Re: [PATCH] 2.5: push BKL out of llseek

[Andrew Morton]

Robert Love wrote:
> 
> On Tue, 2002-01-29 at 20:26, Andrew Morton wrote:
> 
> > Just a little word of caution here.  Remember the
> > apache-flock-synchronisation fiasco, where removal
> > of the BKL halved Apache throughput on 8-way x86.
> >
> > This was because the BKL removal turned serialisation
> > on a quick codepath from a spinlock into a schedule().
> 
> I feared this too, but eventually I decided it was worth it and
> benchmarks backed that up.  If nothing else this is yet-another-excuse
> for locks that can spin-then-sleep.
> 
> I posted dbench results, which show a positive gain even on 2-way for
> multiple client loads.
> 

But dbench does lots of seeking against *different* files,
so removal of a shared lock will help there.

But an application where multiple CPUs lseek and write
the *same* file could take a hit....

(And where's the locking for (non-atomic) i_size in sys_stat())

-