On Thu, Oct 6, 2016 at 5:08 AM, Bart Van Assche
[off-list ref] wrote:

On 10/05/2016 12:11 PM, Sagi Grimberg wrote:

quoted

I was referring to weather we can take srcu in the submission path
conditional of the hctx being STOPPED?

Hello Sagi,

Regarding run-time overhead:
* rcu_read_lock() is a no-op on CONFIG_PREEMPT_NONE kernels and is
  translated into preempt_disable() with preemption enabled. The latter
  function modifies a per-cpu variable.
* Checking BLK_MQ_S_STOPPED before taking an rcu or srcu lock is only
  safe if the BLK_MQ_S_STOPPED flag is tested in such a way that the
  compiler is told to reread the hctx flags (READ_ONCE()) and if the
  compiler and CPU are told not to reorder test_bit() with the
  memory accesses in (s)rcu_read_lock(). To avoid races
  BLK_MQ_S_STOPPED will have to be tested a second time after the lock
  has been obtained, similar to the double-checked-locking pattern.
* srcu_read_lock() reads a word from the srcu structure, disables
  preemption, calls __srcu_read_lock() and re-enables preemption. The
  latter function increments two CPU-local variables and triggers a
  memory barrier (smp_mp()).

We can use srcu read lock for BLOCKING and rcu read lock for non-BLOCKING,
by putting *_read_lock() and *_read_unlock() into two wrappers, which
should minimize the cost of srcu read lock & unlock and the code is still easy
to read & verify.

Swapping srcu_read_lock() and the BLK_MQ_S_STOPPED flag test will make the
code more complicated. Going back to the implementation that calls
rcu_read_lock() if .queue_rq() won't sleep will result in an implementation
that is easier to read and to verify.

Yeah, I agree.

Thanks,
Ming Lei