2.6.5-rc3-mm4 tmpfs, free and free memory reporting

2.6.5-rc3-mm4 tmpfs, free and free memory reporting

[CaT]

I noticed that the OoM killer was being rather brutal to my tasks
even though free was reporting that I had 150meg of ram left. It wasn't
until much later that I realise that I have tmpfs being used as /tmp
and I checked that. Once I cleaned that up a big all was well. The
hassle was, the memory used by tmpfs was being reported as being used
by the cache. That may be so internally but shouldn't it be reported
as actually used ram as it cannot be dumped for processes like a normal
disk cache can and therefore cannot be considered to be 'free' ram.

-- 
    Red herrings strewn hither and yon.

Re: 2.6.5-rc3-mm4 tmpfs, free and free memory reporting

[Hugh Dickins]

On Sun, 25 Apr 2004, CaT wrote:
> I noticed that the OoM killer was being rather brutal to my tasks
> even though free was reporting that I had 150meg of ram left. It wasn't
> until much later that I realise that I have tmpfs being used as /tmp
> and I checked that. Once I cleaned that up a big all was well. The
> hassle was, the memory used by tmpfs was being reported as being used
> by the cache. That may be so internally but shouldn't it be reported
> as actually used ram as it cannot be dumped for processes like a normal
> disk cache can and therefore cannot be considered to be 'free' ram.

If you have swap enabled (do you?) then tmpfs pages get written out
to swap under memory pressure, and so it is like normal disk cache.  

If you don't have swap enabled, yes, there's nowhere else for it to
go; but I'd say perhaps you were then unwise to allow so much of your
memory to be used for tmpfs mounts - not been bumping up that nice
size=50% have you ;-?

But there's certainly scope for suspicion, as to whether the vmscan
algorithms deal effectively with sending tmpfs to swap.  Should
page_mapping_inuse be so reluctant to write out tmpfs swap?  Should
shmem_writepage call swap_writepage directly instead of giving tmpfs
pages another go around the lists?  (Andrea recently found 2.4 cases
which convinced him it should call swap_writepage directly.)

We've made no change in the face of those doubts because nobody was
complaining of current behaviour; and it makes some sense to be a
little reluctant to swap out tmpfs pages - it is supposed to be a
ram-based filesystem, after all.  But if complaints do accumulate,
let's look into changing some decisions there.

Hugh

Re: 2.6.5-rc3-mm4 tmpfs, free and free memory reporting

[CaT]

On Sun, Apr 25, 2004 at 05:50:58PM +0100, Hugh Dickins wrote:
> On Sun, 25 Apr 2004, CaT wrote:
> > I noticed that the OoM killer was being rather brutal to my tasks
> > even though free was reporting that I had 150meg of ram left. It wasn't
> > until much later that I realise that I have tmpfs being used as /tmp
> > and I checked that. Once I cleaned that up a big all was well. The
> > hassle was, the memory used by tmpfs was being reported as being used
> > by the cache. That may be so internally but shouldn't it be reported
> > as actually used ram as it cannot be dumped for processes like a normal
> > disk cache can and therefore cannot be considered to be 'free' ram.
> 
> If you have swap enabled (do you?) then tmpfs pages get written out

Not this time. I had written that in in the first version of the email
but then scrapped that and rewrote and forgot to put it in.

Basically the complaint was that tmpfs was being registered in cached
memory but it wasn't really cache. Now that I'm a bit more awake I
see that as not being that big a deal.

> If you don't have swap enabled, yes, there's nowhere else for it to
> go; but I'd say perhaps you were then unwise to allow so much of your
> memory to be used for tmpfs mounts - not been bumping up that nice
> size=50% have you ;-?

Yes. :) But I would've hit problems anyway as I completely forgot about 
tmpfs when I was trying to figure out why the oom killer was kicking in
and free was reporting high cache usage which would mean that that
ram is available for more permanent use rather then being used by something
that already has it allocated more or less permanently.

> But there's certainly scope for suspicion, as to whether the vmscan
> algorithms deal effectively with sending tmpfs to swap.  Should
> page_mapping_inuse be so reluctant to write out tmpfs swap?  Should

I wouldn't think so. If I haven't used something in my tmpfs mounts
for a while it should be gone to swap as I'd much rather the memory
be used for active caching and other more immediately useful bits.

> We've made no change in the face of those doubts because nobody was
> complaining of current behaviour; and it makes some sense to be a
> little reluctant to swap out tmpfs pages - it is supposed to be a
> ram-based filesystem, after all.  But if complaints do accumulate,
> let's look into changing some decisions there.

Well the above isn't a complaint but rathe rmore of a thought. ;) I
do some unimportant logging to tmpfs so that it's there when I need
it but doesn't actually need to survive a reboot. Chucking that to
swap sooner rather then later would be a good thing for me I reckon.

-- 
    Red herrings strewn hither and yon.