Re: [Qemu-devel] [PATCH] exec: silence hugetlbfs warning under qtest

[Michael S. Tsirkin]

On Tue, Nov 24, 2015 at 02:19:36PM +0000, Daniel P. Berrange wrote:
> On Tue, Nov 24, 2015 at 04:10:54PM +0200, Michael S. Tsirkin wrote:
> > On Mon, Nov 23, 2015 at 06:10:08PM +0000, Daniel P. Berrange wrote:
> > > On Mon, Nov 23, 2015 at 07:01:33PM +0100, Marc-André Lureau wrote:
> > > > On Mon, Nov 23, 2015 at 6:40 PM, Paolo Bonzini <address@hidden> wrote:
> > > > > Before: object-initial, chardev, qtest, object-late (not in the patch)
> > > > >
> > > > > After: chardev, qtest, object-initial, object-late (not in the patch)
> > > > >
> > > > > Objects must be initialized before chardev (except rng-egd) since in 
> > > > > the
> > > > > future chardev will need to use objects, in particular secret objects.
> > > > > Was the swap intentional?
> > > > 
> > > > Yes, without the swap, qtest was not initialized before memory is 
> > > > allocated.
> > > > 
> > > > The alternative I could think of is to check the QTEST_QEMU_BINARY
> > > > variable: 
> > > > http://lists.nongnu.org/archive/html/qemu-devel/2015-11/msg01527.html
> > > 
> > > Why do we not simply delete the warning message about the path not
> > > being on hugetlbfs ? ie, why does QEMU try to force a policy that
> > > a memory-file backend has to be on hugetlbfs, as opposed to on
> > > a plain tmpfs ?  I've previously had user request that we allow
> > > use of plain tmpfs, because they want to use vhost-user without
> > > also using hugepages, and that could be done with plain tmpfs.
> > 
> > Because THP does not work on any other filesystem,
> > so many workloads are much slower.
> > That's why it's a warning, not an error.
> 
> AFAICT this warning message is not in a codepath that is specific to
> use of THP. This is just generic code for allocating guest memory
> backed by a file, which does not have any assumption / prerequisite
> that THP is wanted or enabled. So adding warnings that are specifically
> related to THP is inappropriate.
> 
> The fact that THP only works with a hugetlbfs path is merely a
> documentation item to record against the command line option for
> -mem-path.

I'm worried that things go slow and people don't have a way to find out
why, and documentation isn't the first place people look for when facing
a performance issue.

At the moment we call MADV_HUGEPAGE unconditionally and unfortunately
there's no way to report it's not having the intended effect.

Maybe we want a "don't enable hugepages" flag.
That could also have the effect of suppressing the warnings.

> 
> Regards,
> Daniel
> -- 
> |: http://berrange.com      -o-    http://www.flickr.com/photos/dberrange/ :|
> |: http://libvirt.org              -o-             http://virt-manager.org :|
> |: http://autobuild.org       -o-         http://search.cpan.org/~danberr/ :|
> |: http://entangle-photo.org       -o-       http://live.gnome.org/gtk-vnc :|