On Mon, Nov 08, 2021 at 08:53:36PM +0100, Andrew Lunn wrote: > > I guess I will have to work on this again ASAP or we will end up with > > solution that I don't like. > > > > Nonetheless, what is your opinion about offloading netdev trigger vs > > introducing another trigger? > > It is a solution that fits the general pattern, do it in software, and > offload it if possible. > > However, i'm not sure the software solution actually works very well. > At least for switches. The two DSA drivers which implement > get_stats64() simply copy the cached statistics. The XRS700X updates > its cached values every 3000ms. The ar9331 is the same. Those are the > only two switch drivers which implement get_stats64 and none implement > get_stats. There was also was an issue that get_stats64() cannot > perform blocking calls. I don't remember if that was fixed, but if > not, get_stats64() is going to be pretty useless on switches. No it wasn't, I lost the interest. I feel pretty uneasy hooking up .ndo_get_stats64() to my switches, and basically opening the flood gates for random processes and kernel threads to send SPI transactions back and forth like it's nothing. Latency for programs like ptp4l and phc2sys is actually more important. > > We also need to handle drivers which don't actually implement > dev_get_stats(). That probably means only supporting offloads, all > modes which cannot be offloaded need to be rejected. This is pretty > much the same case of software control of the LEDs is not possible. > Unfortunately, dev_get_stats() does not return -EOPNOTSUPP, you need > to look at dev->netdev_ops->ndo_get_stats64 and > dev->netdev_ops->ndo_get_stats. > > Are you working on Marvell switches? Have you implemented > get_stats64() for mv88e6xxx? How often do you poll the hardware for > the stats? > > Given this, i think we need to bias the API so that it very likely > ends up offloading, if offloading is available. > > Andrew > We could use some of the newer stats APIs exposed by Jakub if .ndo_get_stats64 is not implemented, like ethtool_ops->get_eth_mac_stats. Although.. see above.
