> > Seems like the fallback to /dev should only occur if the device is not > > known to the UV framework. That is: > > > > * Search for <link> in /dev/net/; if it's found, we're done. > > * Check if <link> matches a device name known to linkmgmtd (or > > wherever this information is stored). If so, fail with ENOENT > > (since if the device had a link matching that name, it would've > > already been found). > > * Otherwise, fallback to a scan of /dev. > > > > But maybe I'm missing something. > > > So who will do the second step? libdlpi? Or part of /dev/net/<name> lookup?
I'd think it'd need to be libdlpi -- otherwise, the kernel would be making a policy decision about fallback that seems inappropriate. > I feel both are hacky. I agree it's gross -- but I think the strcmp() stuff would be worse. What would it take it get all of the DLPI devices in /dev/net? Is the problem due to the fact that devices like vni and sppp register themselves as DDI_PSEUDO? Or something else? (Sorry, my memory here is fuzzy.) -- meem
