[ Upstream commit 55d4669ef1 ]
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>
4991cb2d43