Re: [dpdk-dev] [PATCH v2 04/15] net/mlx5: support Rx tunnel type identification

["Xueming(Steven) Li" <xuemingl@mellanox.com>]

> -----Original Message-----
> From: Nélio Laranjeiro <nelio.laranjeiro@6wind.com>
> Sent: Friday, April 13, 2018 4:38 PM
> To: Xueming(Steven) Li <xuemingl@mellanox.com>
> Cc: Shahaf Shuler <shahafs@mellanox.com>; dev@dpdk.org
> Subject: Re: [PATCH v2 04/15] net/mlx5: support Rx tunnel type
> identification
> 
> On Thu, Apr 12, 2018 at 02:27:45PM +0000, Xueming(Steven) Li wrote:
> > > -----Original Message-----
> > > From: Nélio Laranjeiro <nelio.laranjeiro@6wind.com>
> > > Sent: Thursday, April 12, 2018 5:51 PM
> > > To: Xueming(Steven) Li <xuemingl@mellanox.com>
> > > Cc: Shahaf Shuler <shahafs@mellanox.com>; dev@dpdk.org
> > > Subject: Re: [PATCH v2 04/15] net/mlx5: support Rx tunnel type
> > > identification
> > >
> > > On Wed, Apr 11, 2018 at 08:11:50AM +0000, Xueming(Steven) Li wrote:
> > > > Hi Nelio,
> > > >
> > > > > -----Original Message-----
> > > > > From: Nélio Laranjeiro <nelio.laranjeiro@6wind.com>
> > > > > Sent: Tuesday, April 10, 2018 11:17 PM
> > > > > To: Xueming(Steven) Li <xuemingl@mellanox.com>
> > > > > Cc: Shahaf Shuler <shahafs@mellanox.com>; dev@dpdk.org
> > > > > Subject: Re: [PATCH v2 04/15] net/mlx5: support Rx tunnel type
> > > > > identification
> > > > >
> > > > > On Tue, Apr 10, 2018 at 09:34:04PM +0800, Xueming Li wrote:
> > > > > > This patch introduced tunnel type identification based on flow
> rules.
> > > > > > If flows of multiple tunnel types built on same queue,
> > > > > > RTE_PTYPE_TUNNEL_MASK will be returned, bits in flow mark
> > > > > > could be used as tunnel type identifier.
> > > > >
> > > > > I don't see anywhere in this patch where the bits are reserved
> > > > > to identify a flow, nor values which can help to identify it.
> > > > >
> > > > > Is this missing?
> > > > >
> > > > > Anyway we have already very few bits in the mark making it
> > > > > difficult to be used by the user, reserving again some to may
> > > > > lead to remove the mark support from the flows.
> > > >
> > > > Not all users will use multiple tunnel types, this is not included
> > > > in this patch set and left to user decision. I'll update comments
> > > > to make
> > > this clear.
> > >
> > > Thanks,
> > >
> > > > > > Signed-off-by: Xueming Li <xuemingl@mellanox.com>
> > > <snip/>
> > > > > >  /**
> > > > > > + * RXQ update after flow rule creation.
> > > > > > + *
> > > > > > + * @param dev
> > > > > > + *   Pointer to Ethernet device.
> > > > > > + * @param flow
> > > > > > + *   Pointer to the flow rule.
> > > > > > + */
> > > > > > +static void
> > > > > > +mlx5_flow_create_update_rxqs(struct rte_eth_dev *dev, struct
> > > > > > +rte_flow
> > > > > > +*flow) {
> > > > > > +	struct priv *priv = dev->data->dev_private;
> > > > > > +	unsigned int i;
> > > > > > +
> > > > > > +	if (!dev->data->dev_started)
> > > > > > +		return;
> > > > > > +	for (i = 0; i != flow->rss_conf.queue_num; ++i) {
> > > > > > +		struct mlx5_rxq_data *rxq_data = (*priv->rxqs)
> > > > > > +						 [(*flow->queues)[i]];
> > > > > > +		struct mlx5_rxq_ctrl *rxq_ctrl =
> > > > > > +			container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
> > > > > > +		uint8_t tunnel = PTYPE_IDX(flow->tunnel);
> > > > > > +
> > > > > > +		rxq_data->mark |= flow->mark;
> > > > > > +		if (!tunnel)
> > > > > > +			continue;
> > > > > > +		rxq_ctrl->tunnel_types[tunnel] += 1;
> > > > >
> > > > > I don't understand why you need such array, the NIC is unable to
> > > > > return the tunnel type has it returns only one bit saying tunnel.
> > > > > Why don't it store in the priv structure the current configured
> tunnel?
> > > >
> > > > This array is used to count tunnel types bound to queue, if only
> > > > one tunnel type, ptype will report that tunnel type, TUNNEL
> > > > MASK(max
> > > > value) will be returned if multiple types bound to a queue.
> > > >
> > > > Flow rss action specifies queues that binding to tunnel, thus we
> > > > can't assume all queues have same tunnel types, so this is a per
> > > > queue
> > > structure.
> > >
> > > There is something I am missing here, how in the dataplane the PMD
> > > can understand from 1 bit which kind of tunnel the packet is matching?
> >
> > The code under this line is answer, let me post here:
> > 		if (rxq_data->tunnel != flow->tunnel)
> > 			rxq_data->tunnel = rxq_data->tunnel ?
> > 					   RTE_PTYPE_TUNNEL_MASK :
> > 					   flow->tunnel;
> > If no tunnel type associated to rxq, use tunnel type from flow.
> > If a new tunnel type from flow, use RTE_PTYPE_TUNNEL_MASK.
> 
> From my understanding, when in the same queue there are several tunnel
> offloads, the mbuf ptype will contains RTE_PTYPE_TUNNEL_MASK:
> 
>    @@ -1601,7 +1605,7 @@ rxq_cq_to_pkt_type(volatile struct mlx5_cqe *cqe)
>            * bit[7] = outer_l3_type
>            */
>           idx = ((pinfo & 0x3) << 6) | ((ptype & 0xfc00) >> 10);
>   -       return mlx5_ptype_table[idx];
>   +       return mlx5_ptype_table[idx] | rxq->tunnel * !!(idx & (1 << 6));
>    }
> 
> 
> Used by Rx burst functions,
> 
> /* Update packet information. */
>  pkt->packet_type = rxq_cq_to_pkt_type(cqe);
> 
> Is this correct?

