RE: [v4 1/3] ethdev: introduce protocol type based header split

["Ding, Xuan" <xuan.ding@intel.com>]

Hi Andrew,

> -----Original Message-----
> From: Andrew Rybchenko <andrew.rybchenko@oktetlabs.ru>
> Sent: Wednesday, April 20, 2022 11:48 PM
> To: Ding, Xuan <xuan.ding@intel.com>; Wu, WenxuanX
> <wenxuanx.wu@intel.com>; thomas@monjalon.net; Li, Xiaoyun
> <xiaoyun.li@intel.com>; Singh, Aman Deep <aman.deep.singh@intel.com>;
> Zhang, Yuying <yuying.zhang@intel.com>; Zhang, Qi Z <qi.z.zhang@intel.com>
> Cc: dev@dpdk.org; stephen@networkplumber.org;
> mb@smartsharesystems.com; viacheslavo@nvidia.com; Yu, Ping
> <ping.yu@intel.com>; Wang, YuanX <yuanx.wang@intel.com>;
> david.marchand@redhat.com; Ferruh Yigit <ferruhy@xilinx.com>
> Subject: Re: [v4 1/3] ethdev: introduce protocol type based header split
> 
> Hi Xuan,
> 
> On 4/12/22 19:15, Ding, Xuan wrote:
> > Hi Andrew,
> >
> >> -----Original Message-----
> >> From: Andrew Rybchenko <andrew.rybchenko@oktetlabs.ru>
> >> Sent: Thursday, April 7, 2022 6:48 PM
> >> To: Wu, WenxuanX <wenxuanx.wu@intel.com>; thomas@monjalon.net;
> Li,
> >> Xiaoyun <xiaoyun.li@intel.com>; Singh, Aman Deep
> >> <aman.deep.singh@intel.com>; Zhang, Yuying <yuying.zhang@intel.com>;
> >> Zhang, Qi Z <qi.z.zhang@intel.com>
> >> Cc: dev@dpdk.org; stephen@networkplumber.org;
> >> mb@smartsharesystems.com; viacheslavo@nvidia.com; Yu, Ping
> >> <ping.yu@intel.com>; Ding, Xuan <xuan.ding@intel.com>; Wang, YuanX
> >> <yuanx.wang@intel.com>; david.marchand@redhat.com; Ferruh Yigit
> >> <ferruhy@xilinx.com>
> >> Subject: Re: [v4 1/3] ethdev: introduce protocol type based header
> >> split
> >>
> >> On 4/2/22 13:41, wenxuanx.wu@intel.com wrote:
> >>> From: Xuan Ding <xuan.ding@intel.com>
> >>>
> >>> Header split consists of splitting a received packet into two
> >>> separate regions based on the packet content. The split happens
> >>> after the packet header and before the packet payload. Splitting is
> >>> usually between the packet header that can be posted to a dedicated
> >>> buffer and the packet payload that can be posted to a different buffer.
> >>>
> >>> Currently, Rx buffer split supports length and offset based packet split.
> >>> Although header split is a subset of buffer split, configuring
> >>> buffer split based on length is not suitable for NICs that do split
> >>> based on header protocol types. Because tunneling makes the
> >>> conversion from length to protocol type impossible.
> >>>
> >>> This patch extends the current buffer split to support protocol type
> >>> and offset based header split. A new proto field is introduced in
> >>> the rte_eth_rxseg_split structure reserved field to specify header
> >>> protocol type. With Rx offload flag
> RTE_ETH_RX_OFFLOAD_HEADER_SPLIT
> >>> enabled and protocol type configured, PMD will split the ingress
> >>> packets into two separate regions. Currently, both inner and outer
> >>> L2/L3/L4 level header split can be supported.
> >>
> >> RTE_ETH_RX_OFFLOAD_HEADER_SPLIT offload was introduced some time
> ago
> >> to substitute bit-field header_split in struct rte_eth_rxmode. It
> >> allows to enable header split offload with the header size controlled
> >> using split_hdr_size in the same structure.
> >>
> >> Right now I see no single PMD which actually supports
> >> RTE_ETH_RX_OFFLOAD_HEADER_SPLIT with above definition.
> >> Many examples and test apps initialize the field to 0 explicitly. The
> >> most of drivers simply ignore split_hdr_size since the offload is not
> >> advertised, but some double-check that its value is 0.
> >>
> >> I think that it means that the field should be removed on the next
> >> LTS, and I'd say, together with the RTE_ETH_RX_OFFLOAD_HEADER_SPLIT
> offload bit.
> >>
> >> We should not redefine the offload meaning.
> >
> > Yes, you are right. No single PMD supports
> RTE_ETH_RX_OFFLOAD_HEADER_SPLIT now.
> > Previously, I used this flag is to distinguish buffer split and header split.
> > The former supports multi-segments split by length and offset.
> 
> offset is misleading here, since split offset is derived from segment lengths.
> Offset specified in segments is a different thing.

