On Thu, 2018-12-20 at 14:34 -0700, Jens Axboe wrote:

Yeah, I don't think it's bullet proof either, it just closes the gap.
I'm fine with fiddling with the tag iteration. On top of what I sent, we
could have tag iteration hold the RCU read lock, and then we just need
to ensure that the tags are freed with RCU.

Do you mean using call_rcu() to free tags? Would that require to add a
struct rcu_head to every request? Would it be acceptable to increase the
size of struct request with an rcu_head? Additionally, could that reduce
the queue depth if the time between grace periods is larger than the time
between I/O submissions?

Thanks,

Bart.