[libvirt] [PATCH] nodeinfo: fix detection of physical memory on uclibc/musl libc

[Natanael Copa]

On Tue, 15 Apr 2014 12:30:47 +0100
"Daniel P. Berrange" <berrange at redhat.com> wrote:

> On Tue, Apr 15, 2014 at 01:26:27PM +0200, Natanael Copa wrote:
> > On Tue, 15 Apr 2014 10:37:40 +0100
> > "Daniel P. Berrange" <berrange at redhat.com> wrote:
> > 
> > > On Tue, Apr 15, 2014 at 11:31:23AM +0200, Natanael Copa wrote:
> > > > The gnulib's physmem_total will as a fallback report 64MB as total
> > > > memory if sysconf(_SC_PHYS_PAGES) is unimplemented on linux. This makes
> > > > it impossible to detect if physmem_total works or not, so we try first
> > > > the linux only sysinfo(2) before falling back to gnulibs physmem_total.
> > > > 
> > > > This makes the total memory be correctly reported on musl libc and
> > > > uclibc.
> > > > 
> > > > Signed-off-by: Natanael Copa <ncopa at alpinelinux.org>
> > > > ---
> > > >  configure.ac   |  2 +-
> > > >  src/nodeinfo.c | 22 ++++++++++++++++------
> > > >  2 files changed, 17 insertions(+), 7 deletions(-)
> > > 
> > > I think you should instead patch the gnulib physmem_total() function
> > > to support use of  sysinfo() itself. That way all the many users of
> > > GNULIB will work correctly on musl/uclibc, rather than just libvirt.
> > 
> > the intention was to fix all 3 places:
> > 
> > * fix libvirt not trust physmem_total
> > * fix gnulib to use sysinfo and fallback to sysconf
> > * add _SC_PHYS_PAGES support to musl libc
> > 
> > In any case, the gnulib's physmem_total will most likely always fall
> > back to 64MB instead of fail with error since it looks like thats the
> > way it is designed. I doubt they will accept fix for that.
> > 
> > That means that libvirt cannot really trust physmem_total and IMHO it
> > should be avoided.
> 
> The whole point of physmem_total is that libvirt can avoid having to
> have a bunch of different implementations. Any kind of fix that libvirt
> could do, can equally be done in the physmem_total, so there's no reason
> to not rely on physmem_total. So we only need to fix physmem_total, and
> the musl libc.
> 
> The only reason to fix libvirt, is if there's some problem that would
> prevent us getting the fix into gnulib in an acceptable timeframe.

Fair enough.

I posted a bug report and a patch to bugs-gnulib.

-nc