On Thu, 2018-03-01 at 08:16 -0800, Stephen Hemminger wrote: > > I went through a several possible alternatives. > 1. Add a counter in the hash bucket head (like listen already has). > But not namespace aware > 2. Add a percpu counter in network namespace (new struct tcp_netns) > Logical and adds place to move tcp open sockets as well. > But more expensive and several places in code don't have easy > access to namespace.
__this_cpu_inc() should not be very expensive. We always have access to socket namespace when inserting/deleting a bind entry. Since we take the spinlock at that times, inserting __this_cpu_inc(sock_net(sk)->ipv4.bind_counter) would add few instructions. > 3. Counting entries in userspace; defeats the purpose of -s flag. > > Agree it is not an urgent statistic, it is just it got broken; willing to > just drop it. Not counting the fact that it has really no purpose. > > What about adding cond_resched between buckets like other places do? This still is dirtying at least 1MB of memory and performs ~64K atomic operations. Waste of cpu caches and memory bus. If we consider converting the whole /proc/net/socketstat to provide a netns view only, we might use a percpu structure holding all TCP socket counters, to make sure no more than one cache line per cpu is used while folding all the cpus counters.
