On Tuesday, July 8, 2025 4:55 AM, Jakub Kicinski kuba@kernel.org wrote:
On Tue, 1 Jul 2025 12:29:38 +0800 Song Yoong Siang wrote:
|<---sizeof(xdp_meta)--| | | struct xdp_meta rx_desc->address ^ ^ | |+----------+----------------------+------------+------+ | headroom | custom metadata | reserved | data | +----------+----------------------+------------+------+ ^ ^ ^ | | | struct xdp_meta xdp_buff->data_meta xdp_buff->data | | |<---sizeof(xdp_meta)--|
Huh. Did AF_XDP maintainers explicitly sign off on this or it's just how IGC implementation works and nobody noticed?
Previously, IGC do copy out the Rx hwts from metadata area, so no problem when implementing XDP Rx metadata.
After that, net_device_ops.ndo_get_tstamp() is added into IGC to support timestamping from both free-running clock and adjustable clock. The 2 timers are stored in the metadata area, thus causing the issue.
For normal XDP my understanding is that its the driver's responsibility to move the "reserved" stuff out of place before presenting the frame to program.
Is it means that driver needs to move out the "reserved" stuff before XDP program and then move back the stuff after XDP program for certain situation, like XDP_PASS?
IMHO, if driver is allowed to use some portion of the metadata area, then the packet processing will be more efficiency and also align with the "zero-copy" idea.
Any thoughts?