Message-ID: <6d7dd029-5e6e-4718-9d95-7008ab8f0131@intel.com>
Date: Tue, 30 Jul 2024 17:29:42 +0200
From: Alexander Lobakin <aleksander.lobakin@...el.com>
To: Mateusz Polchlopek <mateusz.polchlopek@...el.com>
CC: <intel-wired-lan@...ts.osuosl.org>, <netdev@...r.kernel.org>, Jacob Keller
	<jacob.e.keller@...el.com>, Wojciech Drewek <wojciech.drewek@...el.com>,
	Rahul Rameshbabu <rrameshbabu@...dia.com>, Sunil Goutham
	<sgoutham@...vell.com>, Simon Horman <horms@...nel.org>
Subject: Re: [Intel-wired-lan] [PATCH iwl-next v8 14/14] iavf: add support for
 Rx timestamps to hotpath

From: Mateusz Polchlopek <mateusz.polchlopek@...el.com>
Date: Tue, 30 Jul 2024 05:15:09 -0400

> From: Jacob Keller <jacob.e.keller@...el.com>
> 
> Add support for receive timestamps to the Rx hotpath. This support only
> works when using the flexible descriptor format, so make sure that we
> request this format by default if we have receive timestamp support
> available in the PTP capabilities.
> 
> In order to report the timestamps to userspace, we need to perform
> timestamp extension. The Rx descriptor does actually contain the "40
> bit" timestamp. However, upper 32 bits which contain nanoseconds are
> conveniently stored separately in the descriptor. We could extract the
> 32bits and lower 8 bits, then perform a bitwise OR to calculate the
> 40bit value. This makes no sense, because the timestamp extension
> algorithm would simply discard the lower 8 bits anyways.
> 
> Thus, implement timestamp extension as iavf_ptp_extend_32b_timestamp(),
> and extract and forward only the 32bits of nominal nanoseconds.
> 
> Signed-off-by: Jacob Keller <jacob.e.keller@...el.com>
> Reviewed-by: Wojciech Drewek <wojciech.drewek@...el.com>
> Reviewed-by: Rahul Rameshbabu <rrameshbabu@...dia.com>
> Reviewed-by: Sunil Goutham <sgoutham@...vell.com>
> Reviewed-by: Simon Horman <horms@...nel.org>
> Signed-off-by: Mateusz Polchlopek <mateusz.polchlopek@...el.com>
> ---
>  drivers/net/ethernet/intel/iavf/iavf_main.c |  9 +++
>  drivers/net/ethernet/intel/iavf/iavf_ptp.c  | 69 +++++++++++++++++++++
>  drivers/net/ethernet/intel/iavf/iavf_ptp.h  |  4 ++
>  drivers/net/ethernet/intel/iavf/iavf_txrx.c | 47 ++++++++++++++
>  drivers/net/ethernet/intel/iavf/iavf_type.h |  3 +
>  5 files changed, 132 insertions(+)
> 
> diff --git a/drivers/net/ethernet/intel/iavf/iavf_main.c b/drivers/net/ethernet/intel/iavf/iavf_main.c
> index 61720b27c8f1..03deb3e02279 100644
> --- a/drivers/net/ethernet/intel/iavf/iavf_main.c
> +++ b/drivers/net/ethernet/intel/iavf/iavf_main.c
> @@ -729,6 +729,15 @@ static u8 iavf_select_rx_desc_format(struct iavf_adapter *adapter)
>  	if (!RXDID_ALLOWED(adapter))
>  		return VIRTCHNL_RXDID_1_32B_BASE;
>  
> +	/* Rx timestamping requires the use of flexible NIC descriptors */
> +	if (iavf_ptp_cap_supported(adapter, VIRTCHNL_1588_PTP_CAP_RX_TSTAMP)) {
> +		if (supported_rxdids & BIT(VIRTCHNL_RXDID_2_FLEX_SQ_NIC))
> +			return VIRTCHNL_RXDID_2_FLEX_SQ_NIC;
> +
> +		dev_dbg(&adapter->pdev->dev,
> +			"Unable to negotiate flexible descriptor format.\n");

