On 6/5/15, 3:46 AM, "Thomas Monjalon" <thomas.monjalon at 6wind.com> wrote:
>2015-06-04 22:10, Andrew Harvey: >> On 6/4/15, 7:58 AM, "Stephen Hemminger" <stephen at networkplumber.org> >>wrote: >> >"Andrew Harvey (agh)" <agh at cisco.com> wrote: >> >> I believe that their is value in this interface for software stacks >>not >> >> based on Linux being moved toward DPDK that need simple operations >>like >> >> getting the mac address. Some of these stacks have a dearth of >> >>resources >> >> available and dedicating a core/thread to KNI to get/set a mac >>address >> >> is considered excessive. There are also issues with 32/64 bit kernel >> >> integration >> >> using KNI. If the ethtool interface is not the correct interface >>then >> >> please help me >> >> understand what should/could have been used. If ethtool is considered >> >>'old >> >> and clunky? >> >> Stephen's and your input would be valuable in designing another >> >>interface >> >> with >> >> similar properties. The use-case is pretty simple and there is no >>plans >> >> for moving >> >> anything back into the kernel on the contrary its the complete >>opposite. >> >> >> >> ? Andy >> > >> >We have DPDK API's to do this, and any added wrappers make it bigger. >> >I don't see why calling your ethtool API is better than calling >> >rte_eth* API. >> > >> >If there is a missing functionality in the rte_ethXXX api's for an >> >application then add that. For example: rte_eth_mac_addr_get() >> >> I am getting somewhat confused by your latest comments. Your first >>email >> (referenced below) looked really positive and I found your suggestions >> useful. Your latest post appears to contradict this and now the >>interface >> was there all the time. The wrapper fa?ade provided by the ethtool >> library provide a clean separation of concerns and will allow people to >> migrate from not only KNI but in our case from a legacy system. If a >> software stack has requirements to work with multiple IO abstractions >> then the ethtool approach is attractive. I would speculate that many >> other stacks moving towards dpdk will have similar issues. >> >> Summarizing, for our use-cases the ethtool interface facilitated our >> adoption to dpdk while allowing us to support our legacy IO >>abstractions. > >Stephen and me say the same thing about using the ethdev API. And your would have a point would be valid if dpdk were available to every interface we support (it is not) and on every processor architecture that we support (it is not) and every OS we support (it is not). So to minimize entropy in the code why not leave the client code the same ioctl(fd, ?) and hide the implementation detail in a wrapper library. We have a large legacy code base to move forward and sprinkling special interest code like rte_xxx throughout every client we have is not appropriate at this time. >We don't understand why using a fake ethtool lib would be easier. >Though you are saying it "facilitated [your] adoption to dpdk". >Please could you explain why using an ethtool-like API is easier than >using the existing ethdev API? >In any case, you have to develop a specific backend for DPDK >(rte_ethtool would be also DPDK-specific). > >It seems you already started to use such an ethtool implementation. >Please note that our goal is not to prevent Cisco from upstreaming >(evidence with enic driver integration) but we want to guide you, and >others having the same needs, to the best solution for everybody. >That's why we need to understand what we (or you) are missing. >Maybe that it would be clearer with some code examples (which would >go in the lib documentation if any). > >Thanks