Yes, the length defines the segment length, and the offset defines the data  offset in mbuf.
The usage of length and offset are explained in the comments, but it is somewhat misleading
just from name.

> 
> > The later supports two segments split by proto and offset.
> > At this level, header split is a subset of buffer split.
> 
> IMHO, generic definition of the header split should not limit it by just two
> segments.

Does the header split here refer to the traditional header split?
If so, since you mentioned before we should not redefine the offload meaning,
I will use protocol and mbuf_offset based buffer split in next version.

It is worth noting that the purpose of specifying the split location by protocol is
to divide a packet into two segments. If you want to divide into multiple segments,
it should still be specified by length.

> 
> >
> > Since we shouldn't redefine the meaning of this offload, I will use
> > the RTE_ETH_RX_OFFLOAD_BUFFER_SPLIT flag.
> > The existence of tunnel needs to define a proto field in buffer split,
> > because some PMDs do not support split based on length and offset.
> 
> Not sure that I fully understand, but I'm looking forward to review v5.

Thanks for your comments, I will send a v5, including these main changes:
1. Use protocol and mbuf_offset based buffer split instead of header split.
2. Use RTE_PTYPE* instead of enum rte_eth_rx_header_split_protocol_type.
3. Improve the description of rte_eth_rxseg_split.proto.

Your comments are welcomed. 😊

> 
> >>>
> >>> For example, let's suppose we configured the Rx queue with the
> >>> following segments:
> >>>       seg0 - pool0, off0=2B
> >>>       seg1 - pool1, off1=128B
> >>
> >> Corresponding feature is named Rx buffer split.
> >> Does it mean that protocol type based header split requires Rx buffer
> >> split feature to be supported?
> >
> > Protocol type based header split does not requires Rx buffer split.
> > In previous design, the header split and buffer split are exclusive.
> > Because we only configure one split offload for one RX queue.
> >
> >>
> >>>
> >>> With header split type configured with RTE_ETH_RX_HEADER_SPLIT_UDP,
> >>> the packet consists of MAC_IP_UDP_PAYLOAD will be split like following:
> >>>       seg0 - udp header @ RTE_PKTMBUF_HEADROOM + 2 in mbuf from
> >> pool0
> >>>       seg1 - payload @ 128 in mbuf from pool1
> >>
> >> Is it always outermost UDP? Does it require both UDP over IPv4 and
> >> UDP over
> >> IPv6 to be supported? What will happen if only one is supported? How
> >> application can find out which protocol stack are supported?
> >
> > Both inner and outer UDP are considered.
> > Current design does not distinguish UDP over IPv4 or IPv6.
> > If we want to support granularity like only IPv4 or IPv6 supported,
> > user need add more configurations.

Thanks for your suggestion.
I will improve the documents about the usage of proto based buffer split.

> 
> You should make it clear for application how to use it.
> What happens if unsupported packet is received on an RxQ configured to do
> header split?