1. Remember about pci_*().
2. We usually don't put '.' to the end of kernel messages.

> +	}
> +
>  	/* Warn if the PF does not list support for the default legacy
>  	 * descriptor format. This shouldn't happen, as this is the format
>  	 * used if VIRTCHNL_VF_OFFLOAD_RX_FLEX_DESC is not supported. It is
> diff --git a/drivers/net/ethernet/intel/iavf/iavf_ptp.c b/drivers/net/ethernet/intel/iavf/iavf_ptp.c
> index 7754f4f24052..5fd17f8d1f36 100644
> --- a/drivers/net/ethernet/intel/iavf/iavf_ptp.c
> +++ b/drivers/net/ethernet/intel/iavf/iavf_ptp.c
> @@ -440,6 +440,9 @@ void iavf_ptp_release(struct iavf_adapter *adapter)
>  	}
>  	adapter->aq_required &= ~IAVF_FLAG_AQ_SEND_PTP_CMD;
>  	mutex_unlock(&adapter->ptp.aq_cmd_lock);
> +
> +	adapter->ptp.hwtstamp_config.rx_filter = HWTSTAMP_FILTER_NONE;
> +	iavf_ptp_disable_rx_tstamp(adapter);
>  }
>  
>  /**
> @@ -473,3 +476,69 @@ void iavf_ptp_process_caps(struct iavf_adapter *adapter)
>  		iavf_ptp_disable_rx_tstamp(adapter);
>  	}
>  }
> +
> +/**
> + * iavf_ptp_extend_32b_timestamp - Convert a 32b nanoseconds timestamp to 64b
> + * nanoseconds
> + * @cached_phc_time: recently cached copy of PHC time
> + * @in_tstamp: Ingress/egress 32b nanoseconds timestamp value
> + *
> + * Hardware captures timestamps which contain only 32 bits of nominal
> + * nanoseconds, as opposed to the 64bit timestamps that the stack expects.
> + *
> + * Extend the 32bit nanosecond timestamp using the following algorithm and
> + * assumptions:
> + *
> + * 1) have a recently cached copy of the PHC time
> + * 2) assume that the in_tstamp was captured 2^31 nanoseconds (~2.1
> + *    seconds) before or after the PHC time was captured.
> + * 3) calculate the delta between the cached time and the timestamp
> + * 4) if the delta is smaller than 2^31 nanoseconds, then the timestamp was
> + *    captured after the PHC time. In this case, the full timestamp is just
> + *    the cached PHC time plus the delta.
> + * 5) otherwise, if the delta is larger than 2^31 nanoseconds, then the
> + *    timestamp was captured *before* the PHC time, i.e. because the PHC
> + *    cache was updated after the timestamp was captured by hardware. In this
> + *    case, the full timestamp is the cached time minus the inverse delta.
> + *
> + * This algorithm works even if the PHC time was updated after a Tx timestamp
> + * was requested, but before the Tx timestamp event was reported from
> + * hardware.
> + *
> + * This calculation primarily relies on keeping the cached PHC time up to
> + * date. If the timestamp was captured more than 2^31 nanoseconds after the
> + * PHC time, it is possible that the lower 32bits of PHC time have
> + * overflowed more than once, and we might generate an incorrect timestamp.
> + *
> + * This is prevented by (a) periodically updating the cached PHC time once
> + * a second, and (b) discarding any Tx timestamp packet if it has waited for
> + * a timestamp for more than one second.
> + *
> + * Return: extended timestamp (to 64b)

...but here you actually need a period :D

> + */
> +u64 iavf_ptp_extend_32b_timestamp(u64 cached_phc_time, u32 in_tstamp)
> +{
> +	const u64 mask = GENMASK_ULL(31, 0);
> +	u32 delta;
> +	u64 ns;
> +
> +	/* Calculate the delta between the lower 32bits of the cached PHC
> +	 * time and the in_tstamp value
> +	 */
> +	delta = (in_tstamp - (u32)(cached_phc_time & mask));

`cached_phc_time & GENMASK_ULL(31, 0)` == lower_32_bits(cached_phc_time)

> +
> +	/* Do not assume that the in_tstamp is always more recent than the
> +	 * cached PHC time. If the delta is large, it indicates that the
> +	 * in_tstamp was taken in the past, and should be converted
> +	 * forward.
> +	 */
> +	if (delta > (mask / 2)) {

1. `mask / 2` == S32_MAX.
2. Parenthesis are redundant here.

