Re: swapping when there's a free memory

From: Mikulas Patocka
Date: Fri Apr 30 2010 - 15:13:25 EST




On Tue, 27 Apr 2010, KAMEZAWA Hiroyuki wrote:

> On Mon, 26 Apr 2010 15:33:33 -0700
> Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> wrote:
>
> > On Sun, 25 Apr 2010 09:13:49 +0200
> > Pavel Machek <pavel@xxxxxx> wrote:
> >
> > > Hi!
> > >
> > > > I captured this output of vmstat. The machine was freeing cache and
> > > > swapping out pages even when there was a plenty of free memory.
> > > >
> > > > The machine is sparc64 with 1GB RAM with 2.6.34-rc4. This abnormal
> > > > swapping happened during running spadfsck --- a fsck program for a custom
> > > > filesystem that caches most reads in its internal cache --- so it reads
> > > > buffers and allocates memory at the same time.
> > > >
> > > > Note that sparc64 doesn't have any low/high memory zones, so it couldn't
> > > > be explained by filling one zone and needing to allocate pages in it.
> > >
> > > Fragmented memory + high-order allocation?
> >
> > Yeah, could be. I wonder which slab/slub/slob implementation you're
> > using, and what page sizes it uses for dentries, inodes, etc. Can you
> > have a poke in /prob/slabinfo?

It uses one page-per-slab for dentries and two for inodes. But there was
certainly no dentry or inode-based load --- the machine runs without X
with minimum daemons, there is no major background work. There was just a
process reading 128-kbyte blocks from a raw device and caching them in its
userspace that triggered this. Can it be that kernel uses high-order
allocations for reading from a buffer cache?

> And please /proc/buddyinfo and /proc/zoneinfo when the system is swappy.

It happens rarely, I don't know if I catch it at the right time. The
report I sent, was what I found in a scrollback of vmstat. I didn't catch
it in real time.

> Thanks,
> -Kame

Mikulas
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/