> On Apr 7, 2023, at 2:34 AM, Gleb Smirnoff <gleb...@freebsd.org> wrote: > > Hi, > > recently we had several drivers marked with IFF_KNOWSEPOCH > which reminded me that this flag was supposed to be temporary. > > Here is the change that introduced it e87c4940156. It was > caused by several drivers sending packets from non-interrupt > context and thus triggering NET_EPOCH_ASSERT(). It was about > 10 - 20 drivers having this problem initially and reduced down > to a few after 4426b2e64bd. We had a pretty heated dicussion > back then and I apologize for that. > > May I suggest before entering FreeBSD 14.0-RELEASE cycle we > will get back to what was there before e87c4940156?
It sounds good if only a few drivers need IFF_NEEDSEPOCH. > > To avoid the driver fallout that we used to have back in > early 2020, here is the plan. In ether_input() where we > now conditionally on the IFF_KNOWSEPOCH flag enter/exit the > epoch with INVARIANTS we will also conditionally enter/exit > in case we are supposed to be in the epoch wrt the flag, but > we are not. We will also print a warning once, like "interface > foo0 called if_input without epoch". This handling will be > converted to normal assertion after a couple months. Should also apply to infiniband_input(). BTW, is `in_epoch()` too heavy to replace flag IFF_NEEDSEPOCH or IFF_KNOWSEPOCH ? > > If everybody is fine with this suggestion I will post > a review. Otherwise please express your concerns. Good to make it moving forward ;) Best regards, Zhenlei > > -- > Gleb Smirnoff
