On 3/2/2017 7:07 AM, Andrew Rybchenko wrote: > Implement EF10 (SFN7xxx and SFN8xxx) native datapaths which may be > chosen per device using PCI whitelist device arguments. > > libefx-based datapath implementation is bound to API and structure > imposed by the libefx. It has many indirect function calls to > provide HW abstraction (bad for CPU pipeline) and uses many data > structures: driver Rx/Tx queue, driver event queue, libefx Rx/Tx > queue, libefx event queue, libefx NIC (bad for cache). > > Native datapath implementation is fully separated from control > path to be able to use alternative control path if required > (e.g. kernel-aware). > > Native datapaths show better performance than libefx-based. > > Andrew Rybchenko (13): > net/sfc: callbacks should depend on EvQ usage > net/sfc: emphasis that RSS hash flag is an Rx queue flag > net/sfc: do not use Rx queue control state on datapath > net/sfc: factor out libefx-based Rx datapath > net/sfc: Rx scatter is a datapath-dependent feature > net/sfc: implement EF10 native Rx datapath > net/sfc: factory out libefx-based Tx datapath > net/sfc: VLAN insertion is a datapath dependent feature > net/sfc: TSO is a datapath dependent feature > net/sfc: implement EF10 native Tx datapath > net/sfc: multi-segment support as is Tx datapath features > net/sfc: implement simple EF10 native Tx datapath > net/sfc: support Rx packed stream EF10-specific datapath
Hi Andrew, Overall basic tests for the patchset is good. Only can you please update some patches' title, I commented into the patch already. Thanks, ferruh
