as discussed before:
The root cause for the performance regression is the smp_mb() that was
added into the fast path.

I see two options:
1) switch to full spin_lock()/spin_unlock() for the rare codepath,
  then the fast path doesn't need the smp_mb() anymore.

2) confirm that no arch needs the smp_mb(), then remove it.
  - powerpc is ok after commit
     6262db7c088b ("powerpc/spinlock: Fix spin_unlock_wait()")
  - arm is ok after commit
     d86b8da04dfa ("arm64: spinlock: serialise spin_unlock_wait against 
concurrent lockers")
  - for x86 is ok after commit
     2c6100227116 ("locking/qspinlock: Fix spin_unlock_wait() some more")
  - for the remaining SMP architectures, I don't have a status.

I would prefer the approach 1:
The memory ordering provided by spin_lock()/spin_unlock() is clear.

Attached are patches for approach 1:

- Patch 1 replaces spin_unlock_wait() with spin_lock()/spin_unlock() and
  removes all memory barriers that are then unnecessary.

- Patch 2 adds the hysteresis code: This makes the rare codepath
  extremely rare.
  It also corrects some wrong comments, e.g. regarding switching
  from global lock to per-sem lock (we "must' switch, not we "can"
  switch as written right now).

The patches passed stress-testing.

What do you think?
My initial idea was to aim for 4.10, then we have more time to decide.


Reply via email to