off-by: Julien Meunier <julien.meunier@nokia.com>

Reviewed-by: Ferruh Yigit <ferruh.yigit@intel.com>

Applied to dpdk-next-net/master, thanks.