6.6-stable review patch. If anyone has any objections, please let me know.
------------------
From: Frederic Weisbecker frederic@kernel.org
[ Upstream commit 55d4669ef1b76823083caecfab12a8bd2ccdcf64 ]
When rcu_barrier() calls rcu_rdp_cpu_online() and observes a CPU off rnp->qsmaskinitnext, it means that all accesses from the offline CPU preceding the CPUHP_TEARDOWN_CPU are visible to RCU barrier, including callbacks expiration and counter updates.
However interrupts can still fire after stop_machine() re-enables interrupts and before rcutree_report_cpu_dead(). The related accesses happening between CPUHP_TEARDOWN_CPU and rnp->qsmaskinitnext clearing are _NOT_ guaranteed to be seen by rcu_barrier() without proper ordering, especially when callbacks are invoked there to the end, making rcutree_migrate_callback() bypass barrier_lock.
The following theoretical race example can make rcu_barrier() hang:
CPU 0 CPU 1 ----- ----- //cpu_down() smpboot_park_threads() //ksoftirqd is parked now <IRQ> rcu_sched_clock_irq() invoke_rcu_core() do_softirq() rcu_core() rcu_do_batch() // callback storm // rcu_do_batch() returns // before completing all // of them // do_softirq also returns early because of // timeout. It defers to ksoftirqd but // it's parked </IRQ> stop_machine() take_cpu_down() rcu_barrier() spin_lock(barrier_lock) // observes rcu_segcblist_n_cbs(&rdp->cblist) != 0 <IRQ> do_softirq() rcu_core() rcu_do_batch() //completes all pending callbacks //smp_mb() implied _after_ callback number dec </IRQ>
rcutree_report_cpu_dead() rnp->qsmaskinitnext &= ~rdp->grpmask;
rcutree_migrate_callback() // no callback, early return without locking // barrier_lock //observes !rcu_rdp_cpu_online(rdp) rcu_barrier_entrain() rcu_segcblist_entrain() // Observe rcu_segcblist_n_cbs(rsclp) == 0 // because no barrier between reading // rnp->qsmaskinitnext and rsclp->len rcu_segcblist_add_len() smp_mb__before_atomic() // will now observe the 0 count and empty // list, but too late, we enqueue regardless WRITE_ONCE(rsclp->len, rsclp->len + v); // ignored barrier callback // rcu barrier stall...
This could be solved with a read memory barrier, enforcing the message passing between rnp->qsmaskinitnext and rsclp->len, matching the full memory barrier after rsclp->len addition in rcu_segcblist_add_len() performed at the end of rcu_do_batch().
However the rcu_barrier() is complicated enough and probably doesn't need too many more subtleties. CPU down is a slowpath and the barrier_lock seldom contended. Solve the issue with unconditionally locking the barrier_lock on rcutree_migrate_callbacks(). This makes sure that either rcu_barrier() sees the empty queue or its entrained callback will be migrated.
Signed-off-by: Frederic Weisbecker frederic@kernel.org Signed-off-by: Paul E. McKenney paulmck@kernel.org Signed-off-by: Sasha Levin sashal@kernel.org --- kernel/rcu/tree.c | 10 +++++++--- 1 file changed, 7 insertions(+), 3 deletions(-)
diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c index 8cf6a6fef7965..583cc29080764 100644 --- a/kernel/rcu/tree.c +++ b/kernel/rcu/tree.c @@ -4595,11 +4595,15 @@ void rcutree_migrate_callbacks(int cpu) struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu); bool needwake;
- if (rcu_rdp_is_offloaded(rdp) || - rcu_segcblist_empty(&rdp->cblist)) - return; /* No callbacks to migrate. */ + if (rcu_rdp_is_offloaded(rdp)) + return;
raw_spin_lock_irqsave(&rcu_state.barrier_lock, flags); + if (rcu_segcblist_empty(&rdp->cblist)) { + raw_spin_unlock_irqrestore(&rcu_state.barrier_lock, flags); + return; /* No callbacks to migrate. */ + } + WARN_ON_ONCE(rcu_rdp_cpu_online(rdp)); rcu_barrier_entrain(rdp); my_rdp = this_cpu_ptr(&rcu_data);