Re: [dpdk-dev] [RFC] ethdev: remove all offload API

[Shahaf Shuler <shahafs@mellanox.com>]

Monday, June 11, 2018 2:26 PM, Ananyev, Konstantin:
> Cc: dev@dpdk.org
> Subject: Re: [dpdk-dev] [RFC] ethdev: remove all offload API
> > Subject: Re: [RFC] ethdev: remove all offload API
> >
> > On 6/11/2018 12:00 PM, Shahaf Shuler wrote:
> > > Hi Ferruh,
> > >
> > > Thanks for this patch.
> > >
> > > Monday, June 11, 2018 12:10 PM, Ferruh Yigit:
> > >> Subject: Re: [RFC] ethdev: remove all offload API
> > >>
> > >> On 6/9/2018 9:04 AM, Andrew Rybchenko wrote:
> > >>> On 06/09/2018 01:41 AM, Ferruh Yigit wrote:
> > >>>> Cc: Shahaf Shuler <shahafs@mellanox.com>
> > >>>>
> > >>>> Signed-off-by: Ferruh Yigit <ferruh.yigit@intel.com>
> > >>>> ---
> > >>>
> > >>> <...>
> > >>>
> > >>>> diff --git a/app/test-eventdev/test_perf_common.c
> > >>>> b/app/test-eventdev/test_perf_common.c
> > >>>> index d00f91802..9fe042ffe 100644
> > >>>> --- a/app/test-eventdev/test_perf_common.c
> > >>>> +++ b/app/test-eventdev/test_perf_common.c
> > >>>> @@ -680,13 +680,7 @@ perf_ethdev_setup(struct evt_test *test,
> > >>>> struct
> > >> evt_options *opt)
> > >>>>  			.mq_mode = ETH_MQ_RX_RSS,
> > >>>>  			.max_rx_pkt_len = ETHER_MAX_LEN,
> > >>>>  			.split_hdr_size = 0,
> > >>>> -			.header_split   = 0,
> > >>>> -			.hw_ip_checksum = 0,
> > >>>> -			.hw_vlan_filter = 0,
> > >>>> -			.hw_vlan_strip  = 0,
> > >>>> -			.hw_vlan_extend = 0,
> > >>>>  			.jumbo_frame    = 0,
> > >>>> -			.hw_strip_crc   = 1,
> > >>>
> > >>
> > >> Hi Andrew,
> > >>
> > >>> I have 2 questions here:
> > >>>  1. Why is jumbo_frame kept? There is
> > >> DEV_RX_OFFLOAD_JUMBO_FRAME.
> > >>
> > >> Because there are still some usage of this flag in PMDs, they need
> > >> to be clarified before removing the flag. I am for removing the
> > >> flag in final version, but for this RFC I am not able to find enough time to
> work on PMDs for it.
> > >
> > > Can you elaborate?
> > > Is this something more than just a replace of the jumbo_frame bit with its
> corresponding flag?
> >
> > That was my concern that copy paste won't be enough because some
> > drivers not just use the jumbo_frame but set it based on
> > max_rx_pkt_len etc.., that is why left out .jumbo_frame in the RFC.

I understand the max_rx_pkt_len consideration, but in the end there is some logic which decide if to set/not the jumbo_frame bit. I would expect it just to replace it with the JUMBO_FRAME flag w/o touching the rest of the logic.

Can you refer to some specific file which you encounter it? 

> 
> But max_rx_pkt_len need to be remained (and properly processed anyway)
> no?
> BTW, I always wonder is there any reason to have jumbo_frame flag at all (as
> we do have max_rx_pkt_len anyway)?

I share the same thought, and it was expressed in multiple threads. There is more - {MTU, SCATTER, JUMBO_FRAME}, all are connected and depended only on max_rx_pkt_len and mbuf size. 
However I think it is too much work to include into this RFC and completely orthogonal. 