In fact, the buffer split and rte_flow are used in combination. It is expected that
the received packets will be steering to the RXQ configured with buffer split
offload. So there won't be unsupported packet received on an RXQ.

> 
> >
> > If application want to find out which protocol stack is supported, one
> > way I think is to expose the protocol stack supported by the driver through
> dev_info.
> > Any thoughts are welcomed :)
> 
> dev_info is nice, but very heavily overloaded. We can start from dev_info
> and understand if it should be factored out to a separate API or it is OK to
> have it in dev_info if it just few simple fields.

I'm also thinking exposing the protocol stack by dev_info is heavy.
We can configure all the protocol stack, and driver supports
part of the stacks. For protocols driver not supported, driver can returns the error.
What do you think of this design?

Regards,
Xuan

> 
> >>
> >>>
> >>> The memory attributes for the split parts may differ either - for
> >>> example the mempool0 and mempool1 belong to dpdk memory and
> >> external
> >>> memory, respectively.
> >>>
> >>> Signed-off-by: Xuan Ding <xuan.ding@intel.com>
> >>> Signed-off-by: Yuan Wang <yuanx.wang@intel.com>
> >>> Signed-off-by: Wenxuan Wu <wenxuanx.wu@intel.com>
> >>> Reviewed-by: Qi Zhang <qi.z.zhang@intel.com>
> >>> ---
> >>>    lib/ethdev/rte_ethdev.c | 34 ++++++++++++++++++++++-------
> >>>    lib/ethdev/rte_ethdev.h | 48
> >> +++++++++++++++++++++++++++++++++++++++--
> >>>    2 files changed, 72 insertions(+), 10 deletions(-)
> >>>
> >>> diff --git a/lib/ethdev/rte_ethdev.c b/lib/ethdev/rte_ethdev.c index
> >>> 29a3d80466..29adcdc2f0 100644
> >>> --- a/lib/ethdev/rte_ethdev.c
> >>> +++ b/lib/ethdev/rte_ethdev.c
> >>> @@ -1661,6 +1661,7 @@ rte_eth_rx_queue_check_split(const struct
> >> rte_eth_rxseg_split *rx_seg,
> >>>    		struct rte_mempool *mpl = rx_seg[seg_idx].mp;
> >>>    		uint32_t length = rx_seg[seg_idx].length;
> >>>    		uint32_t offset = rx_seg[seg_idx].offset;
> >>> +		uint16_t proto = rx_seg[seg_idx].proto;
> >>>
> >>>    		if (mpl == NULL) {
> >>>    			RTE_ETHDEV_LOG(ERR, "null mempool pointer\n");
> >> @@ -1694,13
> >>> +1695,29 @@ rte_eth_rx_queue_check_split(const struct
> >> rte_eth_rxseg_split *rx_seg,
> >>>    		}
> >>>    		offset += seg_idx != 0 ? 0 : RTE_PKTMBUF_HEADROOM;
> >>>    		*mbp_buf_size = rte_pktmbuf_data_room_size(mpl);
> >>> -		length = length != 0 ? length : *mbp_buf_size;
> >>> -		if (*mbp_buf_size < length + offset) {
> >>> -			RTE_ETHDEV_LOG(ERR,
> >>> -				       "%s mbuf_data_room_size %u < %u
> >> (segment length=%u + segment offset=%u)\n",
> >>> -				       mpl->name, *mbp_buf_size,
> >>> -				       length + offset, length, offset);
> >>> -			return -EINVAL;
> >>> +		if (proto == RTE_ETH_RX_HEADER_SPLIT_NONE) {
> >>> +			/* Check buffer split. */
> >>> +			length = length != 0 ? length : *mbp_buf_size;
> >>> +			if (*mbp_buf_size < length + offset) {
> >>> +				RTE_ETHDEV_LOG(ERR,
> >>> +					"%s mbuf_data_room_size %u < %u
