> On Fri, 18 Oct 2024 05:50:20 +0000 > Chaoyong He <chaoyong...@corigine.com> wrote: > > > > On Fri, 18 Oct 2024 03:21:28 +0000 > > > Chaoyong He <chaoyong...@corigine.com> wrote: > > > > > > > > RTE_PKTMBUF_HEADROOM); } > > > > > > + > > > > > > > > > > Not sure why this is needed? What is the problem with the original > code? > > > > > Are you trying to force packets to be segmented? > > > > > > > > Actually, we are trying to force packets *not* segmented by making > > > > the > > > mbuf size large enough to hold the packets. > > > > > > > > In our user case, we start l3fwd app with parameter '--max-pkt-len > > > > 4000', > > > and obviously the original logic with RTE_MBUF_DEFAULT_DATAROOM > mbuf > > > size will cause the packets to be segmented. > > > > Which is not what we want, so we add this new '--mbuf-size=4096' > > > parameter, the mbuf size will large enough to hold even the largest > > > packet. > > > > > > > > Do you think this make sense? > > > > > > Maybe query the driver, and use the max_rx_pkt_len as input to > > > deciding the right mbuf size. > > > > Sorry, I am not quite understanding here. > > I can't find 'max_rx_pkt_len' in l3fwd app, instead it's exist testpmd app. > > Could you please explain a little more about the advice? > > In rte_eth_dev_info, I meant the field max_rx_bufsize and there is also > max_rx_pktlen. > > > > > > If max-pkt-len was 4000 and driver can only take 2K buffers, then > > > use 2K mbuf size. > > > If max-pkt-len was 1500 then use mtu + headroom and round up
Oh, I understand what you mean now, thanks for the clarification. But the solution you suppose is not flexible enough, thus can't satisfy our needs. Follow your example and consider this situation: If max-pkt-len was 4000 and driver can only take 2K buffers, then use 2K mbuf size. But we want to measure the performance when the mbuf size is 1024 and 512. Then there is no way to do this in your solution, I suppose? But with our '--mbuf-size' parameter, we can easily do that. Thanks for your hint, we realized our solution also has a little problem, which not consider the 'max_rx_bufsize' of rte_eth_dev_info, and we will fix that in the next version patch.
