[xdp-hints] Re: [RFC net-next v4 2/8] xsk: add TX timestamp and TX checksum offload support

[Willem de Bruijn <willemdebruijn.kernel@gmail.com>]

Jesper Dangaard Brouer wrote:
> 
> 
> On 25/07/2023 22.58, Willem de Bruijn wrote:
> > Stanislav Fomichev wrote:
> >> This change actually defines the (initial) metadata layout
> >> that should be used by AF_XDP userspace (xsk_tx_metadata).
> >> The first field is flags which requests appropriate offloads,
> >> followed by the offload-specific fields. The supported per-device
> >> offloads are exported via netlink (new xsk-flags).
> >>
> >> The offloads themselves are still implemented in a bit of a
> >> framework-y fashion that's left from my initial kfunc attempt.
> >> I'm introducing new xsk_tx_metadata_ops which drivers are
> >> supposed to implement. The drivers are also supposed
> >> to call xsk_tx_metadata_request/xsk_tx_metadata_complete in
> >> the right places. Since xsk_tx_metadata_{request,_complete}
> >> are static inline, we don't incur any extra overhead doing
> >> indirect calls.
> >>
> >> The benefit of this scheme is as follows:
> >> - keeps all metadata layout parsing away from driver code
> >> - makes it easy to grep and see which drivers implement what
> >> - don't need any extra flags to maintain to keep track of that
> >>    offloads are implemented; if the callback is implemented - the offload
> >>    is supported (used by netlink reporting code)
> >>
> >> Two offloads are defined right now:
> >> 1. XDP_TX_METADATA_CHECKSUM: skb-style csum_start+csum_offset
> >> 2. XDP_TX_METADATA_TIMESTAMP: writes TX timestamp back into metadata
> >>     area upon completion (tx_timestamp field)
> >>
> >> The offloads are also implemented for copy mode:
> >> 1. Extra XDP_TX_METADATA_CHECKSUM_SW to trigger skb_checksum_help; this
> >>     might be useful as a reference implementation and for testing
> >> 2. XDP_TX_METADATA_TIMESTAMP writes SW timestamp from the skb
> >>     destructor (note I'm reusing hwtstamps to pass metadata pointer)
> >>
> >> The struct is forward-compatible and can be extended in the future
> >> by appending more fields.
> >>
> >> Signed-off-by: Stanislav Fomichev <sdf@google.com>
> >> ---
> >>   Documentation/netlink/specs/netdev.yaml | 19 ++++++++
> >>   include/linux/netdevice.h               | 27 +++++++++++
> >>   include/linux/skbuff.h                  |  5 ++-
> >>   include/net/xdp_sock.h                  | 60 +++++++++++++++++++++++++
> >>   include/net/xdp_sock_drv.h              | 13 ++++++
> >>   include/uapi/linux/if_xdp.h             | 35 +++++++++++++++
> >>   include/uapi/linux/netdev.h             | 15 +++++++
> >>   net/core/netdev-genl.c                  | 12 ++++-
> >>   net/xdp/xsk.c                           | 38 ++++++++++++++++
> >>   net/xdp/xsk_queue.h                     |  2 +-
> >>   tools/include/uapi/linux/if_xdp.h       | 50 ++++++++++++++++++---
> >>   11 files changed, 268 insertions(+), 8 deletions(-)
> >>
> >> diff --git a/Documentation/netlink/specs/netdev.yaml b/Documentation/netlink/specs/netdev.yaml
> >> index e41015310a6e..bf9c1cc32614 100644
> >> --- a/Documentation/netlink/specs/netdev.yaml
> >> +++ b/Documentation/netlink/specs/netdev.yaml
> >> @@ -42,6 +42,19 @@ name: netdev
> >>           doc:
> >>             This feature informs if netdev implements non-linear XDP buffer
> >>             support in ndo_xdp_xmit callback.
> >> +  -
> >> +    type: flags
> >> +    name: xsk-flags
> >> +    render-max: true
> >> +    entries:
> >> +      -
> >> +        name: tx-timestamp
> >> +        doc:
> >> +          HW timestamping egress packets is supported by the driver.
> >> +      -
> >> +        name: tx-checksum
> >> +        doc:
> >> +          L3 checksum HW offload is supported by the driver.
> >>   
> >>   attribute-sets:
> >>     -
> >> @@ -68,6 +81,12 @@ name: netdev
> >>           type: u32
> >>           checks:
> >>             min: 1
> >> +      -
> >> +        name: xsk-features
> >> +        doc: Bitmask of enabled AF_XDP features.
> >> +        type: u64
> >> +        enum: xsk-flags
> >> +        enum-as-flags: true
> >>   
> >>   operations:
> >>     list:
> >> diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
> >> index 11652e464f5d..8b40c80557aa 100644
> >> --- a/include/linux/netdevice.h
> >> +++ b/include/linux/netdevice.h
> >> @@ -1660,6 +1660,31 @@ struct xdp_metadata_ops {
> >>   			       enum xdp_rss_hash_type *rss_type);
> >>   };
> >>   
> >> +/*
> >> + * This structure defines the AF_XDP TX metadata hooks for network devices.
> >> + * The following hooks can be defined; unless noted otherwise, they are
> >> + * optional and can be filled with a null pointer.
> >> + *
> >> + * int (*tmo_request_timestamp)(void *priv)