You got the point.

> 
> There is another strange point here,
> 
>  +       [PTYPE_IDX(RTE_PTYPE_TUNNEL_VXLAN)] = RTE_PTYPE_TUNNEL_VXLAN |
>  +                                             RTE_PTYPE_L4_UDP,
> 
> According to the RFC 7348 [1] having a VXLAN with an outer IPv6 is
> possible.  How do you handle it?

The answer was hide in the code you pasted:

    @@ -1601,7 +1605,7 @@ rxq_cq_to_pkt_type(volatile struct mlx5_cqe *cqe)
            * bit[7] = outer_l3_type
            */
           idx = ((pinfo & 0x3) << 6) | ((ptype & 0xfc00) >> 10);
   -       return mlx5_ptype_table[idx];
   +       return mlx5_ptype_table[idx] | rxq->tunnel * !!(idx & (1 << 6));

In comment, Bit 7 is outer L3 type from CQE, PTYPE will be retrieved from 
mlx5_ptype_table lookup.

> 
> > > <snip/>
> > > > > > @@ -2334,9 +2414,9 @@ mlx5_flow_stop(struct rte_eth_dev *dev,
> > > > > > struct mlx5_flows *list)  {
> > > > > >  	struct priv *priv = dev->data->dev_private;
> > > > > >  	struct rte_flow *flow;
> > > > > > +	unsigned int i;
> > > > > >
> > > > > >  	TAILQ_FOREACH_REVERSE(flow, list, mlx5_flows, next) {
> > > > > > -		unsigned int i;
> > > > > >  		struct mlx5_ind_table_ibv *ind_tbl = NULL;
> > > > > >
> > > > > >  		if (flow->drop) {
> > > > > > @@ -2382,6 +2462,16 @@ mlx5_flow_stop(struct rte_eth_dev *dev,
> > > > > > struct
> > > > > mlx5_flows *list)
> > > > > >  		DRV_LOG(DEBUG, "port %u flow %p removed", dev->data-
> > > >port_id,
> > > > > >  			(void *)flow);
> > > > > >  	}
> > > > > > +	/* Cleanup Rx queue tunnel info. */
> > > > > > +	for (i = 0; i != priv->rxqs_n; ++i) {
> > > > > > +		struct mlx5_rxq_data *q = (*priv->rxqs)[i];
> > > > > > +		struct mlx5_rxq_ctrl *rxq_ctrl =
> > > > > > +			container_of(q, struct mlx5_rxq_ctrl, rxq);
> > > > > > +
> > > > > > +		memset((void *)rxq_ctrl->tunnel_types, 0,
> > > > > > +		       sizeof(rxq_ctrl->tunnel_types));
> > > > > > +		q->tunnel = 0;
> > > > > > +	}
> > > > > >  }
> > > > >
> > > > > This hunk does not handle the fact the Rx queue array may have
> > > > > some holes i.e. the application is allowed to ask for 10 queues
> > > > > and only initialise some.  In such situation this code will
> segfault.
> > > >
> > > > In other words, "q" could be NULL, correct? I'll add check for this.
> > >
> > > Correct.
> > >
> > > > BTW, there should be an action item to add such check in rss/queue
> > > > flow
> > > creation.
> > >
> > > As it is the responsibility of the application/user to make rule
> > > according to what it has configured, it has not been added.  It can
> > > still be added, but it cannot be considered as a fix.
> > >
> > > > > It should only memset the Rx queues making part of the flow not
> > > > > the
> > > others.
> > > >
> > > > Clean this(decrease tunnel_types counter of each queue) from each
> > > > flow would be time consuming.
> > >
> > > Considering flows are already relying on syscall to communicate with
> > > the kernel, the extra cycles consumption to only clear the queues
> > > making part of this flow is neglectable.
> > >
> > > By the way in the same function the mark is cleared only for the
> > > queues making part of the flow, the same loop can be used to clear
> > > those tunnel informations at the same time.
> > >
> > > > If an error happened, counter will not be cleared and such state
> > > > will impact tunnel type after port start again.
> > >
> > > Unless an implementation error which other kind of them do you fear
> > > to happen?
> >
> > Mark of rxq simply reset to 0, this field is counter, the final target
> > is to clear field value, so my code should be straight forward and
> > error free 😊
> >
> > From a quick look, this function could be much simple that what it is
> today:
> > 1. clean verb flow and hrex where possible, despite of flow type.
> > 2. clean rxq state: mark and tunnel_types.
> 
> Ok.
> 
> Thanks,
> 
> [1]
> https://emea01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fdpdk.or
> g%2Fpatch%2F37965&data=02%7C01%7Cxuemingl%40mellanox.com%7Cc3de5d7cfc85463
> 41a6808d5a119c916%7Ca652971c7d2e4d9ba6a4d149256f461b%7C0%7C0%7C63659205449
> 1963568&sdata=Eoq30ySZ8gRhemDG6BDawVqFvWB1gI85GpcYIMwl32Q%3D&reserved=0
> 
> --
> Nélio Laranjeiro
> 6WIND