Re: [PATCH] i40e: fix the wrong PTP frequency calculation

From: Jacob Keller
Date: Tue Jun 27 2023 - 16:21:11 EST




On 6/26/2023 7:26 PM, Yajun Deng wrote:
> The new adjustment should be based on the base frequency, not the
> I40E_PTP_40GB_INCVAL in i40e_ptp_adjfine().
>
> This issue was introduced in commit 3626a690b717 ("i40e: use
> mul_u64_u64_div_u64 for PTP frequency calculation"), and was fixed in
> commit 1060707e3809 ("ptp: introduce helpers to adjust by scaled
> parts per million"). However the latter is a new feature and hasn't been
> backported to the stable releases.
>
> This issue affects both v6.0 and v6.1 versions, and the v6.1 version is
> an LTS version.
>
> Fixes: 3626a690b717 ("i40e: use mul_u64_u64_div_u64 for PTP frequency calculation")
> Cc: <stable@xxxxxxxxxxxxxxx> # 6.1
> Signed-off-by: Yajun Deng <yajun.deng@xxxxxxxxx>
> ---
> drivers/net/ethernet/intel/i40e/i40e_ptp.c | 4 ++--
> 1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/net/ethernet/intel/i40e/i40e_ptp.c b/drivers/net/ethernet/intel/i40e/i40e_ptp.c
> index ffea0c9c82f1..97a9efe7b713 100644
> --- a/drivers/net/ethernet/intel/i40e/i40e_ptp.c
> +++ b/drivers/net/ethernet/intel/i40e/i40e_ptp.c
> @@ -361,9 +361,9 @@ static int i40e_ptp_adjfine(struct ptp_clock_info *ptp, long scaled_ppm)
> 1000000ULL << 16);
>
> if (neg_adj)
> - adj = I40E_PTP_40GB_INCVAL - diff;
> + adj = freq - diff;
> else
> - adj = I40E_PTP_40GB_INCVAL + diff;
> + adj = freq + diff;
>
> wr32(hw, I40E_PRTTSYN_INC_L, adj & 0xFFFFFFFF);
> wr32(hw, I40E_PRTTSYN_INC_H, adj >> 32);

This straight forward fix makes sense. However, it wasn't obvious to me
without context why the 3626a690b717 ("i40e: use mul_u64_u64_div_u64 for
PTP frequency calculation") was where the fault got introduced. Thus,
here is that context for anyone else who failed to spot it just looking
at shrunk patch diffs.

--> code before that commit <---
> /**
> * i40e_ptp_adjfreq - Adjust the PHC frequency
> * @ptp: The PTP clock structure
> * @ppb: Parts per billion adjustment from the base
> *
> * Adjust the frequency of the PHC by the indicated parts per billion from the
> * base frequency.
> **/
> static int i40e_ptp_adjfreq(struct ptp_clock_info *ptp, s32 ppb)
> {
> struct i40e_pf *pf = container_of(ptp, struct i40e_pf, ptp_caps);
> struct i40e_hw *hw = &pf->hw;
> u64 adj, freq, diff;
> int neg_adj = 0;
>
> if (ppb < 0) {
> neg_adj = 1;
> ppb = -ppb;
> }
>
> freq = I40E_PTP_40GB_INCVAL;

frequency is left just as base I40E_PTP_40GB_INCVAL.

> freq *= ppb;
> diff = div_u64(freq, 1000000000ULL);
>
> if (neg_adj)
> adj = I40E_PTP_40GB_INCVAL - diff;
> else
> adj = I40E_PTP_40GB_INCVAL + diff;
>

So the base here can't be freq since we modify freq above using *=, but
using I40E_PTP_40GB_INCVAL is consistent.

> /* At some link speeds, the base incval is so large that directly
> * multiplying by ppb would result in arithmetic overflow even when
> * using a u64. Avoid this by instead calculating the new incval
> * always in terms of the 40GbE clock rate and then multiplying by the
> * link speed factor afterwards. This does result in slightly lower
> * precision at lower link speeds, but it is fairly minor.
> */
> smp_mb(); /* Force any pending update before accessing. */
> adj *= READ_ONCE(pf->ptp_adj_mult);
>

Finally, the multiply is applied last. This affects the combined base +
difference, and is done in order to avoid overflowing the *= used in the
original implementation.

> wr32(hw, I40E_PRTTSYN_INC_L, adj & 0xFFFFFFFF);
> wr32(hw, I40E_PRTTSYN_INC_H, adj >> 32);
>
> return 0;
> }


---> code after that commit <---
> /**
> * i40e_ptp_adjfreq - Adjust the PHC frequency
> * @ptp: The PTP clock structure
> * @ppb: Parts per billion adjustment from the base
> *
> * Adjust the frequency of the PHC by the indicated parts per billion from the
> * base frequency.
> **/
> static int i40e_ptp_adjfreq(struct ptp_clock_info *ptp, s32 ppb)
> {
> struct i40e_pf *pf = container_of(ptp, struct i40e_pf, ptp_caps);
> struct i40e_hw *hw = &pf->hw;
> u64 adj, freq, diff;
> int neg_adj = 0;
>
> if (ppb < 0) {
> neg_adj = 1;
> ppb = -ppb;
> }
>
> smp_mb(); /* Force any pending update before accessing. */
> freq = I40E_PTP_40GB_INCVAL * READ_ONCE(pf->ptp_adj_mult);
> diff = mul_u64_u64_div_u64(freq, (u64)ppb,
> 1000000000ULL);
>

Here, we assign freq to be the I40E_PTP_40GB_INCVAL times the
ptp_adj_mult value, and then we don't modify it, instead using
mul_u64_u64_div_u64.

> if (neg_adj)
> adj = I40E_PTP_40GB_INCVAL - diff;
> else
> adj = I40E_PTP_40GB_INCVAL + diff;
>

But then the diff is applied on the wrong value, and no multiplication
is done afterwards.

> wr32(hw, I40E_PRTTSYN_INC_L, adj & 0xFFFFFFFF);
> wr32(hw, I40E_PRTTSYN_INC_H, adj >> 32);
>
> return 0;
> }

---> current version with adjust_by_scaled_ppm <---
> /**
> * i40e_ptp_adjfine - Adjust the PHC frequency
> * @ptp: The PTP clock structure
> * @scaled_ppm: Scaled parts per million adjustment from base
> *
> * Adjust the frequency of the PHC by the indicated delta from the base
> * frequency.
> *
> * Scaled parts per million is ppm with a 16 bit binary fractional field.
> **/
> static int i40e_ptp_adjfine(struct ptp_clock_info *ptp, long scaled_ppm)
> {
> struct i40e_pf *pf = container_of(ptp, struct i40e_pf, ptp_caps);
> struct i40e_hw *hw = &pf->hw;
> u64 adj, base_adj;
>
> smp_mb(); /* Force any pending update before accessing. */
> base_adj = I40E_PTP_40GB_INCVAL * READ_ONCE(pf->ptp_adj_mult);
>
> adj = adjust_by_scaled_ppm(base_adj, scaled_ppm);
>

Using adjust_by_scaled_ppm correctly performs the calculation and uses
the base adjustment, so there's no error here.

> wr32(hw, I40E_PRTTSYN_INC_L, adj & 0xFFFFFFFF);
> wr32(hw, I40E_PRTTSYN_INC_H, adj >> 32);
>
> return 0;
> }


Thanks for finding and fixing this mistake. I think its the simplest fix
to get into the stable kernel that are broken, since taking the
adjust_by_scaled_ppm version would require additional patches.

Reviewed-by: Jacob Keller <jacob.e.keller@xxxxxxxxx>