Re: [Qemu-devel] [PULL 02/29] ppc/spapr: Receive and store device tree blob from SLOF

[Greg Kurz]

On Wed, 27 Mar 2019 19:38:27 +1100
David Gibson <address@hidden> wrote:

> On Wed, Mar 27, 2019 at 08:09:26AM +0100, Greg Kurz wrote:
> > On Wed, 27 Mar 2019 11:56:10 +1100
> > David Gibson <address@hidden> wrote:
> >   
> > > On Tue, Mar 26, 2019 at 08:09:53AM +0100, Greg Kurz wrote:  
> > > > On Tue, 26 Mar 2019 10:47:15 +1100
> > > > David Gibson <address@hidden> wrote:
> > > >     
> > > > > On Mon, Mar 25, 2019 at 05:33:21PM +0100, Greg Kurz wrote:    
> > > > > > On Mon, 25 Mar 2019 11:53:47 +1100
> > > > > > David Gibson <address@hidden> wrote:
> > > > > >       
> > > > > > > On Sun, Mar 24, 2019 at 12:03:54AM -0400, Brad Smith wrote:      
> > > > > > > > Now that I am checking out 4.0.0 rc's I see this diff is broken 
> > > > > > > > and
> > > > > > > > depends on a function libfdt does not expose. The breakage is
> > > > > > > > hidden by the fallback check in the configure script.        
> > > > > > > 
> > > > > > > Ah, bother.  That keeps happening, unfortunately.  I think it's
> > > > > > > because so many people use libfdt embedded, rather than as a 
> > > > > > > shared
> > > > > > > library that we tend not to notice.
> > > > > > >       
> > > > > > 
> > > > > > It's a bit more complicated. I do have latest libfdt packages on my 
> > > > > > laptop:
> > > > > > 
> > > > > > libfdt-1.4.7-2.fc28.x86_64
> > > > > > libfdt-devel-1.4.7-2.fc28.x86_64
> > > > > > 
> > > > > > but I still end up using the embedded one and the build doesn't spot
> > > > > > the missing symbols.      
> > > > > 
> > > > > Sorry, I wasn't clear.  I wasn't meaning in the context of qemu, but
> > > > > for dtc generally.  A large portion of the users are things like
> > > > > u-boot that have to use dtc embedded, rather than as a shared
> > > > > library.  That's why we tend not to notice missing symbols from the
> > > > > version script upstream.
> > > > >     
> > > > 
> > > > Ok, I get it.
> > > >     
> > > > > > This happens because of several reasons:
> > > > > > 
> > > > > > 1) configure unconditionally falls back to embedded if an error 
> > > > > > occurs with
> > > > > >    the system packages. And, as reported by Brad, the current 1.4.7 
> > > > > > packages
> > > > > >    are broken indeed:
> > > > > > 
> > > > > > $ objdump -T /usr/lib64/libfdt-1.4.7.so | grep fdt_check_full
> > > > > > $ 
> > > > > > 
> > > > > > 2) when building embedded, we only build the archive, not the 
> > > > > > shared lib.
> > > > > >       
> > > > > > > I guess we should figure out how to force the testsuite to link
> > > > > > > against the shared library rather than static to test for this.  
> > > > > > > I'll
> > > > > > > look into it if I have time (which is a big if).
> > > > > > >       
> > > > > > 
> > > > > > I think we should rather build the embedded shared library using
> > > > > > the 'libfdt' rule of the top-level makefile of the dtc sub-module
> > > > > > and have QEMU to be linked against this share library instead of
> > > > > > the static one. AFAIK, this is what gcc does when it finds both
> > > > > > .a and .so.      
> > > > > 
> > > > > Actually, I don't think this is a good idea.  It means the resulting
> > > > > qemu build would have to be installed with the libfdt image as well.
> > > > > As well as complicating the install path, that means that the qemu
> > > > > build will now actively conflict with a packaged libfdt, rather than
> > > > > merely suboptimally failing to use it.    
> > > > 
> > > > Yes you're right: the resulting QEMU shouldn't be installed, especially
> > > > if it has a RPATH poiting to the build directory.
> > > > 
> > > > This being said, if someone wants to build AND install QEMU, shouldn't
> > > > she rely exclusively on installed libfdt packages ? In other words,
> > > > shouldn't the embedded libfdt be a QEMU developper only thing ? What
> > > > are the real life use cases for embedded libfdt ?    
> > > 
> > > I don't think we should insist on that, although it's certainly the
> > > way distros will usually work.  If someone wants to build and install
> > > qemu locally, I don't think we should insist they first work out how
> > > to install a new enough libfdt for it to use.
> > > 
> > > Likewise a limited purpose distro for whom qemu is the only user of
> > > libfdt might not want to package it separately.
> > >   
> > 
> > Fair enough. There's still one minor fix to make in configure for
> > the tarball case though:
> > 
> >           # Not a git build & no libfdt found, prompt for system install
> >           error_exit "DTC (libfdt) version >= 1.4.2 not present." \
> >                      "Please install the DTC (libfdt) devel package"
> > 
> > Which version should that be ? Do you plan to release 1.5.0 or
> > whatever anytime soon ?  
> 
> I released 1.5.0 about 3 weeks ago.
> 

Hmm... I do see the version bumping patch in https://github.com/dgibson/dtc/
but no trace of 1.5.0 in https://github.com/dgibson/dtc/releases... so
I'm not sure if 1.5.0 is actually released. If it is released indeed, then
I guess we need a 1.5.1 with the version.lds fix.

pgphQcUwn6b7p.pgp
Description: OpenPGP digital signature