Re: [RFC] [PATCH v5 0/3] fadvise: support POSIX_FADV_NOREUSE

[Andrea Righi]

On Tue, Feb 14, 2012 at 03:22:20PM -0800, Andrew Morton wrote:
> On Tue, 14 Feb 2012 23:59:22 +0100
> Andrea Righi <andrea@xxxxxxxxxxxxxxx> wrote:
> 
> > On Tue, Feb 14, 2012 at 01:33:37PM -0800, Andrew Morton wrote:
> > > On Sun, 12 Feb 2012 01:21:35 +0100
> > > Andrea Righi <andrea@xxxxxxxxxxxxxxx> wrote:
> > > 
> > > > The new proposal is to implement POSIX_FADV_NOREUSE as a way to perform a real
> > > > drop-behind policy where applications can mark certain intervals of a file as
> > > > FADV_NOREUSE before accessing the data.
> > > 
> > > I think you and John need to talk to each other, please.  The amount of
> > > duplication here is extraordinary.
> > 
> > Yes, definitely. I'm currently reviewing and testing the John's patch
> > set. I was even considering to apply my patch set on top of the John's
> > patch, or at least propose my tree-based approach to manage the list of
> > the POSIX_FADV_VOLATILE ranges.
> 
> Cool.
> 
> > > 
> > > Both patchsets add fields to the address_space (and hence inode), which
> > > is significant - we should convince ourselves that we're getting really
> > > good returns from a feature which does this.
> > > 
> > > 
> > > 
> > > Regarding the use of fadvise(): I suppose it's a reasonable thing to do
> > > in the long term - if the feature works well, popular data streaming
> > > applications will eventually switch over.  But I do think we should
> > > explore interfaces which don't require modification of userspace source
> > > code.  Because there will always be unconverted applications, and the
> > > feature becomes available immediately.
> > > 
> > > One such interface would be to toss the offending application into a
> > > container which has a modified drop-behind policy.  And here we need to
> > > drag out the crystal ball: what *is* the best way of tuning application
> > > pagecache behaviour?  Will we gravitate towards containerization, or
> > > will we gravitate towards finer-tuned fadvise/sync_page_range/etc
> > > behaviour?  Thus far it has been the latter, and I don't think that has
> > > been a great success.
> > > 
> > > Finally, are the problems which prompted these patchsets already
> > > solved?  What happens if you take the offending streaming application
> > > and toss it into a 16MB memcg?  That *should* avoid perturbing other
> > > things running on that machine.
> > 
> > Moving the streaming application into a 16MB memcg can be dangerous in
> > some cases... the application might start to do "bad" things, like
> > swapping (if the memcg can swap) or just fail due to OOMs.
> 
> Well OK, maybe there are problems with the current implementation.  But
> are they unfixable problems?  Is the right approach to give up on ever
> making containers useful for this application and to instead go off and
> implement a new and separate feature?
> 
> > > And yes, a container-based approach is pretty crude, and one can
> > > envision applications which only want modified reclaim policy for one
> > > particualr file.  But I suspect an application-wide reclaim policy
> > > solves 90% of the problems.
> > 
> > I really like the container-based approach. But for this we need a
> > better file cache control in the memory cgroup; now we have the
> > accounting of file pages, but there's no way to limit them.
> 
> Again, if/whem memcg becomes sufficiently useful for this application
> we're left maintaining the obsolete POSIX_FADVISE_NOREUSE for ever.

Yes, totally agree. For the future a memcg-based solution is probably
the best way to go.

This reminds me to the old per-memcg dirty memory discussion
(http://thread.gmane.org/gmane.linux.kernel.mm/67114), cc'ing Greg.

Maybe the generic feature to provide that could solve both problems is
a better file cache isolation in memcg.

Thanks,
-Andrea
--
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/