Hi Andre, "Guedes, Andre" <andre.gue...@intel.com> writes:
> Hi Vinicius, > >> On Sep 23, 2019, at 10:04 PM, Vinicius Costa Gomes >> <vinicius.go...@intel.com> wrote: >> >> The problem happens because that when offloading is enabled, the cbs >> instance is not added to the list. >> >> Also, the current code doesn't handle correctly the case when offload >> is disabled without removing the qdisc: if the link speed changes the >> credit calculations will be wrong. When we create the cbs instance >> with offloading enabled, it's not added to the notification list, when >> later we disable offloading, it's not in the list, so link speed >> changes will not affect it. >> >> The solution for both issues is the same, add the cbs instance being >> created unconditionally to the global list, even if the link state >> notification isn't useful "right now". > > I believe we could fix both issues described above and still don’t > notify the qdisc about link state if we handled the list > insertion/removal in cbs_change() instead. > > Reading the cbs code more carefully, it seems it would be beneficial > to refactor the offload handling. For example, we currently init the > qdisc_watchdog even if it’s not useful when offload is enabled. Now, > we’re going to notify the qdisc even if it’s not useful too. I like your idea, but even after reading your email and the code a couple of times, I couldn't come up with anything quickly that wouldn't complicate things (i.e. add more code), I would need to experiment a bit. (btw, qdisc_watchdog_init() is just initializing some fields in a struct, and the notification part should be quite rare in practice). So my suggestion is to keep this patch as is, as it solves a real crash that a colleague faced. Later, we can try and simplify things even more. Cheers, -- Vinicius P.S.: I think I am still a bit traumatized but getting the init() and destroy() right were the hardest parts when we were trying to uptream this. That's why I am hesitant about adding more code to those flows.
