[dpdk-dev] [RFC PATCH 3/3] net/pcap: support hardware Tx timestamps

[Olivier Matz]

On Thu, Jun 25, 2020 at 03:01:19PM -0400, Vivien Didelot wrote:
> When hardware timestamping is enabled on Rx path, system time should
> no longer be used to calculate the timestamps when dumping packets.
> 
> Instead, use the value stored by the driver in mbuf->timestamp
> and assume it is already converted to nanoseconds (otherwise the
> application may edit the packet headers itself afterwards).
> 
> Signed-off-by: Vivien Didelot <vivien.didelot at gmail.com>
> Signed-off-by: Patrick Keroulas <patrick.keroulas at radio-canada.ca>
> ---
>  doc/guides/rel_notes/release_20_08.rst |  1 +
>  drivers/net/pcap/rte_eth_pcap.c        | 30 +++++++++++++++-----------
>  2 files changed, 18 insertions(+), 13 deletions(-)
> 
> diff --git a/doc/guides/rel_notes/release_20_08.rst b/doc/guides/rel_notes/release_20_08.rst
> index a67015519..cd1ca987f 100644
> --- a/doc/guides/rel_notes/release_20_08.rst
> +++ b/doc/guides/rel_notes/release_20_08.rst
> @@ -61,6 +61,7 @@ New Features
>    Updated PCAP driver with new features and improvements, including:
>  
>    * Support software Tx nanosecond timestamps precision.
> +  * Support hardware Tx timestamps.
>  
>  * **Updated Mellanox mlx5 driver.**
>  
> diff --git a/drivers/net/pcap/rte_eth_pcap.c b/drivers/net/pcap/rte_eth_pcap.c
> index 13a3d0ac7..3d80b699b 100644
> --- a/drivers/net/pcap/rte_eth_pcap.c
> +++ b/drivers/net/pcap/rte_eth_pcap.c
> @@ -290,19 +290,23 @@ eth_null_rx(void *queue __rte_unused,
>  #define NSEC_PER_SEC	1000000000L
>  
>  static inline void
> -calculate_timestamp(struct timeval *ts) {
> -	uint64_t cycles;
> -	struct timeval cur_time;
> +calculate_timestamp(const struct rte_mbuf *mbuf, struct timeval *ts) {
> +	if (mbuf->ol_flags & PKT_RX_TIMESTAMP) {
> +		ts->tv_sec = mbuf->timestamp / NSEC_PER_SEC;
> +		ts->tv_usec = mbuf->timestamp % NSEC_PER_SEC;
> +	} else {
> +		uint64_t cycles = rte_get_timer_cycles() - start_cycles;
> +		struct timeval cur_time = {
> +			.tv_sec = cycles / hz,
> +			.tv_usec = (cycles % hz) * NSEC_PER_SEC / hz,
> +		};
>  
> -	cycles = rte_get_timer_cycles() - start_cycles;
> -	cur_time.tv_sec = cycles / hz;
> -	cur_time.tv_usec = (cycles % hz) * NSEC_PER_SEC / hz;
> -
> -	ts->tv_sec = start_time.tv_sec + cur_time.tv_sec;
> -	ts->tv_usec = start_time.tv_usec + cur_time.tv_usec;
> -	if (ts->tv_usec >= NSEC_PER_SEC) {
> -		ts->tv_usec -= NSEC_PER_SEC;
> -		ts->tv_sec += 1;
> +		ts->tv_sec = start_time.tv_sec + cur_time.tv_sec;
> +		ts->tv_usec = start_time.tv_usec + cur_time.tv_usec;
> +		if (ts->tv_usec >= NSEC_PER_SEC) {
> +			ts->tv_usec -= NSEC_PER_SEC;
> +			ts->tv_sec += 1;
> +		}
>  	}
>  }
>  

I don't get why you expect that timestamp to be in nanoseconds.
The conversion is done in librte_pdump (in the previous patch),
but it won't work if the library is not used, right?

Out of curiosity, can you explain your motivation for using the hardware
timestamp? Is it faster? More accurate? (knowing it timestamps the Rx
operation, not the Tx)