Re: New filesystem for Linux

["Paul E. McKenney" <paulmck@us.ibm.com>]

On Thu, Nov 02, 2006 at 03:15:33PM -0800, Linus Torvalds wrote:
> 
> 
> On Thu, 2 Nov 2006, Mikulas Patocka wrote:
> > 
> > * There is a rw semaphore that is locked for read for nearly all operations
> > and locked for write only rarely. However locking for read causes cache line
> > pingpong on SMP systems. Do you have an idea how to make it better?

[ . . . ]

> (Seqlocks could be changed to drop the first requirement, although it 
> could cause some serious starvation issues, so I'm not sure it's a good 
> idea. For RCU the atomic nature is pretty much designed-in.)

I can't help putting in a plug for SRCU, which is in the 2.6.19-rc series,
and which allows readers to sleep.  (http://lwn.net/Articles/202847/)

SRCU allows readers and writers to run concurrently (as do all forms of
RCU).  If this is a problem, it might be worth looking into Gautham
Shenoy's reader-writer lock built on top of RCU.  (A version for hotplug
may be found at http://lkml.org/lkml/2006/10/26/73.)  This approach
keeps the reader-writer-lock semantics, but gets rid of cache thrashing.
That said, writers have to wait for a grace period.

And as Linus pointed out, if you have disk I/O involved, you probably
won't notice normal reader-writer-lock overhead.

							Thanx, Paul