> >> (segment length=%u + segment offset=%u)\n",
> >>> +					mpl->name, *mbp_buf_size,
> >>> +					length + offset, length, offset);
> >>> +				return -EINVAL;
> >>> +			}
> >>> +		} else {
> >>> +			/* Check header split. */
> >>> +			if (length != 0) {
> >>> +				RTE_ETHDEV_LOG(ERR, "segment length
> >> should be set to zero in header split\n");
> >>> +				return -EINVAL;
> >>> +			}
> >>> +			if (*mbp_buf_size < offset) {
> >>> +				RTE_ETHDEV_LOG(ERR,
> >>> +					"%s mbuf_data_room_size %u < %u
> >> segment offset)\n",
> >>> +					mpl->name, *mbp_buf_size,
> >>> +					offset);
> >>> +				return -EINVAL;
> >>> +			}
> >>>    		}
> >>>    	}
> >>>    	return 0;
> >>> @@ -1778,7 +1795,8 @@ rte_eth_rx_queue_setup(uint16_t port_id,
> >> uint16_t rx_queue_id,
> >>>    		rx_seg = (const struct rte_eth_rxseg_split *)rx_conf->rx_seg;
> >>>    		n_seg = rx_conf->rx_nseg;
> >>>
> >>> -		if (rx_conf->offloads & RTE_ETH_RX_OFFLOAD_BUFFER_SPLIT)
> >> {
> >>> +		if (rx_conf->offloads & RTE_ETH_RX_OFFLOAD_BUFFER_SPLIT
> >> ||
> >>> +			rx_conf->offloads &
> >> RTE_ETH_RX_OFFLOAD_HEADER_SPLIT) {
> >>>    			ret = rte_eth_rx_queue_check_split(rx_seg, n_seg,
> >>>    							   &mbp_buf_size,
> >>>    							   &dev_info);
> >>> diff --git a/lib/ethdev/rte_ethdev.h b/lib/ethdev/rte_ethdev.h index
> >>> 04cff8ee10..e8371b98ed 100644
> >>> --- a/lib/ethdev/rte_ethdev.h
> >>> +++ b/lib/ethdev/rte_ethdev.h
> >>> @@ -1197,12 +1197,31 @@ struct rte_eth_txmode {
> >>>     *     - pool from the last valid element
> >>>     *     - the buffer size from this pool
> >>>     *     - zero offset
> >>> + *
> >>> + * Header split is a subset of buffer split. The split happens
> >>> + after the
> >>> + * packet header and before the packet payload. For PMDs that do
> >>> + not
> >>> + * support header split configuration by length, the location of
> >>> + the split
> >>> + * needs to be specified by the header protocol type. While for
> >>> + buffer split,
> >>> + * this field should not be configured.
> >>> + *
> >>> + * If RTE_ETH_RX_OFFLOAD_HEADER_SPLIT flag is set in offloads
> >>> + field,
> >>> + * the PMD will split the received packets into two separate regions:
> >>> + * - The header buffer will be allocated from the memory pool,
> >>> + *   specified in the first array element, the second buffer, from the
> >>> + *   pool in the second element.
> >>> + *
> >>> + * - The lengths do not need to be configured in header split.
> >>> + *
> >>> + * - The offsets from the segment description elements specify
> >>> + *   the data offset from the buffer beginning except the first mbuf.
> >>> + *   The first segment offset is added with RTE_PKTMBUF_HEADROOM.
> >>>     */
> >>>    struct rte_eth_rxseg_split {
> >>>    	struct rte_mempool *mp; /**< Memory pool to allocate segment
> >> from. */
> >>>    	uint16_t length; /**< Segment data length, configures split point. */
> >>>    	uint16_t offset; /**< Data offset from beginning of mbuf data buffer.
> >> */