> >> + *     This function is called when AF_XDP frame requested egress timestamp.
> >> + *
> >> + * int (*tmo_fill_timestamp)(void *priv)
> >> + *     This function is called when AF_XDP frame, that had requested
> >> + *     egress timestamp, received a completion. The hook needs to return
> >> + *     the actual HW timestamp.
> >> + *
> >> + * int (*tmo_request_timestamp)(u16 csum_start, u16 csum_offset, void *priv)
> > 
> > typo: tmo_request_checksum
> > 
> >> + *     This function is called when AF_XDP frame requested HW checksum
> >> + *     offload. csum_start indicates position where checksumming should start.
> >> + *     csum_offset indicates position where checksum should be stored.
> >> + *
> >> + */
> >> +struct xsk_tx_metadata_ops {
> >> +	void	(*tmo_request_timestamp)(void *priv);
> >> +	u64	(*tmo_fill_timestamp)(void *priv);
> >> +	void	(*tmo_request_checksum)(u16 csum_start, u16 csum_offset, void *priv);
> >> +};
> >> +
> >>   /**
> >>    * enum netdev_priv_flags - &struct net_device priv_flags
> >>    *
> >> @@ -1844,6 +1869,7 @@ enum netdev_ml_priv_type {
> >>    *	@netdev_ops:	Includes several pointers to callbacks,
> >>    *			if one wants to override the ndo_*() functions
> >>    *	@xdp_metadata_ops:	Includes pointers to XDP metadata callbacks.
> >> + *	@xsk_tx_metadata_ops:	Includes pointers to AF_XDP TX metadata callbacks.
> >>    *	@ethtool_ops:	Management operations
> >>    *	@l3mdev_ops:	Layer 3 master device operations
> >>    *	@ndisc_ops:	Includes callbacks for different IPv6 neighbour
> >> @@ -2100,6 +2126,7 @@ struct net_device {
> >>   	unsigned long long	priv_flags;
> >>   	const struct net_device_ops *netdev_ops;
> >>   	const struct xdp_metadata_ops *xdp_metadata_ops;
> >> +	const struct xsk_tx_metadata_ops *xsk_tx_metadata_ops;
> >>   	int			ifindex;
> >>   	unsigned short		gflags;
> >>   	unsigned short		hard_header_len;
> >> diff --git a/include/linux/skbuff.h b/include/linux/skbuff.h
> >> index faaba050f843..5febc1a5131e 100644
> >> --- a/include/linux/skbuff.h
> >> +++ b/include/linux/skbuff.h
> >> @@ -581,7 +581,10 @@ struct skb_shared_info {
> >>   	/* Warning: this field is not always filled in (UFO)! */
> >>   	unsigned short	gso_segs;
> >>   	struct sk_buff	*frag_list;
> >> -	struct skb_shared_hwtstamps hwtstamps;
> >> +	union {
> >> +		struct skb_shared_hwtstamps hwtstamps;
> >> +		struct xsk_tx_metadata *xsk_meta;
> >> +	};
> >>   	unsigned int	gso_type;
> >>   	u32		tskey;
> >>   
> >> diff --git a/include/net/xdp_sock.h b/include/net/xdp_sock.h
> >> index 467b9fb56827..288fa58c4665 100644
> >> --- a/include/net/xdp_sock.h
> >> +++ b/include/net/xdp_sock.h
> >> @@ -90,6 +90,54 @@ int xsk_generic_rcv(struct xdp_sock *xs, struct xdp_buff *xdp);
> >>   int __xsk_map_redirect(struct xdp_sock *xs, struct xdp_buff *xdp);
> >>   void __xsk_map_flush(void);
> >>   
> >> +/**
> >> + *  xsk_tx_metadata_request - Evaluate AF_XDP TX metadata at submission
> >> + *  and call appropriate xsk_tx_metadata_ops operation.
> >> + *  @meta: pointer to AF_XDP metadata area
> >> + *  @ops: pointer to struct xsk_tx_metadata_ops
> >> + *  @priv: pointer to driver-private aread
> >> + *
> >> + *  This function should be called by the networking device when
> >> + *  it prepares AF_XDP egress packet.
> >> + */
> >> +static inline void xsk_tx_metadata_request(const struct xsk_tx_metadata *meta,
> >> +					   const struct xsk_tx_metadata_ops *ops,
> >> +					   void *priv)
> >> +{
> >> +	if (!meta)
> >> +		return;
> >> +
> >> +	if (ops->tmo_request_timestamp)
> >> +		if (meta->flags & XDP_TX_METADATA_TIMESTAMP)
> >> +			ops->tmo_request_timestamp(priv);
> >> +
> >> +	if (ops->tmo_request_checksum)
> >> +		if (meta->flags & XDP_TX_METADATA_CHECKSUM)
> >> +			ops->tmo_request_checksum(meta->csum_start, meta->csum_offset, priv);
> >> +}
> > 
> > Might be cheaper to test the flag in the hot cacheline before
> > dereferencing ops?
> > 
> 
> I was thinking the same thing, but I was wrong, see below, as these ops
> deref's are optimized out by the compiler.
> 
> > Also, just add these functions to net_device_ops directly,
> > rather than dereferencing another pointer to xsk_tx_metadata_ops?
> > 
> 
> After the ASM/objdump discussion[1] in this thread, I think Stanislav's
> code approach here is actually optimal. Because when the 'const' ops are
> defined locally in the same file that use/does the inlining of
> xsk_tx_metadata_request() then the compilers (both llvm and gcc) are
> smart enough to inline and do dead-code elimination.  E.g. I noticed
> that generated ASM code eliminated `if (ops->tmo_request_timestamp)`
> code path in mlx5 as its not implemented.
> 
>   [1] 
> https://lore.kernel.org/all/cce9db50-8c9d-ea97-cb88-171fa46cc064@redhat.com/
> 
> --Jesper
> 

Agreed. That is pretty slick :-) I had entirely missed that.