Hello, Frederic!
> Le Thu, Mar 05, 2026 at 11:59:15AM +0100, Uladzislau Rezki a écrit :
> > On Tue, Mar 03, 2026 at 03:45:58PM -0500, Joel Fernandes wrote:
> > > On Mon, 02 Mar 2026 11:04:04 +0100, Uladzislau Rezki (Sony) wrote:
> > >
> > > > * The latch is cleared only when the pending requests are fully
> > > > drained(nr == 0);
> > >
> > > > +static void rcu_sr_normal_add_req(struct rcu_synchronize *rs)
> > > > +{
> > > > + long nr;
> > > > +
> > > > + llist_add((struct llist_node *) &rs->head, &rcu_state.srs_next);
> > > > + nr = atomic_long_inc_return(&rcu_sr_normal_count);
> > > > +
> > > > + /* Latch: only when flooded and if unlatched. */
> > > > + if (nr >= RCU_SR_NORMAL_LATCH_THR)
> > > > + (void)atomic_cmpxchg(&rcu_sr_normal_latched, 0, 1);
> > > > +}
> > >
> > > I think there is a stuck-latch race here. Once llist_add() places the
> > > entry in srs_next, the GP kthread can pick it up and fire
> > > rcu_sr_normal_complete() before the latching cmpxchg runs. If the last
> > > in-flight completion drains count to zero in that window, the unlatch
> > > cmpxchg(latched, 1, 0) fails (latched is still 0 at that moment), and
> > > then the latching cmpxchg(latched, 0, 1) fires anyway — with count=0:
> > >
> > > CPU 0 (add_req, count just hit 64) GP kthread
> > > ---------------------------------- ----------
> > > llist_add() <-- entry now in srs_next
> > > inc_return() --> nr = 64
> > > [preempted]
> > > rcu_sr_normal_complete() x64:
> > > dec_return -> count:
> > > 64..1..0
> > > count==0:
> > > cmpxchg(latched, 1, 0)
> > > --> FAILS (latched still
> > > 0)
> > > [resumes]
> > > cmpxchg(latched, 0, 1) --> latched = 1
> > >
> > > Final state: count=0, latched=1 --> STUCK LATCH
> > >
> > > All subsequent synchronize_rcu() callers see latched==1 and take the
> > > fallback path (not counted). With no new SR-normal callers,
> > > rcu_sr_normal_complete() is never reached again, so the unlatch
> > > cmpxchg(latched, 1, 0) never fires. The latch is permanently stuck.
> > >
> > > This requires preemption for a full GP duration between llist_add() and
> > > the cmpxchg, which is probably more likely on PREEMPT_RT or heavily loaded
> > > systems.
> > >
> > > The fix: move the cmpxchg *before* llist_add(), so the entry is not
> > > visible to the GP kthread until after the latch is already set.
> > >
> > > That should fix it, thoughts?
> > >
> > Yes and thank you!
> >
> > We can improve it even more by removing atomic_cmpxchg() in
> > the rcu_sr_normal_add_req() function, because only one context
> > sees the (nr == RCU_SR_NORMAL_LATCH_THR) condition:
> >
> > <snip>
> > diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> > index 86dc88a70fd0..72b340940e11 100644
> > --- a/kernel/rcu/tree.c
> > +++ b/kernel/rcu/tree.c
> > @@ -1640,7 +1640,7 @@ static struct workqueue_struct *sync_wq;
> >
> > /* Number of in-flight synchronize_rcu() calls queued on srs_next. */
> > static atomic_long_t rcu_sr_normal_count;
> > -static atomic_t rcu_sr_normal_latched;
> > +static int rcu_sr_normal_latched; /* 0/1 */
> >
> > static void rcu_sr_normal_complete(struct llist_node *node)
> > {
> > @@ -1662,7 +1662,7 @@ static void rcu_sr_normal_complete(struct llist_node
> > *node)
> > * drained and if it has been latched.
> > */
> > if (nr == 0)
> > - (void)atomic_cmpxchg(&rcu_sr_normal_latched, 1, 0);
> > + (void)cmpxchg(&rcu_sr_normal_latched, 1, 0);
> > }
> >
> > static void rcu_sr_normal_gp_cleanup_work(struct work_struct *work)
> > @@ -1808,14 +1808,22 @@ static bool rcu_sr_normal_gp_init(void)
> >
> > static void rcu_sr_normal_add_req(struct rcu_synchronize *rs)
> > {
> > - long nr;
> > + /*
> > + * Increment before publish to avoid a complete
> > + * vs enqueue race on latch.
> > + */
> > + long nr = atomic_long_inc_return(&rcu_sr_normal_count);
> >
> > - llist_add((struct llist_node *) &rs->head, &rcu_state.srs_next);
> > - nr = atomic_long_inc_return(&rcu_sr_normal_count);
> > + /*
> > + * Latch on threshold crossing. (nr == RCU_SR_NORMAL_LATCH_THR)
> > + * can be true only for one context, avoiding contention on the
> > + * write path.
> > + */
> > + if (nr == RCU_SR_NORMAL_LATCH_THR)
> > + WRITE_ONCE(rcu_sr_normal_latched, 1);
>
> Isn't it still racy?
>
> rcu_sr_normal_add_req rcu_sr_normal_complete
> --------------------- ----------------------
> nr =
> atomic_long_dec_return(&rcu_sr_normal_count);
> // nr == 0
> ======= PREEMPTION
> =======
> // 64 tasks doing synchronize_rcu()
> rcu_sr_normal_add_req()
> WRITE_ONCE(rcu_sr_normal_latched, 1);
>
> cmpxchg(&rcu_sr_normal_latched, 1, 0);
>
>
> Also more generally there is nothing that orders the WRITE_ONCE() with the
> cmpxchg.
>
Yep that i know. This is rather "relaxed" mechanism rather than
a strictly ordered. The race you described can happen but i do not
find it as a problem because as noted it is relaxed policy flag.
But WRITE_ONCE() i can replace by:
if (nr == RCU_SR_NORMAL_LATCH_THR)
cmpxchg(&rcu_sr_normal_latched, 0, 1);
>
> Is it possible to remove rcu_sr_normal_latched and simply deal with
> comparisons
> between rcu_sr_normal_count and RCU_SR_NORMAL_LATCH_THR?
>
It is. But the idea with latch is a bit different then just checking
threshold. The main goal is to detect flood and lath the path until
__all__ users are flushed. I.e. it implements hysteresis to prevent
repeated switches around the threshold.
With your proposal behaviour becomes different.
Thoughts?
--
Uladzislau Rezki
