On Sat, Sep 30, 2023 at 4:15 AM Frederic Weisbecker <frederic@xxxxxxxxxx> wrote: > > Le Thu, Sep 28, 2023 at 05:39:17PM -0400, Joel Fernandes a écrit : > > If srcu_invoke_callbacks() was really called for the rdp, I would have > > expected rcu_segcblist_advance() to advance all those pending > > callbacks to 304. > > > > I posit that probably srcu_invoke_callbacks() is not even being called > > in the first place, otherwise why is the DONE segment count still 0? > > DONE should not be 0, otherwise callbacks will not run. The segcblist > > of the rdp seems out of date perhaps due to rcu_invoke_callbacks() not > > being called due to some other bug (or a workqueue/timer issue causing > > the srcu_invoke_callbacks() to not run?). > > Speaking of, I'm looking at srcu_gp_end() and the comment after releasing > srcu_gp_mutex says: > > "/* A new grace period can start at this point. But only one. */" > > But what ensures that? > I think ->srcu_cb_mutex ensures that. /* Prevent more than one additional grace period. */ mutex_lock(&sup->srcu_cb_mutex); Thanks Neeraj > Because if we get preempted here and more than 3 grace periods elapse, then > ->srcu_have_cbs[the_same_idx] could get overriden and then all sorts of race > may lead to srcu_invoke_callbacks() to not be invoked. > > I'm not asking too loudly though because I'm sure I'm missing something obvious, > I see a lot of mirages these days...
