Message-ID: <20081224185302.GA16467@Krystal>
Date:	Wed, 24 Dec 2008 13:53:03 -0500
From:	Mathieu Desnoyers <mathieu.desnoyers@...ymtl.ca>
To:	Rusty Russell <rusty@...tcorp.com.au>
Cc:	David Miller <davem@...emloft.net>, rostedt@...dmis.org,
	akpm@...ux-foundation.org, linux-kernel@...r.kernel.org,
	paulus@...ba.org, benh@...nel.crashing.org,
	linux-ia64@...r.kernel.org, linux-s390@...r.kernel.org,
	Christoph Lameter <christoph@...eter.com>,
	"Paul E. McKenney" <paulmck@...ux.vnet.ibm.com>,
	Martin Bligh <mbligh@...gle.com>
Subject: Re: local_add_return

* Rusty Russell (rusty@...tcorp.com.au) wrote:
> On Tuesday 23 December 2008 05:13:28 Mathieu Desnoyers wrote:
> > > I can be convinced, but I'll need more than speculation.  Assuming
> > > local_long_atomic_t, can you produce a patch which uses it somewhere else?
> > 
> > I had this patch applying over Christoph Lameter's vm tree last
> > February. It did accelerate the slub fastpath allocator by using
> > cmpxchg_local rather than disabling interrupts. cmpxchg_local is not
> > using the local_t type, but behaves similarly to local_cmpxchg.
> > 
> > http://lkml.org/lkml/2008/2/28/568
> 
> OK, I'll buy that.  So we split local_t into a counter and an atomic type.
> 
> > I know that
> > local_counter_long_t and local_atomic_long_t are painful to write, but
> > that would follow the current atomic_t vs atomic_long_t semantics. Hm ?
> 
> OK, I've looked at how they're used, to try to figure out whether long
> is the right thing.  Counters generally want to be long, but I was in doubt
> about atomics; yet grep shows that atomic_long_t is quite popular.  Then
> I hit struct nfs_iostats which would want a u64 and a long.  I don't think
> we want local_counter_u64 etc.
> 
> Just thinking out loud, perhaps a new *type* is the wrong direction?  How
> about a set of macros which take a fundamental type, such as:
> 
> 	DECLARE_LOCAL_COUNTER(type, name);
> 	local_counter_inc(type, addr);
> 	...
> 	DECLARE_LOCAL_ATOMIC(type, name);
> 	local_atomic_add_return(type, addr);
> 
> This allows pointers, u32, u64, long, etc.  If a 32-bit arch can't do 64-bit
> local_counter_inc easily, at least the hairy 64-bit code can be eliminated at
> compile time.
> 
> Or maybe that's overdesign?
> Rusty.

Yeah, I also thought of this, but I am not sure every architecture
provides primitives to modify u16 or u8 data atomically like x86 does.
But yes, I remember hearing Christoph Lameter being interested to use
unsigned char or short atomic counters for the vm allocator in the past.
The rationale was mostly that he wanted to keep a counter in a very
small data type, expecting to "poll" the counter periodically (e.g.
every X counter increment) and sum the total somewhere else.

So I think it would be the right design in the end if we want to allow
wider use of such atomic primitives for counters w/o interrupts
disabled. And I would propose we use a BUILD_BUG_ON() when the
architecture does not support an atomic operation on a specific type.
We should also document which type sizes are supported portably and
which are architecture-specific.

Or, as you say, maybe it's overdesign ? If we have to pick something
simple, just supporting "long" would be a good start.

Mathieu

-- 
Mathieu Desnoyers
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists