Re: [dpdk-dev] [PATCH v2 3/4] net/ixgbe: convert to new Rx offloads API

["Ananyev, Konstantin" <konstantin.ananyev@intel.com>]

Hi Wei,

> 
> Hi, Konstantin
> Thanks for your feedback.
> 
> > -----Original Message-----
> > From: Ananyev, Konstantin
> > Sent: Thursday, March 15, 2018 5:48 AM
> > To: Dai, Wei <wei.dai@intel.com>; Lu, Wenzhuo <wenzhuo.lu@intel.com>
> > Cc: dev@dpdk.org
> > Subject: RE: [PATCH v2 3/4] net/ixgbe: convert to new Rx offloads API
> >
> > Hi Wei,
> >
> > > -----Original Message-----
> > > From: Dai, Wei
> > > Sent: Wednesday, March 7, 2018 1:06 PM
> > > To: Lu, Wenzhuo <wenzhuo.lu@intel.com>; Ananyev, Konstantin
> > > <konstantin.ananyev@intel.com>
> > > Cc: dev@dpdk.org; Dai, Wei <wei.dai@intel.com>
> > > Subject: [PATCH v2 3/4] net/ixgbe: convert to new Rx offloads API
> > >
> > > Ethdev Rx offloads API has changed since:
> > > commit ce17eddefc20 ("ethdev: introduce Rx queue offloads API") This
> > > commit support the new Rx offloads API.
> > >
> > > Signed-off-by: Wei Dai <wei.dai@intel.com>
> > > ---
> > >  drivers/net/ixgbe/ixgbe_ethdev.c          |  93 +++++++++--------
> > >  drivers/net/ixgbe/ixgbe_ipsec.c           |   8 +-
> > >  drivers/net/ixgbe/ixgbe_rxtx.c            | 163
> > ++++++++++++++++++++++++++----
> > >  drivers/net/ixgbe/ixgbe_rxtx.h            |   3 +
> > >  drivers/net/ixgbe/ixgbe_rxtx_vec_common.h |   2 +-
> > >  drivers/net/ixgbe/ixgbe_rxtx_vec_neon.c   |   2 +-
> > >  6 files changed, 205 insertions(+), 66 deletions(-)
> > >
> > > +uint64_t
> > > +ixgbe_get_rx_queue_offloads(struct rte_eth_dev *dev) {
> > > +	uint64_t offloads;
> > > +	struct ixgbe_hw *hw =
> > > +IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
> > > +
> > > +	offloads = DEV_RX_OFFLOAD_HEADER_SPLIT;
> >
> > As I can see I ixgbe all header_split code is enabled only if
> > RTE_HEADER_SPLIT_ENABLE is on.
> > It is off by default and I doubt anyone really using it these days.
> > So I think the best thing would be not to advertise
> > DEV_RX_OFFLOAD_HEADER_SPLIT for ixgbe at all, and probably remove
> > related code.
> > If you'd prefer to keep it, then at least we should set that capability only at
> > #ifdef RTE_HEADER_SPLIT_ENABLE.
> > Another thing - 	it should be per port, not per queue.
> > Thought I think better is just to remove it completely.
> I will set this header splitting capability in #ifdef RTE_HEADER_SPLIT_ENABLE in my next patch set.
> I think it is a per queue capability as it can be configured on the register IXGBE_SRRCTL of every Rx queue
> In this code line: IXGBE_WRITE_REG(hw, IXGBE_SRRCTL(rxq->reg_idx), srrctl); in ixgbe_dev_rx_init( ).
> Same case is also in the code line: IXGBE_WRITE_REG(hw, IXGBE_VFSRRCTL(i), srrctl); in ixgbevf_dev_rx_init( ).

Yes, HW can enable/disable it on a per queue basis.
Though it affects rx function selection, and as right now we have one rx function per device -
That's why it looks to me more like a per port offload.
Though I believe these days ixgbe PMD doesn't support it properly anyway 
(we always set rxd.hdr_addr to zero) - so probably better to remove it at all.

> 
> > > +static int
> > > +ixgbe_check_rx_queue_offloads(struct rte_eth_dev *dev, uint64_t
> > > +requested) {
> > > +	uint64_t port_offloads = dev->data->dev_conf.rxmode.offloads;
> > > +	uint64_t queue_supported = ixgbe_get_rx_queue_offloads(dev);
> > > +	uint64_t port_supported = ixgbe_get_rx_port_offloads(dev);
> > > +
> > > +	if ((requested & (queue_supported | port_supported)) != requested)
> > > +		return 0;
> > > +
> > > +	if ((port_offloads ^ requested) & port_supported)
> >
> > Could you explain a bit more what are you cheking here?
> > As I can see:
> >  (port_offloads ^ requested) - that's a diff between already set and newly
> > requested offloads.
> > Then you check if that diff consists of supported by port offloads, and if yes
> > you return an error?
> > Konstantin
> >
> This function is similar to mlx4_check_rx_queue_offloads() in mlx4 driver.
> As the git log message in the commit ce17eddefc20285bbfe575bdc07f42f0b20f34cb say
> that a per port capability should has same setting (enabling or disabling) on both port
> configuration via rte_eth_dev_configure( ) and queue configuration via rte_eth_rx_queue_setup( ).
> This function check if this requirement is matched or not.
> It also check offloading request is supported as a per port or a per queue capability or not.
> If above checking is pass, it return 1 else return 0.

Ok, let be more specific here.
Let say:
requested == DEV_RX_OFFLOAD_VLAN_STRIP;
port_offloads == DEV_RX_OFFLOAD_IPV4_CKSUM;
port_supported = (DEV_RX_OFFLOAD_IPV4_CKSUM  |
		   DEV_RX_OFFLOAD_UDP_CKSUM   |
		   DEV_RX_OFFLOAD_TCP_CKSUM   |
		   DEV_RX_OFFLOAD_CRC_STRIP   |
		   DEV_RX_OFFLOAD_JUMBO_FRAME |
		   DEV_RX_OFFLOAD_SCATTER);

(port_offloads ^ requested) == DEV_RX_OFFLOAD_VLAN_STRIP | DEV_RX_OFFLOAD_IPV4_CKSUM;
(port_offloads ^ requested) & port_supported == DEV_RX_OFFLOAD_IPV4_CKSUM;
And that function will return failure, while as I understand it shouldn't - requested queue offload is valid.

Konstantin