> >>> -	uint32_t reserved; /**< Reserved field. */
> >>> +	uint16_t proto; /**< header protocol type, configures header split
> >>> +point. */
> >>
> >> I realize that you don't want to use here enum defined above to save
> >> some reserved space, but description must refer to the enum
> >> rte_eth_rx_header_split_protocol_type.
> >
> > Thanks for your suggestion, will fix it in next version.
> >
> >>
> >>> +	uint16_t reserved; /**< Reserved field. */
> >>
> >> As far as I can see the structure is experimental. So, it should not
> >> be the problem to extend it, but it is a really good question raised
> >> by Stephen in RFC
> >> v1 discussion.
> >> Shouldn't we require that all reserved fields are initialized to zero
> >> and ignored on processing? Frankly speaking I always thought so, but
> >> failed to find the place were it is documented.
> >
> > Yes, it can be documented. By default is should be zero, and we can
> > configure it to enable protocol type based buffer split.
> >
> >>
> >> @Thomas, @David, @Ferruh?
> >>
> >>>    };
> >>>
> >>>    /**
> >>> @@ -1212,7 +1231,7 @@ struct rte_eth_rxseg_split {
> >>>     * A common structure used to describe Rx packet segment properties.
> >>>     */
> >>>    union rte_eth_rxseg {
> >>> -	/* The settings for buffer split offload. */
> >>> +	/* The settings for buffer split and header split offload. */
> >>>    	struct rte_eth_rxseg_split split;
> >>>    	/* The other features settings should be added here. */
> >>>    };
> >>> @@ -1664,6 +1683,31 @@ struct rte_eth_conf {
> >>>    			     RTE_ETH_RX_OFFLOAD_QINQ_STRIP)
> >>>    #define DEV_RX_OFFLOAD_VLAN
> >> RTE_DEPRECATED(DEV_RX_OFFLOAD_VLAN)
> >>> RTE_ETH_RX_OFFLOAD_VLAN
> >>>
> >>> +/**
> >>> + * @warning
> >>> + * @b EXPERIMENTAL: this enum may change without prior notice.
> >>> + * This enum indicates the header split protocol type  */ enum
> >>> +rte_eth_rx_header_split_protocol_type {
> >>> +	RTE_ETH_RX_HEADER_SPLIT_NONE = 0,
> >>> +	RTE_ETH_RX_HEADER_SPLIT_MAC,
> >>> +	RTE_ETH_RX_HEADER_SPLIT_IPV4,
> >>> +	RTE_ETH_RX_HEADER_SPLIT_IPV6,
> >>> +	RTE_ETH_RX_HEADER_SPLIT_L3,
> >>> +	RTE_ETH_RX_HEADER_SPLIT_TCP,
> >>> +	RTE_ETH_RX_HEADER_SPLIT_UDP,
> >>> +	RTE_ETH_RX_HEADER_SPLIT_SCTP,
> >>> +	RTE_ETH_RX_HEADER_SPLIT_L4,
> >>> +	RTE_ETH_RX_HEADER_SPLIT_INNER_MAC,
> >>> +	RTE_ETH_RX_HEADER_SPLIT_INNER_IPV4,
> >>> +	RTE_ETH_RX_HEADER_SPLIT_INNER_IPV6,
> >>> +	RTE_ETH_RX_HEADER_SPLIT_INNER_L3,
> >>> +	RTE_ETH_RX_HEADER_SPLIT_INNER_TCP,
> >>> +	RTE_ETH_RX_HEADER_SPLIT_INNER_UDP,
> >>> +	RTE_ETH_RX_HEADER_SPLIT_INNER_SCTP,
> >>> +	RTE_ETH_RX_HEADER_SPLIT_INNER_L4,
> >>
> >> Enumeration members should be documented. See my question in the
> >> patch description.
> >
> > Thanks for your detailed comments, questions are answered accordingly.
> >
> > Best Regards,
> > Xuan
> >
> >>
> >>> +};
> >>> +
> >>>    /*
> >>>     * If new Rx offload capabilities are defined, they also must be
> >>>     * mentioned in rte_rx_offload_names in rte_ethdev.c file.
> >