On 5/20/21 3:35 PM, Gaetan Rivet wrote: > Joining pthreads makes the caller quiescent. It should register as such, > as joined threads may wait on an RCU callback executing before quitting, > deadlocking the caller.
Hi, Gaetan. This patch doesn't look right to me. The problem is that users of this function has no idea that the quiescent state will be entered by this function. And this is really hard to track down, because it can be called very deep inside some separate part of the code base that user at the top level might not even know about. For example, a lot of call chains in ovsdb-server may lead to xpthread_join called from ovsdb/log.c. Even though ovsdb-server is single-threaded now, insertion of the ovsrcu_quiesce_start() into xpthread_join() will effectively mean that ovsdb-server will never be able to use RCU if it will become multi-threaded someday, e.g. it will not be able to use CMAPs. So, instead of doing that, callers should enter quiescent state before joining, and this should be done at the highest level of a call chain possible. We have the same thing with cond_wait() implementation, you may add similar comment to the join() function as we have for cond_wait(). Best regards, Ilya Maximets. _______________________________________________ dev mailing list d...@openvswitch.org https://mail.openvswitch.org/mailman/listinfo/ovs-dev
