Hi, I like this. It seems more robust and easier extended than all the environment variable.
We should also keep the basic portname. Ideally io_ops should be in performance order so if I write generic code that uses en0, it'll try netmap, then sock_mmap and finally fallback to standard sockets if others won't work. And/ Or a way to manually specify fallbacks: NETMAP::SOCKET_MMAP::SOCKET:eth0 Nicolas On 11/12/2015 02:38 PM, Bala Manoharan wrote: > Hi, > > I went through the same issue as well.. > In ODP we have defined that the dev-name is implementation specific > and hence I like the idea to add socket type to dev-name something > like "SOCKET_MMAP:eth0"/ "NETMAP: eth0" and this maps well with the > ODP definition for device name. > > Regards, > Bala > > On 12 November 2015 at 18:23, Christophe Milard > <christophe.mil...@linaro.org> wrote: >> Hi, >> >> Today, as far as I can see (at least for linux generic), the type of pktio >> selected at open time depends on: >> 1) the priority order defined in pktio_io/io_ops.c (currentely: netmap mmap >> socket and usual sockets) >> 2)A set of environment variable (ODP_PKTIO_DISABLE_NETMAP, >> ODP_PKTIO_DISABLE_SOCKET_MMAP...) from which the order defined in 1) can be >> changed... >> 3)Some mechanism taking the device name as a hint (looback, pcap...) >> >> I am not sure I really understand why this is done so, and how these >> mechanism can be extented: >> >> I can see a future in which a odp application will need to talk (read send >> and receive packet) to different type of pktio at the same time. Possibly >> one to the linux jernel (Ravineet's work), some normal sockets (possibly >> netmapped) and some NICs (possibly many) -my problem-.... and surely some >> more to come. >> >> Aren't we missing a proper pktio selection mechanism? Can/Should the >> application be really pktio-type (un)aware?. >> >> A first hack could be having a dev name like: "NIC:<pci_address>" to start a >> PMD driver on the nic at that address, but by choosing specific dev name the >> application becomes well aware of the pktio devices anyway.... >> The underlying question is also whether ODP should be performing PCI >> enumeration and how NIC should be selected... >> >> Christophe. >> >> >> >> > _______________________________________________ > lng-odp mailing list > lng-odp@lists.linaro.org > https://lists.linaro.org/mailman/listinfo/lng-odp _______________________________________________ lng-odp mailing list lng-odp@lists.linaro.org https://lists.linaro.org/mailman/listinfo/lng-odp
