On Tue, Oct 13, 2020 at 01:20:08AM +0200, Frederic Weisbecker wrote:
> On Wed, Sep 23, 2020 at 11:22:09AM -0400, Joel Fernandes (Google) wrote:
> > +/* Return number of callbacks in a segment of the segmented callback list. */
> > +static void rcu_segcblist_add_seglen(struct rcu_segcblist *rsclp, int seg, long v)
> > +{
> > +#ifdef CONFIG_RCU_NOCB_CPU
> > + smp_mb__before_atomic(); /* Up to the caller! */
> > + atomic_long_add(v, &rsclp->seglen[seg]);
> > + smp_mb__after_atomic(); /* Up to the caller! */
> > +#else
> > + smp_mb(); /* Up to the caller! */
> > + WRITE_ONCE(rsclp->seglen[seg], rsclp->seglen[seg] + v);
> > + smp_mb(); /* Up to the caller! */
> > +#endif
> > +}
>
> I know that these "Up to the caller" comments come from the existing len
> functions but perhaps we should explain a bit more against what it is ordering
> and what it pairs to.
>
> Also why do we need one before _and_ after?
>
> And finally do we have the same ordering requirements than the unsegmented len
> field?
Hi Paul and Neeraj,
Would be nice to discuss this on the call. I actually borrowed the memory
barriers from add_len() just to be safe, but I think Frederic's points are
valid. Would be nice if we can go over all the usecases and discuss which
memory barriers are needed. Thanks for your help!
Another thought: inc_len() calls add_len() which already has smp_mb(), so
callers of inc_len also do not need memory barriers I think.
thanks,
- Joel
> > +
> > +/* Move from's segment length to to's segment. */
> > +static void rcu_segcblist_move_seglen(struct rcu_segcblist *rsclp, int from, int to)
> > +{
> > + long len;
> > +
> > + if (from == to)
> > + return;
> > +
> > + len = rcu_segcblist_get_seglen(rsclp, from);
> > + if (!len)
> > + return;
> > +
> > + rcu_segcblist_add_seglen(rsclp, to, len);
> > + rcu_segcblist_set_seglen(rsclp, from, 0);
> > +}
> > +
> [...]
> > @@ -245,6 +283,7 @@ void rcu_segcblist_enqueue(struct rcu_segcblist *rsclp,
> > struct rcu_head *rhp)
> > {
> > rcu_segcblist_inc_len(rsclp);
> > + rcu_segcblist_inc_seglen(rsclp, RCU_NEXT_TAIL);
> > smp_mb(); /* Ensure counts are updated before callback is enqueued. */
>
> Since inc_len and even now inc_seglen have two full barriers embracing the add up,
> we can probably spare the above smp_mb()?
>
> > rhp->next = NULL;
> > WRITE_ONCE(*rsclp->tails[RCU_NEXT_TAIL], rhp);
> > @@ -274,27 +313,13 @@ bool rcu_segcblist_entrain(struct rcu_segcblist *rsclp,
> > for (i = RCU_NEXT_TAIL; i > RCU_DONE_TAIL; i--)
> > if (rsclp->tails[i] != rsclp->tails[i - 1])
> > break;
> > + rcu_segcblist_inc_seglen(rsclp, i);
> > WRITE_ONCE(*rsclp->tails[i], rhp);
> > for (; i <= RCU_NEXT_TAIL; i++)
> > WRITE_ONCE(rsclp->tails[i], &rhp->next);
> > return true;
> > }
> >
> > @@ -403,6 +437,7 @@ void rcu_segcblist_advance(struct rcu_segcblist *rsclp, unsigned long seq)
> > if (ULONG_CMP_LT(seq, rsclp->gp_seq[i]))
> > break;
> > WRITE_ONCE(rsclp->tails[RCU_DONE_TAIL], rsclp->tails[i]);
> > + rcu_segcblist_move_seglen(rsclp, i, RCU_DONE_TAIL);
>
> Do we still need the same amount of full barriers contained in add() called by move() here?
> It's called in the reverse order (write queue then len) than usual. If I trust the comment
> in rcu_segcblist_enqueue(), the point of the barrier is to make the length visible before
> the new callback for rcu_barrier() (although that concerns len and not seglen). But here
> above, the unsegmented length doesn't change. I could understand a write barrier between
> add_seglen(x, i) and set_seglen(0, RCU_DONE_TAIL) but I couldn't find a paired couple either.
>
> > }
> >
> > /* If no callbacks moved, nothing more need be done. */
> > @@ -423,6 +458,7 @@ void rcu_segcblist_advance(struct rcu_segcblist *rsclp, unsigned long seq)
> > if (rsclp->tails[j] == rsclp->tails[RCU_NEXT_TAIL])
> > break; /* No more callbacks. */
> > WRITE_ONCE(rsclp->tails[j], rsclp->tails[i]);
> > + rcu_segcblist_move_seglen(rsclp, i, j);
>
> Same question here (feel free to reply "same answer" :o)
>
> Thanks!
