Re: [RFC PATCH bpf-next v1 3/3] net: Add additional bit to support userspace timestamp type

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

Abhishek Chauhan wrote:
> tstamp_type can be real, mono or userspace timestamp.
> 
> This commit adds userspace timestamp and sets it if there is
> valid transmit_time available in socket coming from userspace.
> 
> To make the design scalable for future needs this commit bring in
> the change to extend the tstamp_type:1 to tstamp_type:2 to support
> userspace timestamp.
> 
> Link: https://lore.kernel.org/netdev/bc037db4-58bb-4861-ac31-a361a93841d3@linux.dev/
> Signed-off-by: Abhishek Chauhan <quic_abchauha@quicinc.com>
> ---
>  include/linux/skbuff.h | 19 +++++++++++++++++--
>  net/ipv4/ip_output.c   |  2 +-
>  net/ipv4/raw.c         |  2 +-
>  net/ipv6/ip6_output.c  |  2 +-
>  net/ipv6/raw.c         |  2 +-
>  net/packet/af_packet.c |  6 +++---
>  6 files changed, 24 insertions(+), 9 deletions(-)
> 
> diff --git a/include/linux/skbuff.h b/include/linux/skbuff.h
> index 6160185f0fe0..2f91a8a2157a 100644
> --- a/include/linux/skbuff.h
> +++ b/include/linux/skbuff.h
> @@ -705,6 +705,9 @@ typedef unsigned char *sk_buff_data_t;
>  enum skb_tstamp_type {
>  	SKB_TSTAMP_TYPE_RX_REAL = 0,    /* A RX (receive) time in real */
>  	SKB_TSTAMP_TYPE_TX_MONO = 1,    /* A TX (delivery) time in mono */
> +	SKB_TSTAMP_TYPE_TX_USER = 2,    /* A TX (delivery) time and its clock
> +									 * is in skb->sk->sk_clockid.
> +									 */

Weird indentation?

More fundamentally: instead of defining a type TX_USER, can we use a
real clockid (e.g., CLOCK_TAI) based on skb->sk->sk_clockid? Rather
than store an id that means "go look at sk_clockid".

>  };
>  
>  /**
> @@ -830,6 +833,9 @@ enum skb_tstamp_type {
>   *		delivery_time in mono clock base (i.e. EDT).  Otherwise, the
>   *		skb->tstamp has the (rcv) timestamp at ingress and
>   *		delivery_time at egress.
> + *		delivery_time in mono clock base (i.e., EDT) or a clock base chosen
> + *		by SO_TXTIME. If zero, skb->tstamp has the (rcv) timestamp at
> + *		ingress.
>   *	@napi_id: id of the NAPI struct this skb came from
>   *	@sender_cpu: (aka @napi_id) source CPU in XPS
>   *	@alloc_cpu: CPU which did the skb allocation.
> @@ -960,7 +966,7 @@ struct sk_buff {
>  	/* private: */
>  	__u8			__mono_tc_offset[0];
>  	/* public: */
> -	__u8			tstamp_type:1;	/* See SKB_MONO_DELIVERY_TIME_MASK */
> +	__u8			tstamp_type:2;	/* See SKB_MONO_DELIVERY_TIME_MASK */
>  #ifdef CONFIG_NET_XGRESS
>  	__u8			tc_at_ingress:1;	/* See TC_AT_INGRESS_MASK */
>  	__u8			tc_skip_classify:1;

With pahole, does this have an effect on sk_buff layout?

> @@ -4274,7 +4280,16 @@ static inline void skb_set_delivery_time(struct sk_buff *skb, ktime_t kt,
>  					enum skb_tstamp_type tstamp_type)
>  {
>  	skb->tstamp = kt;
> -	skb->tstamp_type = kt && tstamp_type;
> +
> +	if (skb->tstamp_type)
> +		return;
> +

Why bail if a type is already set? And what if
skb->tstamp_type != tstamp_type? Should skb->tstamp then not be
written to (i.e., the test moved up), and perhaps a rate limited
warning.

> +	if (kt && tstamp_type == SKB_TSTAMP_TYPE_TX_MONO)
> +		skb->tstamp_type = SKB_TSTAMP_TYPE_TX_MONO;
> +
> +	if (kt && tstamp_type == SKB_TSTAMP_TYPE_TX_USER)
> +		skb->tstamp_type = SKB_TSTAMP_TYPE_TX_USER;

Simpler

    if (kt)
        skb->tstamp_type = tstamp_type;