Re: [PATCH net-next] octeontx2-af: return tsc value along with PTP clock

[Vadim Fedorenko <vadim.fedorenko@linux.dev>]

On 06/08/2026 17:16, nshettyj@marvell.com wrote:
> From: Harman Kalra <hkalra@marvell.com>
> 
> This patch updates the existing PTP_OP_GET_CLOCK mbox message to
> return the timestamp counter value, tsc (cntvct_el0 or pmccntr_el0)
> along with the PTP HI clock value.
> 
> In some debugging scenarios, a user might need to read the PTP HI
> clock value in the fastpath to know how many ticks have been spent
> since the reception of the packet (as the packet reception tick value
> is already appended to the packet by CGX). If the PTP_OP_GET_CLOCK
> mbox message is sent every time the user wants to record the PTP HI
> clock value, it will bring down performance to a great extent as mbox
> communication is a very expensive process.
> 
> To handle this, the PTP HI clock can be derived from the timestamp
> counter (tsc, which could be running at 100MHz or system freq) using
> two parameters: freq multiplier (ratio of frequencies of PTP HI clock
> and tsc) and clock delta (by how much tsc is lagging from PTP HI
> clock).
> 
> By returning both the PTP_HI value and the tsc value simultaneously
> via PTP_OP_GET_CLOCK, the consumer can calculate these parameters
> without being affected by mbox propagation delay:
>       freq_mult = (freq of PTP HI clock) / (freq of tsc counter)
>       clk_delta = (PTP_HI clock value / freq_mult) - (tsc val)
> 
> Now, whenever the user wants to know the PTP HI clock in the fastpath,
> it can be derived from the local tsc counter:
>       PTP_HI val = (tsc value + clk_delta) * freq_mult
> 
> Signed-off-by: Harman Kalra <hkalra@marvell.com>
> Signed-off-by: Nitin Shetty J <nshettyj@marvell.com>
> ---
>   .../net/ethernet/marvell/octeontx2/af/mbox.h  |  1 +
>   .../net/ethernet/marvell/octeontx2/af/ptp.c   | 26 ++++++++++++++++---
>   2 files changed, 23 insertions(+), 4 deletions(-)
> 
> diff --git a/drivers/net/ethernet/marvell/octeontx2/af/mbox.h b/drivers/net/ethernet/marvell/octeontx2/af/mbox.h
> index 73f743e4a83d..21c3d3d5018e 100644
> --- a/drivers/net/ethernet/marvell/octeontx2/af/mbox.h
> +++ b/drivers/net/ethernet/marvell/octeontx2/af/mbox.h
> @@ -2010,6 +2010,7 @@ struct ptp_req {
>   	struct mbox_msghdr hdr;
>   	u8 op;
>   	s64 scaled_ppm;
> +	u8 is_pmu;
>   	u64 thresh;
>   	u64 period;
>   	int pps_on;
> diff --git a/drivers/net/ethernet/marvell/octeontx2/af/ptp.c b/drivers/net/ethernet/marvell/octeontx2/af/ptp.c
> index 58e62e955554..0b66a86246b6 100644
> --- a/drivers/net/ethernet/marvell/octeontx2/af/ptp.c
> +++ b/drivers/net/ethernet/marvell/octeontx2/af/ptp.c
> @@ -356,10 +356,27 @@ static int ptp_adjfine(struct ptp *ptp, long scaled_ppm)
>   	return 0;
>   }
>   
> -static int ptp_get_clock(struct ptp *ptp, u64 *clk)
> +static inline u64 get_tsc(bool is_pmu)
>   {
> -	/* Return the current PTP clock */
> -	*clk = ptp->read_ptp_tstmp(ptp);
> +#if defined(CONFIG_ARM64)
> +	return is_pmu ? read_sysreg(pmccntr_el0) : read_sysreg(cntvct_el0);
> +#else
> +	return 0;
> +#endif
> +}
> +
> +static int ptp_get_clock(struct ptp *ptp, bool is_pmu, u64 *clk, u64 *tsc)
> +{
> +	u64 end, start;
> +	u8 retries = 0;
> +
> +	do {
> +		start = get_tsc(0);
> +		*tsc = get_tsc(is_pmu);
> +		*clk = ptp->read_ptp_tstmp(ptp);
> +		end = get_tsc(0);
> +		retries++;
> +	} while (((end - start) > 50) && retries < 5);

didn't fully get why do you need this cycle? what are you trying to
avoid?

>   
>   	return 0;
>   }
> @@ -636,7 +653,8 @@ int rvu_mbox_handler_ptp_op(struct rvu *rvu, struct ptp_req *req,
>   		err = ptp_adjfine(rvu->ptp, req->scaled_ppm);
>   		break;
>   	case PTP_OP_GET_CLOCK:
> -		err = ptp_get_clock(rvu->ptp, &rsp->clk);
> +		err = ptp_get_clock(rvu->ptp, req->is_pmu, &rsp->clk,
> +				    &rsp->tsc);
>   		break;
>   	case PTP_OP_GET_TSTMP:
>   		err = ptp_get_tstmp(rvu->ptp, &rsp->clk);