Re: Pls apply this spinlock patch to the kernel

[Linus Torvalds]

On Mon, 29 Oct 2001, Juergen Doelle wrote:
>
> I had created a patch for improving the spinlock behavior on IA32 SMP
> systems for file system work load created with dbench
> (ftp://samba.org/pub/tridge/dbench). The work load is a mix of create,
> delete, write, and read operations executed from a scalable number of
> clients. It is mainly handled in buffer cache.

Fair enough. However, I wonder why you didn't just use the existing
(unaligned) types, and then on a lock-by-lock basis just mark them
aligned. That implies no code-changes.

Something like this should do it:

	.. regular "spinlock_t" type

	#define cachealign \
	 __attribute__((section("aligned"),__aligned__(SMP_CACHELINE_SIZE)))

(use a separate section so that subsequent data structures are also
guaranteed to be aligned - otherwise you might get false sharing from
non-aligned data structures that follow this one).

Eh?

Yes, we already try to do something like this, but due to the false
sharing with other stuff it doesn't _guarantee_ an exclusive cacheline.
Sometimes that is what you want (ie once you get the lock, it _can_ be
advantageous to have the hottest data structure associated with the lock
be in the same cacheline)

		Linus