> Konstantin
> 
> >
> > >
> > >>
> > >>>  2. Why is hw_strip_crc=1 discarded? Yes, I remember plans to make
> > >>> it
> > >>>      default behaviour and introduce flag to keep CRC, but right
> > >>> now the
> > >>>      patch looks like mixture of two changes which is not good.
> > >>
> > >> Yes it is wrong, app should replace "".hw_strip_crc=1 with KEEP_CRC
> offload.
> > >> Since both are RFC, kind of hard to maintain, but I think good to
> > >> create a dependency from this patch to KEEP_CRC one.
> > >>
> > >>>
> > >>> There are more cases in the patch where hw_strip_crc=1 is simply
> > >> discarded.
> > >>>
> > >>> <...>
> > >>>
> > >>>> diff --git a/drivers/net/sfc/sfc_ethdev.c
> > >>>> b/drivers/net/sfc/sfc_ethdev.c index 1b6499f85..ee8ae5b9f 100644
> > >>>> --- a/drivers/net/sfc/sfc_ethdev.c
> > >>>> +++ b/drivers/net/sfc/sfc_ethdev.c
> > >>>> @@ -1089,7 +1089,6 @@ sfc_tx_queue_info_get(struct rte_eth_dev
> > >> *dev,
> > >>>> uint16_t tx_queue_id,
> > >>>>
> > >>>>  	memset(qinfo, 0, sizeof(*qinfo));
> > >>>>
> > >>>> -	qinfo->conf.txq_flags = txq_info->txq->flags;
> > >>>>  	qinfo->conf.offloads = txq_info->txq->offloads;
> > >>>>  	qinfo->conf.tx_free_thresh = txq_info->txq->free_thresh;
> > >>>>  	qinfo->conf.tx_deferred_start = txq_info->deferred_start;
> diff
> > >>>> --git a/drivers/net/sfc/sfc_rx.c b/drivers/net/sfc/sfc_rx.c index
> > >>>> cc76a5b15..58a0df583 100644
> > >>>> --- a/drivers/net/sfc/sfc_rx.c
> > >>>> +++ b/drivers/net/sfc/sfc_rx.c
> > >>>> @@ -1446,7 +1446,6 @@ sfc_rx_check_mode(struct sfc_adapter *sa,
> > >> struct rte_eth_rxmode *rxmode)
> > >>>>  	if (~rxmode->offloads & DEV_RX_OFFLOAD_CRC_STRIP) {
> > >>>>  		sfc_warn(sa, "FCS stripping cannot be disabled -
> always on");
> > >>>>  		rxmode->offloads |= DEV_RX_OFFLOAD_CRC_STRIP;
> > >>>> -		rxmode->hw_strip_crc = 1;
> > >>>>  	}
> > >>>>
> > >>>>  	return rc;
> > >>>> diff --git a/drivers/net/sfc/sfc_tx.c b/drivers/net/sfc/sfc_tx.c
> > >>>> index 1bcc2c697..6d42a1a65 100644
> > >>>> --- a/drivers/net/sfc/sfc_tx.c
> > >>>> +++ b/drivers/net/sfc/sfc_tx.c
> > >>>> @@ -171,7 +171,6 @@ sfc_tx_qinit(struct sfc_adapter *sa, unsigned
> > >>>> int
> > >> sw_index,
> > >>>>  	txq->free_thresh =
> > >>>>  		(tx_conf->tx_free_thresh) ? tx_conf-
> >tx_free_thresh :
> > >>>>  		SFC_TX_DEFAULT_FREE_THRESH;
> > >>>> -	txq->flags = tx_conf->txq_flags;
> > >>>>  	txq->offloads = offloads;
> > >>>>
> > >>>>  	rc = sfc_dma_alloc(sa, "txq", sw_index,
> > >>>> EFX_TXQ_SIZE(txq_info->entries), @@ -182,7 +181,6 @@
> > >> sfc_tx_qinit(struct sfc_adapter *sa, unsigned int sw_index,
> > >>>>  	memset(&info, 0, sizeof(info));
> > >>>>  	info.max_fill_level = txq_max_fill_level;
> > >>>>  	info.free_thresh = txq->free_thresh;
> > >>>> -	info.flags = tx_conf->txq_flags;
> > >>>>  	info.offloads = offloads;
> > >>>>  	info.txq_entries = txq_info->entries;
> > >>>>  	info.dma_desc_size_max = encp-
> >enc_tx_dma_desc_size_max;
> > >> @@ -431,18
