On 09/10, Eric W. Biederman wrote: > > Ok. I think I see the where the confusion is. We were looking > at different parts of the puzzle. But I we need to resolve this > to make certain I didn't do something clever and racy. Yes, I think we misunderstood each other :) > As for the rest of your suggestion it would not be hard to be able to > follow a struct pid pointer in an rcu safe way, and we do in the pid > hash table. In other contexts so far I always have other variables > that need to be updated in concert, so there isn't a point in coming > up with a lockless implementation. I believe vt_pid is the only > case that I have run across where this is a problem and I have > at least preliminary patches for every place where signals are > sent. > > Updating this old code is painful. No, no, we shouldn't change the old code, it is fine. Just in case, to avoid any possible confusion. put_pid(pid) has the following restrictions. The caller should ensure that any other possible reference to this pid "owns" it (did get_pid()). So we can add a new helper, put_pid_rcu(). It is ok if this pid is used in parallel under rcu_read_lock() without bumping pid->count. Contrary, the only restriction those users must not call get_pid(pid). But yes, you are right, I don't see an immediate usage of put_pid_rcu(). Oleg.