Re: [net PATCH] octeontx2-pf: Recalculate UDP checksum for ptp 1-step sync packet

[Paolo Abeni <pabeni@redhat.com>]

On Mon, 2023-02-20 at 17:50 +0530, Sai Krishna wrote:
> From: Geetha sowjanya <gakula@marvell.com>
> 
> When checksum offload is disabled in the driver via ethtool,
> the PTP 1-step sync packets contain incorrect checksum, since
> the stack calculates the checksum before driver updates
> PTP timestamp field in the packet. This results in PTP packets
> getting dropped at the other end. This patch fixes the issue by
> re-calculating the UDP checksum after updating PTP
> timestamp field in the driver.
> 
> Fixes: 2958d17a8984 ("octeontx2-pf: Add support for ptp 1-step mode on CN10K silicon")
> Signed-off-by: Geetha sowjanya <gakula@marvell.com>
> Signed-off-by: Hariprasad Kelam <hkelam@marvell.com>
> Signed-off-by: Sunil Kovvuri Goutham <sgoutham@marvell.com>
> Signed-off-by: Sai Krishna <saikrishnag@marvell.com>
> ---
>  .../marvell/octeontx2/nic/otx2_txrx.c         | 78 ++++++++++++++-----
>  1 file changed, 59 insertions(+), 19 deletions(-)
> 
> diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c
> index ef10aef3cda0..67345a3e2bba 100644
> --- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c
> +++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c
> @@ -10,6 +10,7 @@
>  #include <net/tso.h>
>  #include <linux/bpf.h>
>  #include <linux/bpf_trace.h>
> +#include <net/ip6_checksum.h>
>  
>  #include "otx2_reg.h"
>  #include "otx2_common.h"
> @@ -699,7 +700,7 @@ static void otx2_sqe_add_ext(struct otx2_nic *pfvf, struct otx2_snd_queue *sq,
>  
>  static void otx2_sqe_add_mem(struct otx2_snd_queue *sq, int *offset,
>  			     int alg, u64 iova, int ptp_offset,
> -			     u64 base_ns, int udp_csum)
> +			     u64 base_ns, bool udp_csum_crt)
>  {
>  	struct nix_sqe_mem_s *mem;
>  
> @@ -711,7 +712,7 @@ static void otx2_sqe_add_mem(struct otx2_snd_queue *sq, int *offset,
>  
>  	if (ptp_offset) {
>  		mem->start_offset = ptp_offset;
> -		mem->udp_csum_crt = udp_csum;
> +		mem->udp_csum_crt = !!udp_csum_crt;
>  		mem->base_ns = base_ns;
>  		mem->step_type = 1;
>  	}
> @@ -986,10 +987,11 @@ static bool otx2_validate_network_transport(struct sk_buff *skb)
>  	return false;
>  }
>  
> -static bool otx2_ptp_is_sync(struct sk_buff *skb, int *offset, int *udp_csum)
> +static bool otx2_ptp_is_sync(struct sk_buff *skb, int *offset, bool *udp_csum_crt)
>  {
>  	struct ethhdr *eth = (struct ethhdr *)(skb->data);
>  	u16 nix_offload_hlen = 0, inner_vhlen = 0;
> +	bool udp_hdr_present = false, is_sync;
>  	u8 *data = skb->data, *msgtype;
>  	__be16 proto = eth->h_proto;
>  	int network_depth = 0;
> @@ -1029,45 +1031,83 @@ static bool otx2_ptp_is_sync(struct sk_buff *skb, int *offset, int *udp_csum)
>  		if (!otx2_validate_network_transport(skb))
>  			return false;
>  
> -		*udp_csum = 1;
>  		*offset = nix_offload_hlen + skb_transport_offset(skb) +
>  			  sizeof(struct udphdr);
> +		udp_hdr_present = true;
> +
>  	}
>  
>  	msgtype = data + *offset;
> -
>  	/* Check PTP messageId is SYNC or not */
> -	return (*msgtype & 0xf) == 0;
> +	is_sync =  ((*msgtype & 0xf) == 0) ? true : false;

the above could be:

	is_sync = !(*msgtype & 0xf);

possibly more readable.

