> On Tue, 29 Oct 2024 15:48:05 +0200 > Gur Stavi <gur.st...@huawei.com> wrote: > > > net_af_packet PMD multi "queue" support relies on Linux FANOUT > capability. > > Linux FANOUT is a SW based load balancer that is similar to HW RSS > which is > > more common for DPDK PMDs. Instead of multiple HW descriptor queues, AF > PACKET > > uses multiple sockets. > > HW RSS will typically drop a packet if its selected RX queue is empty. > However, > > Linux FANOUT, as a SW load balancer, can be configured to avoid this > packet > > drop by rolling over to the next socket. > > This rollover functionality was ALWAYS enabled in net_af_packet. It is > > surrounded by ifdef, but only to allow compilation on ancient Linux > versions > > that did not have it. > > > > Since DPDK applications are usually designed for HW based PMDs, this > rollover > > functionality, which the developers are likely unaware of, could be > confusing. > > > > Another option that is part of Linux FANOUT is DEFRAG that instructs > Linux to > > compose complete IP packet out of fragments before delivering it to the > PACKET > > socket. Again, this behavior typically does not exist for HW based PMDs > and may > > confuse users. > > > > This patch adds 2 options to control these features: > > rollover=[0|1],defrag=[0|1] > > For backward compatibility both features are enabled by default even > though most > > users will probably want both of them disabled. > > > > Signed-off-by: Gur Stavi <gur.st...@huawei.com> > > --- > > Makes sense to expose kernel options. But have all combinations been > tested?
I tested the command line parsing of the new options. I tested that the PMD is initialized successfully with all combinations. I did not test if Linux actually behaves according to the man page. I did not test traffic at all.
