[dpdk-dev] mlx5 flow create/destroy behaviour

[Nélio Laranjeiro]

Hi Allain,

My attempt to reproduce it was a failure, may be I missed something,
please see below,

On Tue, Mar 28, 2017 at 12:42:05PM +0000, Legacy, Allain wrote:
> Hi,
> I am setting up an experiment to gauge the usability of the flow API
> and the flow marking behavior of the CX4.   I am working from v17.02.
> I am seeing some unpredictable behavior that I am unsure of the cause. 
> 
> This is the layout of the test:
>    
>    2 x CX4 (15b3:1015) 
>       + 1 port used on each
>    A test application with 1 core, and 1 queue/port
>    Traffic generator attached to each port
>       + 500 unique src+dst MAC address combinations sent from each port
>       + All traffic is VLAN tagged (1 VLAN per port)
> 
> The test application examines packets as they are received on each
> port.  It sets up flow rules and calls rte_flow_create() for each new
> layer2 flow that it observes.    The flow patterns are of the form
> ETH+VLAN+END where ETH matches src+dst+type=vlan, VLAN matches the
> port's VLAN ID.  The flow actions are of the form MARK+QUEUE+END where
> MARK assigns a unique integer to each flow and, and QUEUE assigns the
> flow to queue_id=0 (since the test app only has 1 queue per port).

If I understand correctly, your application is adding 500 rules like:

 flow create 0 ingress pattern eth src is <smac> dst is <dmac> / vlan vid is <vid> / end action mark id is <id> / queue index 0 / end

> Once the flows are setup, the application then checks that ingress
> packets are properly marked with the intended unique integer specified
> in the MARK action.

It is sending packets to verify this?

> The traffic is run for a short period of time and then stopped.  Once
> the traffic is stopped the application removes the flow rules by
> calling rte_flow_destroy().    There is no guarantee that the order of
> the destroys resembles in any way the order of the creates.   (I
> mention this because of this warning in rte_flow.h:  "This function is
> only guaranteed to succeed if handles are destroyed in reverse order
> of their creation.").   All of the calls to rte_flow_destroy()
> succeed. 
> 
> When I run this test after the NIC has been reset there are no issues.

What do you mean by "reset"?

> All calls to rte_flow_create()/rte_flow_destroy() succeed and all
> packets have a valid mark ID that corresponds to the unique integer
> assigned to that src+dst+vlan grouping.

In mlx5 PMD rte_flow_destroy() always returns success as the destruction
should never fail.
Can you compile in debug mode (by setting CONFIG_RTE_LIBRTE_MLX5_DEBUG
to "y")?  Then you should have as many print for the creation rules than
the destroyed ones.

> The problem happens when I run this test for a second or third time
> without first resetting the NIC.  On subsequent test runs I still see
> no errors in create/destroy API calls but packets are no longer marked
> by the hardware.  In some test runs none of the flows have valid mark
> id values, and other test runs have some percentage of flows with
> valid mark id values while others do not.   The behavior seems
> inconsistent but if I reset the NIC the behavior goes back to working
> for 1 test run and then starts behaving incorrectly again on
> subsequent runs.
> 
> I should note that in subsequent test runs the MAC addresses are the
> same as previous runs, and but the mapping from unique integer to
> src+dst+vlan are different each time.
> 
> Is this behavior consistent with your experience using the device
> and/or API?

No I did not face such issue, the behavior was consistent, but I never
tried to generate so many rules in the past.


Thanks,

-- 
Nélio Laranjeiro
6WIND