Hi Alan,

On 07/08/2017 06:21 PM, Alan Stern wrote:

Pardon me for barging in, but I found this whole interchange extremely
confusing...

On Sat, 8 Jul 2017, Ingo Molnar wrote:

quoted

* Paul E. McKenney [off-list ref] wrote:

quoted

On Sat, Jul 08, 2017 at 10:35:43AM +0200, Ingo Molnar wrote:

quoted

* Manfred Spraul [off-list ref] wrote:

quoted

Hi Ingo,

On 07/07/2017 10:31 AM, Ingo Molnar wrote:

quoted

There's another, probably just as significant advantage: queued_spin_unlock_wait()
is 'read-only', while spin_lock()+spin_unlock() dirties the lock cache line. On
any bigger system this should make a very measurable difference - if
spin_unlock_wait() is ever used in a performance critical code path.

At least for ipc/sem:
Dirtying the cacheline (in the slow path) allows to remove a smp_mb() in the
hot path.
So for sem_lock(), I either need a primitive that dirties the cacheline or
sem_lock() must continue to use spin_lock()/spin_unlock().

This statement doesn't seem to make sense.  Did Manfred mean to write
"smp_mb()" instead of "spin_lock()/spin_unlock()"?

Option 1:
     fastpath:
         spin_lock(local_lock)
         smp_mb(); [[1]]
         smp_load_acquire(global_flag);
     slow path:
         global_flag = 1;
         smp_mb();
         <spin_unlock_wait_without_cacheline_dirtying>

Option 2:
     fastpath:
         spin_lock(local_lock);
         smp_load_acquire(global_flag)
     slow path:
         global_flag = 1;
         spin_lock(local_lock);spin_unlock(local_lock).

Rational:
The ACQUIRE from spin_lock is at the read of local_lock, not at the write.
i.e.: Without the smp_mb() at [[1]], the CPU can do:
         read local_lock;
         read global_flag;
         write local_lock;
For Option 2, the smp_mb() is not required, because fast path and slow 
path acquire the same lock.

quoted

quoted

quoted

Technically you could use spin_trylock()+spin_unlock() and avoid the lock acquire
spinning on spin_unlock() and get very close to the slow path performance of a
pure cacheline-dirtying behavior.

This is even more confusing.  Did Ingo mean to suggest using
"spin_trylock()+spin_unlock()" in place of "spin_lock()+spin_unlock()"
could provide the desired ordering guarantee without delaying other
CPUs that may try to acquire the lock?  That seems highly questionable.

I agree :-)

--
     Manfred