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

[Andrew Morton]

Nigel Gamble wrote:
> 
> On 29 Jan 2002, 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().
> 
> Yes, but the other factor to consider here is why did the extra schedule
> take place at all?  I think this is a actually a scheduler issue, and
> I'm hoping that the new scheduler will behave better in this case.  A
> call to schedule() should not happen unless the woken process has a
> higher priority than the process that did the unlock, but the old
> scheduler evidently always calculated this to be the case.  But we
> really want the process that did the unlock to continue running (until
> the end of its timeslice, if not preempted or blocked before then), just
> as it would when the lock was a spinlock.  It would be interesting to
> see whether the new scheduler gets this right.
> 
> Am I remembering the problem correctly?
> 

I don't think so :)

The problem was that the semaphore was highly contended, so the
losing process was explicitly scheduling away.

This doesn't necessarily mean that it was a long-held lock.  In
this case, it was a short-held lock, but it was also very *frequently*
being held and released.   This is a scenario where a spinlock is
heaps more appropriate than a semaphore.

I don't think we need any locking at all in the default lseek()
path, btw.  Apart from the non-atomic i_size thing, which is
only an issue for 32-bit machines.

-