Maciej Fijalkowski <maciej.fijalkowski@xxxxxxxxx> writes: > Under rare circumstances there might be a situation where a requirement > of having a XDP Tx queue per core could not be fulfilled and some of the > Tx resources would have to be shared between cores. This yields a need > for placing accesses to xdp_rings array onto critical section protected > by spinlock. > > Design of handling such scenario is to at first find out how many queues > are there that XDP could use. Any number that is not less than the half > of a count of cores of platform is allowed. XDP queue count < cpu count > is signalled via new VSI state ICE_VSI_XDP_FALLBACK which carries the > information further down to Rx rings where new ICE_TX_XDP_LOCKED is set > based on the mentioned VSI state. This ring flag indicates that locking > variants for getting/putting xdp_ring need to be used in fast path. > > For XDP_REDIRECT the impact on standard case (one XDP ring per CPU) can > be reduced a bit by providing a separate ndo_xdp_xmit and swap it at > configuration time. However, due to the fact that net_device_ops struct > is a const, it is not possible to replace a single ndo, so for the > locking variant of ndo_xdp_xmit, whole net_device_ops needs to be > replayed. > > It has an impact on performance (1-2 %) of a non-fallback path as > branches are introduced. I generally feel this is the right approach, although the performance impact is a bit unfortunately, obviously. Maybe it could be avoided by the use of static_branch? I.e., keep a global refcount of how many netdevs are using the locked path and only activate the check in the fast path while that refcount is >0? -Toke
