On 7/19/2024 10:10 PM, Jeff Johnson wrote: > On 7/14/2024 7:38 PM, Baochen Qiang wrote: >> In transmit path, it is likely that the iova is not aligned to PCIe TLP >> max payload size, which is 128 for WCN7850. Normally in such cases hardware >> is expected to split the packet into several parts in a manner such that >> they, other than the first one, have aligned iova. However due to hardware >> limitations, WCN7850 does not behave like that properly with some specific >> unaligned iova in transmit path. This easily results in target hang in a >> KPI transmit test: packet send/receive failure, WMI command send timeout >> etc. Also fatal error seen in PCIe level: >> >> ... >> Capabilities: ... >> ... >> DevSta: ... FatalErr+ ... >> ... >> ... >> >> Work around this by manually moving/reallocating payload buffer such that >> we can map it to a 128 bytes aligned iova. The moving requires sufficient >> head room or tail room in skb: for the former we can do ourselves a favor >> by asking some extra bytes when registering with mac80211, while for the >> latter we can do nothing. >> >> Moving/reallocating buffer consumes additional CPU cycles, but the good news >> is that an aligned iova increases PCIe efficiency. In my tests on some X86 >> platforms the KPI results are almost consistent. >> >> Since this is seen only with WCN7850, add a new hardware parameter to >> differentiate from others. > > I asked for expert opinion on this patch and received the following response. > Baochen, can you take a look at this suggestion? > >> Aligning headers is sometimes done, but it appears the driver >> doesn't support scatter gather? I think the author may want to advertise right, ath12k does not support SG currently. >> scatter and linearize manually in the driver, to a correct offset. is there an existing skb API or API combinations which can do that for me? I checked __skb_linearize() and it does not take an 'offset' argument. >> Because now core is linearizing the skb in validate_xmit_skb() >> and then the driver moves it a second time.. > >
