Am 22.05.2018 um 12:50 schrieb Benjamin Drung: > Am Dienstag, den 22.05.2018, 12:36 +0200 schrieb Michael Biebl: >> Am 22.05.2018 um 12:27 schrieb Benjamin Drung: >>> What do you suggest to do instead? >> >> For that I know too little how rdma works, so I can't really answer >> that. > > Mellanox ConnectX architecture is: mlx4_core is the lower level > PCI driver which register on the PCI id, and protocol specific drivers > are depended on it: mlx4_en - for Ethernet and mlx4_ib for Infiniband. > NIC could have multiple ports which can change their type dynamically. > We use the request_module() call to load the relevant protocol driver > when needed: on loading time or at port type change event. The ib_ipoib > kernel modules provides IP over Infiniband support. > > The order is as following: > 1) The kernel loads mlx4_core on boot (due to device detection) > 2) mlx4_core loads mlx4_ib > 3) udev triggers rdma-load-modules@infiniband.service > 4) rdma-load-modules@infiniband.service loads ib_ipoib (can be > configured in /etc/rdma)
Afaics there is a race condition here. How do you guarantee that ib_ipoib is loaded before systemd-udev-settle.service is started? > 5) networking.service brings up the IPoIB devices Not sure if ifupdown is the best fit for such a configuration, at least not networking.service. Maybe allow-hotplug would be a better fit or another networking configuration system which is more dynamic and does listen to uevents? Michael -- Why is it that all of the instruments seeking intelligent life in the universe are pointed away from Earth?
signature.asc
Description: OpenPGP digital signature