> +		/* reverse the delta calculation here */
> +		delta = ((u32)(cached_phc_time & mask) - in_tstamp);

Here as well.

> +		ns = cached_phc_time - delta;
> +	} else {
> +		ns = cached_phc_time + delta;
> +	}

1. (u32) casts are not needed since your mask ends with bit 31.
2. (cached_phc_time & mask) is used two times. Perhaps you needed

	u32 low = lower_32_bits(cached_phc_time);

	delta = in_tstamp - low;
	if (delta > S32_MAX)
		ns = cached_phc_time - (low - in_tstamp);
	else
		ns = cached_phc_time + delta;


> +
> +	return ns;
> +}

[...]

> diff --git a/drivers/net/ethernet/intel/iavf/iavf_txrx.c b/drivers/net/ethernet/intel/iavf/iavf_txrx.c
> index 997fd0d520a9..8d74549c3535 100644
> --- a/drivers/net/ethernet/intel/iavf/iavf_txrx.c
> +++ b/drivers/net/ethernet/intel/iavf/iavf_txrx.c
> @@ -1085,6 +1085,52 @@ static void iavf_flex_rx_hash(const struct iavf_ring *ring,
>  	}
>  }
>  
> +/**
> + * iavf_flex_rx_tstamp - Capture Rx timestamp from the descriptor
> + * @rx_ring: descriptor ring
> + * @rx_desc: specific descriptor
> + * @skb: skb currently being received
> + *
> + * Read the Rx timestamp value from the descriptor and pass it to the stack.
> + *
> + * This function only operates on the VIRTCHNL_RXDID_2_FLEX_SQ_NIC flexible
> + * descriptor writeback format.
> + */
> +static void iavf_flex_rx_tstamp(const struct iavf_ring *rx_ring,
> +				const struct iavf_rx_desc *rx_desc,
> +				struct sk_buff *skb)
> +{
> +	struct skb_shared_hwtstamps *skb_tstamps;
> +	struct iavf_adapter *adapter;
> +	__le64 qw2 = rx_desc->qw2;
> +	__le64 qw3 = rx_desc->qw3;
> +	bool tstamp_valid;
> +	u32 tstamp;
> +	u64 ns;
> +
> +	/* Skip processing if timestamps aren't enabled */
> +	if (!(rx_ring->flags & IAVF_TXRX_FLAGS_HW_TSTAMP))
> +		return;
> +
> +	/* Check if this Rx descriptor has a valid timestamp */
> +	tstamp_valid = le64_get_bits(qw2, IAVF_PTP_40B_TSTAMP_VALID);
> +	if (!tstamp_valid)
> +		return;

Read/convert descriptor qwords when you actually need them.

	if (!(rx_ring->flags & HW_TSTAMP))
		return;

	if (!le64_get_bits(rx_desc->qw2, TSTAMP_VALID))
		return;

	tstamp = le64_get_bits(rx_desc->qw3, ...

> +
> +	adapter = netdev_priv(rx_ring->netdev);
> +
> +	/* the ts_low field only contains the valid bit and sub-nanosecond
> +	 * precision, so we don't need to extract it.
> +	 */
> +	tstamp = le64_get_bits(qw3, IAVF_RXD_FLEX_QW3_TSTAMP_HIGH_M);
> +	ns = iavf_ptp_extend_32b_timestamp(adapter->ptp.cached_phc_time,
> +					   tstamp);
> +
> +	skb_tstamps = skb_hwtstamps(skb);
> +	memset(skb_tstamps, 0, sizeof(*skb_tstamps));
> +	skb_tstamps->hwtstamp = ns_to_ktime(ns);

Ah, so many times I've seen this pattern :D

&skb_shared_hwtstamps is 2 fields. One field you always initialize
above, the other one needs to be cleared. IOW, just do

	*skb_hwtstamps(skb) = (struct skb_shared_hwtstamps){
		.hwtstamp	= ns_to_ktime(ns),
	};

That's it!

Thanks,
Olek

Powered by blists - more mailing lists