> > >>>> +429,10 @@ sfc_tx_qstart(struct sfc_adapter *sa, unsigned int
> > >>>> +sw_index)
> > >>>>  	if (rc != 0)
> > >>>>  		goto fail_ev_qstart;
> > >>>>
> > >>>> -	/*
> > >>>> -	 * The absence of ETH_TXQ_FLAGS_IGNORE is associated
> with a
> > >> legacy
> > >>>> -	 * application which expects that IPv4 checksum offload is
> enabled
> > >>>> -	 * all the time as there is no legacy flag to turn off the offload.
> > >>>> -	 */
> > >>>> -	if ((txq->offloads & DEV_TX_OFFLOAD_IPV4_CKSUM) ||
> > >>>> -	    (~txq->flags & ETH_TXQ_FLAGS_IGNORE))
> > >>>> +	if (txq->offloads & DEV_TX_OFFLOAD_IPV4_CKSUM)
> > >>>>  		flags |= EFX_TXQ_CKSUM_IPV4;
> > >>>>
> > >>>> -	if ((txq->offloads &
> DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM) ||
> > >>>> -	    ((~txq->flags & ETH_TXQ_FLAGS_IGNORE) &&
> > >>>> -	     (offloads_supported &
> > >> DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM)))
> > >>>> +	if (txq->offloads & DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM)
> > >>>>  		flags |= EFX_TXQ_CKSUM_INNER_IPV4;
> > >>>>
> > >>>>  	if ((txq->offloads & DEV_TX_OFFLOAD_TCP_CKSUM) || @@
> -453,16
> > >> +443,7
> > >>>> @@ sfc_tx_qstart(struct sfc_adapter *sa, unsigned int sw_index)
> > >>>>  			flags |= EFX_TXQ_CKSUM_INNER_TCPUDP;
> > >>>>  	}
> > >>>>
> > >>>> -	/*
> > >>>> -	 * The absence of ETH_TXQ_FLAGS_IGNORE is associated
> with a
> > >> legacy
> > >>>> -	 * application. In turn, the absence of
> ETH_TXQ_FLAGS_NOXSUMTCP
> > >> is
> > >>>> -	 * associated specifically with a legacy application which
> expects
> > >>>> -	 * both TCP checksum offload and TSO to be enabled
> because the
> > >> legacy
> > >>>> -	 * API does not provide a dedicated mechanism to control
> TSO.
> > >>>> -	 */
> > >>>> -	if ((txq->offloads & DEV_TX_OFFLOAD_TCP_TSO) ||
> > >>>> -	    ((~txq->flags & ETH_TXQ_FLAGS_IGNORE) &&
> > >>>> -	     (~txq->flags & ETH_TXQ_FLAGS_NOXSUMTCP)))
> > >>>> +	if (txq->offloads & DEV_TX_OFFLOAD_TCP_TSO)
> > >>>>  		flags |= EFX_TXQ_FATSOV2;
> > >>>>
> > >>>>  	rc = efx_tx_qcreate(sa->nic, sw_index, 0, &txq->mem,
> > >>>
> > >>> net/sfc changes looks good.
> > >>> Plus 'struct sfc_txq -> flags' (drivers/net/sfc/sfc_tx.h) and
> > >>> 'struct sfc_dp_tx_qcreate_info -> flags'
> > >>> (drivers/net/sfc/sfc_dp_tx.h) should be removed since there are not
> used now.
> > >>>
> > >>> If finally rxmode.jumbo_frame is removed, it should removed from
> > >>> net/sfc as well (but compiler will help to find it in any case).
> > >>>
> > >>> After applying the patch:
> > >>> $ git grep ETH_TXQ_FLAGS
> > >>> drivers/net/fm10k/fm10k.h:#define FM10K_SIMPLE_TX_FLAG
> > >>> ((uint32_t)ETH_TXQ_FLAGS_NOMULTSEGS | \
> > >>> drivers/net/fm10k/fm10k.h:
> > >>> ETH_TXQ_FLAGS_NOOFFLOADS)
> > >>
> > >> Thanks, will remove this too.
> > >>
> > >>>
> > >>> In general I think that we should do it ASAP. Also it will
> > >>> guarantee that new PMDs do not use corresponding structure
> members etc.
> > >>
> > >> +1, +1
> > >