> +	if (is_sync) {
> +		if (udp_hdr_present)
> +			*udp_csum_crt = true;

or:
		*udp_csum_crt = udp_hdr_present;

that will make the code more compact.

> +	} else {
> +		*offset = 0;
> +	}
> +
> +	return is_sync;
>  }
>  
>  static void otx2_set_txtstamp(struct otx2_nic *pfvf, struct sk_buff *skb,
>  			      struct otx2_snd_queue *sq, int *offset)
>  {
> +	struct ethhdr	*eth = (struct ethhdr *)(skb->data);
>  	struct ptpv2_tstamp *origin_tstamp;
> -	int ptp_offset = 0, udp_csum = 0;
> +	bool udp_csum_crt = false;
> +	unsigned int udphoff;
>  	struct timespec64 ts;
> +	int ptp_offset = 0;
> +	__wsum skb_csum;
>  	u64 iova;
>  
>  	if (unlikely(!skb_shinfo(skb)->gso_size &&
>  		     (skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP))) {
> -		if (unlikely(pfvf->flags & OTX2_FLAG_PTP_ONESTEP_SYNC)) {
> -			if (otx2_ptp_is_sync(skb, &ptp_offset, &udp_csum)) {
> -				origin_tstamp = (struct ptpv2_tstamp *)
> -						((u8 *)skb->data + ptp_offset +
> -						 PTP_SYNC_SEC_OFFSET);
> -				ts = ns_to_timespec64(pfvf->ptp->tstamp);
> -				origin_tstamp->seconds_msb = htons((ts.tv_sec >> 32) & 0xffff);
> -				origin_tstamp->seconds_lsb = htonl(ts.tv_sec & 0xffffffff);
> -				origin_tstamp->nanoseconds = htonl(ts.tv_nsec);
> -				/* Point to correction field in PTP packet */
> -				ptp_offset += 8;
> +		if (unlikely(pfvf->flags & OTX2_FLAG_PTP_ONESTEP_SYNC &&
> +			     otx2_ptp_is_sync(skb, &ptp_offset, &udp_csum_crt))) {
> +			origin_tstamp = (struct ptpv2_tstamp *)
> +					((u8 *)skb->data + ptp_offset +
> +					 PTP_SYNC_SEC_OFFSET);
> +			ts = ns_to_timespec64(pfvf->ptp->tstamp);
> +			origin_tstamp->seconds_msb = htons((ts.tv_sec >> 32) & 0xffff);
> +			origin_tstamp->seconds_lsb = htonl(ts.tv_sec & 0xffffffff);
> +			origin_tstamp->nanoseconds = htonl(ts.tv_nsec);
> +			/* Point to correction field in PTP packet */
> +			ptp_offset += 8;
> +
> +			/* When user disables hw checksum, stack calculates the csum,
> +			 * but it does not cover ptp timestamp which is added later.
> +			 * Recalculate the checksum manually considering the timestamp.
> +			 */
> +			if (udp_csum_crt) {
> +				struct udphdr *uh = udp_hdr(skb);
> +
> +				if (skb->ip_summed != CHECKSUM_PARTIAL && uh->check != 0) {
> +					udphoff = skb_transport_offset(skb);
> +					uh->check = 0;
> +					skb_csum = skb_checksum(skb, udphoff, skb->len - udphoff,
> +								0);
> +					if (ntohs(eth->h_proto) == ETH_P_IPV6)
> +						uh->check = csum_ipv6_magic(&ipv6_hdr(skb)->saddr,
> +									    &ipv6_hdr(skb)->daddr,
> +									    skb->len - udphoff,
> +									    ipv6_hdr(skb)->nexthdr,
> +									    skb_csum);
> +					else
> +						uh->check = csum_tcpudp_magic(ip_hdr(skb)->saddr,
> +									      ip_hdr(skb)->daddr,
> +									      skb->len - udphoff,
> +									      IPPROTO_UDP,
> +									      skb_csum);

Have you considered incremental csum updates instead? Possibly the code
could be simpler and likely faster - not sure if the latter part is
relevant in this case.

Cheers,

Paolo