Re: fs/locks.c BKL removal

[Andrew Morton]

Dave Hansen wrote:
> 
> As Linus pointed out, a semaphore is probably the wrong way to go.
> The only things that really needs to be protected are the list
> operations themselves.
> 

It was I who put the BKL back into locks.c, much to
Matthew's disgust...

The problem was that replacing the BKL with a semaphore
seriously damaged Apache thoughput on 8-way.  Apache
was using flock()-based synchronisation and replacing
a spin with a schedule just killed it.

So.. Apache isn't doing that any more, but it is an
instructive case.  Replacing the BKL with a semaphore
can sometimes be a very bad thing.

See http://www.uwsg.iu.edu/hypermail/linux/kernel/0010.3/ -
search for "scalability"

-