The current code can lose RCU callbacks at shutdown time, which can
result in hangs. This lossage can happen as follows:
o A thread invokes call_rcu_data_free(), which executes up through
the wake_call_rcu_thread(). At this point, the call_rcu_data
structure has been drained of callbacks, but is still on the
call_rcu_data_list. Note that this thread does not hold the
call_rcu_mutex.
o Another thread invokes rcu_barrier(), which traverses the
call_rcu_data_list under the protection of call_rcu_mutex,
a list which still includes the above newly drained structure.
This thread therefore adds a callback to the newly drained
call_rcu_data structure. It then releases call_rcu_mutex and
enters a mystifying loop that does futex stuff.
o The first thread finishes executing call_rcu_data_free(),
which acquires call_rcu_mutex just long enough to remove the
newly drained call_rcu_data structure from call_rcu_data_list.
Which causes one of the rcu_barrier() invocation's callbacks to
be leaked.
o The second thread's rcu_barrier() invocation never returns
resulting in a hang.
This commit therefore changes call_rcu_data_free() to acquire
call_rcu_mutex before checking the call_rcu_data structure for callbacks.
In the case where there are no callbacks, call_rcu_mutex is held across
both the check and the removal from call_rcu_data_list, thus preventing
rcu_barrier() from adding a callback in the meantime. In the case where
there are callbacks, call_rcu_mutex must be momentarily dropped across
the call to get_default_call_rcu_data(), which can itself acquire
call_rcu_mutex. This momentary drop is not a problem because any
callbacks that rcu_barrier() might queue during that period of time will
be moved to the default call_rcu_data structure, and the lock will be
held across the full time including moving those callbacks and removing
the call_rcu_data structure that was passed into call_rcu_data_free()
from call_rcu_data_list.
With this fix, a several-hundred-CPU test successfully completes more
than 5,000 executions. Without this fix, it fails within a few tens
of executions. Although the failures happen more quickly on larger
systems, in theory this could happen on a single-CPU system, courtesy
of preemption.
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: Stephen Hemminger <stephen@networkplumber.org>
Cc: Alan Stern <stern@rowland.harvard.edu>
Cc: Lai Jiangshan <jiangshanlai@gmail.com>
Cc: <lttng-dev@lists.lttng.org>
Cc: <linux-kernel@vger.kernel.org>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0)
(void) poll(NULL, 0, 1);
}
+ call_rcu_lock(&call_rcu_mutex);
if (!cds_wfcq_empty(&crdp->cbs_head, &crdp->cbs_tail)) {
- /* Create default call rcu data if need be */
+ call_rcu_unlock(&call_rcu_mutex);
+ /* Create default call rcu data if need be. */
+ /* CBs queued here will be handed to the default list. */
(void) get_default_call_rcu_data();
+ call_rcu_lock(&call_rcu_mutex);
__cds_wfcq_splice_blocking(&default_call_rcu_data->cbs_head,
&default_call_rcu_data->cbs_tail,
&crdp->cbs_head, &crdp->cbs_tail);
wake_call_rcu_thread(default_call_rcu_data);
}
- call_rcu_lock(&call_rcu_mutex);
cds_list_del(&crdp->list);
call_rcu_unlock(&call_rcu_mutex);