On Tue, Oct 15, 2019 at 09:42:49AM -0700, Wei Wang wrote: > On Tue, Oct 15, 2019 at 7:45 AM David Ahern <dsah...@gmail.com> wrote: > > > > On 10/14/19 1:26 PM, Martin Lau wrote: > > > > > > AFAICT, even for the route that are affected by > > > fib6_update_sernum_upto_root(), > > > I don't see the RTF_PCPU route is re-created. v6 sk does > > > dst_check() => re-lookup the fib6 => > > > found the same RTF_PCPU (but does not re-create it) => > > > update the sk with new cookie in ip6_dst_store() > > > > Hmm... That is a good point. Why does v4 need to recreate the dst > cache even though the route itself is not changed? > Now that I think about it, I agree with Martin's previous comment: it > probably is because v4 code does not cache rt->rt_genid into the > socket and every user of the rt is sharing the same rt_genid stored in > the route itself. Exactly :) If no re-create, dst_dev_put() can be avoided. The root cause is not really related to the global NS rt_genid. A granular rt_genid may help to reduce the race on dst_dev_put() but it will still happen. (that aside, improving the NS rt_genid would still be great).
Thinking more about it, this issue should not be limited to input. I think you fix is right. > > > > > That's fine. The pcpu cache is per nexthop (fib6_nh) for a specific > > gateway/device. > > > > The invalidate forces another lookup for the intended destination after > > the change to the fib. If the lookup resolves to the same fib entry and > > nexthop, then re-using the same cached dst/rt6_info is ok.
