From mboxrd@z Thu Jan  1 00:00:00 1970
From: Eric Dumazet <eric.dumazet@gmail.com>
Subject: Re: [PATCH 15/17] fs: inode per-cpu last_ino allocator
Date: Thu, 30 Sep 2010 07:36:22 +0200
Message-ID: <1285824982.5211.675.camel@edumazet-laptop>
References: <1285762729-17928-1-git-send-email-david@fromorbit.com>
	 <1285762729-17928-16-git-send-email-david@fromorbit.com>
	 <20100929215312.5fcb6976.akpm@linux-foundation.org>
Mime-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: QUOTED-PRINTABLE
Cc: Dave Chinner <david@fromorbit.com>, linux-fsdevel@vger.kernel.org,
	linux-kernel@vger.kernel.org
To: Andrew Morton <akpm@linux-foundation.org>
Return-path: <linux-kernel-owner@vger.kernel.org>
In-Reply-To: <20100929215312.5fcb6976.akpm@linux-foundation.org>
Sender: linux-kernel-owner@vger.kernel.org
List-Id: linux-fsdevel.vger.kernel.org

Le mercredi 29 septembre 2010 =C3=A0 21:53 -0700, Andrew Morton a =C3=A9=
crit :
> On Wed, 29 Sep 2010 22:18:47 +1000 Dave Chinner <david@fromorbit.com>=
 wrote:
>=20
> > From: Eric Dumazet <dada1@cosmosbay.com>
> >=20

Please note my new email address, thanks.

> > last_ino was converted to an atomic variable to allow the inode_loc=
k
> > to go away. However, contended atomics do not scale on large
> > machines, and new_inode() triggers excessive contention in such
> > situations.
> >=20
> > Solve this problem by providing to each cpu a per_cpu variable,
> > feeded by the shared last_ino, but once every 1024 allocations.
> > This reduces contention on the shared last_ino, and give same
> > spreading ino numbers than before (i.e. same wraparound after 2^32
> > allocations).
> >=20
> > [npiggin: some extra commenting and use of defines]
> >=20
> > ...
> > =20
> > +#ifdef CONFIG_SMP
> > +#define LAST_INO_BATCH 1024
> > +/*
> > + * Each cpu owns a range of LAST_INO_BATCH numbers.
> > + * 'shared_last_ino' is dirtied only once out of LAST_INO_BATCH al=
locations,
> > + * to renew the exhausted range.
> > + *
> > + * This does not significantly increase overflow rate because ever=
y CPU can
> > + * consume at most LAST_INO_BATCH-1 unused inode numbers. So there=
 is
> > + * NR_CPUS*(LAST_INO_BATCH-1) wastage. At 4096 and 1024, this is ~=
0.1% of the
> > + * 2^32 range, and is a worst-case. Even a 50% wastage would only =
increase
> > + * overflow rate by 2x, which does not seem too significant.
> > + *
> > + * On a 32bit, non LFS stat() call, glibc will generate an EOVERFL=
OW
> > + * error if st_ino won't fit in target struct field. Use 32bit cou=
nter
> > + * here to attempt to avoid that.
> > + */
> > +static DEFINE_PER_CPU(unsigned int, last_ino);
> > +static atomic_t shared_last_ino;
> > +
> > +static unsigned int last_ino_get(void)
> > +{
> > +	unsigned int *p =3D &get_cpu_var(last_ino);
> > +	unsigned int res =3D *p;
> > +
> > +	if (unlikely((res & (LAST_INO_BATCH-1)) =3D=3D 0))
> > +		res =3D (unsigned int)atomic_add_return(LAST_INO_BATCH,
> > +				&shared_last_ino) - LAST_INO_BATCH;
>=20
> May as well remove the "- LAST_INO_BATCH" there, I think.  It'll skew
> the results a tad at startup, but why does that matter?


Because on x86, atomic_add_return(val, ptr) uses xadd() + val

So, using "atomic_add_return(val, ptr) - val" removes one instruction ;=
)

>=20
> > +	*p =3D ++res;
> > +	put_cpu_var(last_ino);
> > +	return res;
> > +}
> > +#else
> > +static unsigned int last_ino_get(void)
> > +{
> > +	static unsigned int last_ino;
> > +
> > +	return ++last_ino;
> > +}
>=20
> This is racy with CONFIG_PREEMPT on some architectures, I suspect.  I=
'd
> suggest conversion to atomic_t with, of course, an explanatory commen=
t ;)
>=20

Thanks, I'll rework the patch !

I am pretty happy to see some interest on this patch serie,
eventually :)