On 12/05, Song, Yoong Siang wrote:
On Wednesday, December 4, 2024 11:49 PM, Stanislav Fomichev stfomichev@gmail.com wrote:
On 12/04, Song Yoong Siang wrote:
Set tx_type to HWTSTAMP_TX_ON to enable hardware timestamping for all outgoing packets.
Besides, set XDP_UMEM_TX_METADATA_LEN flag to reserve tx_metadata_len bytes of per-chunk metadata.
XDP_UMEM_TX_METADATA_LEN is missing after d5e726d9143c ("xsk: Require XDP_UMEM_TX_METADATA_LEN to actuate tx_metadata_len"), so that make sense. Maybe add a fixes tag?
Sure. I will add the fixes tag and submit with "PATCH bpf" prefix in next version.
And I don't see mlx5 looking at HWTSTAMP_TX anywhere in the drivers, so I'm assuming that's why I didn't need HWTSTAMP_TX_ON during my tests.. Which device are you testing against? I do see some hwts_tx_en checks in the stfmmac at least... Can you add these details to the commit message and respin?
I am testing on stmmac and igc drivers. You are right, stmmac needs it for hwts_tx_en check. Besides, igc needs it to set IGC_RING_FLAG_TX_HWTSTAMP flag.
Without this patch, user will need to manually enable tx hwts using command: sudo hwstamp_ctl -i eth0 -t 1 -r 1 after start xdp_hw_metadata.
Therefore, adding HWTSTAMP_TX_ON is not a bug fix solution. I will separate this as another new patch to "PATCH bpf-next" and provide detail in commit message.
Btw, is mlx5 driver always enable Tx HWTS?
I don't remember doing anything special to enable it. And looking at the code I also don't see any conditionals on HWTSTAMP_TX_ON.