Re: [PATCH rcu 0/11] Add light-weight readers for SRCU

[Kent Overstreet]

On Tue, Sep 03, 2024 at 03:13:40PM GMT, Paul E. McKenney wrote:
> On Tue, Sep 03, 2024 at 05:38:05PM -0400, Kent Overstreet wrote:
> > On Tue, Sep 03, 2024 at 09:32:51AM GMT, Paul E. McKenney wrote:
> > > Hello!
> > > 
> > > This series provides light-weight readers for SRCU.  This lightness
> > > is selected by the caller by using the new srcu_read_lock_lite() and
> > > srcu_read_unlock_lite() flavors instead of the usual srcu_read_lock() and
> > > srcu_read_unlock() flavors.  Although this passes significant rcutorture
> > > testing, this should still be considered to be experimental.
> > 
> > This avoids memory barriers, correct?
> 
> Yes, there are no smp_mb() invocations in either srcu_read_lock_lite()
> or srcu_read_unlock_lite().  As usual, nothing comes for free, so the
> overhead is moved to the update side, and amplified, in the guise of
> the at least two calls to synchronize_rcu().
> 
> > > There are a few restrictions:  (1) If srcu_read_lock_lite() is called
> > > on a given srcu_struct structure, then no other flavor may be used on
> > > that srcu_struct structure, before, during, or after.  (2) The _lite()
> > > readers may only be invoked from regions of code where RCU is watching
> > > (as in those regions in which rcu_is_watching() returns true).  (3)
> > > There is no auto-expediting for srcu_struct structures that have
> > > been passed to _lite() readers.  (4) SRCU grace periods for _lite()
> > > srcu_struct structures invoke synchronize_rcu() at least twice, thus
> > > having longer latencies than their non-_lite() counterparts.  (5) Even
> > > with synchronize_srcu_expedited(), the resulting SRCU grace period
> > > will invoke synchronize_rcu() at least twice, as opposed to invoking
> > > the IPI-happy synchronize_rcu_expedited() function.  (6)  Just as with
> > > srcu_read_lock() and srcu_read_unlock(), the srcu_read_lock_lite() and
> > > srcu_read_unlock_lite() functions may not (repeat, *not*) be invoked
> > > from NMI handlers (that is what the _nmisafe() interface are for).
> > > Although one could imagine readers that were both _lite() and _nmisafe(),
> > > one might also imagine that the read-modify-write atomic operations that
> > > are needed by any NMI-safe SRCU read marker would make this unhelpful
> > > from a performance perspective.
> > 
> > So if I'm following, this should work fine for bcachefs, and be a nifty
> > small perforance boost.
> 
> Glad you like it!
> 
> > Can I give you an account for my test cluster? If you'd like, we can
> > convert bcachefs to it and point it at your branch.
> 
> Thank you, but I will pass on more accounts.  I have a fair amount of
> hardware at my disposal.  ;-)

Well - bcachefs might be a good torture test; if you patch bcachefs to
use the new API point me at a branch and I'll point the CI at it