On Thu, May 03, 2018 at 07:50:41PM +0100, Dr. David Alan Gilbert wrote:
> * Stefan Hajnoczi (stefanha@redhat.com) wrote:
> > On Thu, May 03, 2018 at 04:16:41PM +0200, Cédric Le Goater wrote:
> > > Coming back to the initial motivation that Peter pointed out, would
> > > the goal to be able to run vcpus of different architectures ? It would
> > > certainly be interesting to model a platform, specially if we can 
> > > synchronize the execution in some ways and find timing issues.
> > 
> > Yes, there is demand for heterogenous guests.  People have worked on
> > this problem in the past but nothing mergeable came out of it.
> > 
> > The main issue with a monolithic binary is that today, QEMU builds
> > target-specific object files (see Makefile.target).  That means the same
> > C source file is recompiled for each target with different #defines.  We
> > cannot easily link these "duplicate" object files into a single binary
> > because the symbol names would collide.
> > 
> > It would be necessary to refactor target-specific #ifdefs.  Here is a
> > trivial example from arch_init.c:
> > 
> >   #ifdef TARGET_SPARC
> >   int graphic_width = 1024;
> >   int graphic_height = 768;
> >   int graphic_depth = 8;
> >   #else
> >   int graphic_width = 800;
> >   int graphic_height = 600;
> >   int graphic_depth = 32;
> >   #endif
> > 
> > This can be converted into a runtime check:
> > 
> >   static void init_graphic_resolution(void)
> >   {
> >       if (arch_type == QEMU_ARCH_SPARC) {
> >           graphic_width = 1024;
> >           graphic_height = 768;
> >           graphic_depth = 8;
> >       } else {
> >           graphic_width = 800;
> >           graphic_height = 600;
> >           graphic_depth = 32;
> >       }
> >   }
> >   target_init(init_graphic_resolution)
> > 
> > I'm assuming target_init() registers functions that will be called once
> > arch_type has been set.
> > 
> > Of course the meaning of arch_type in a heterogenous system is different
> > since there can be multiple CPUs.  It would mean the overall board (e.g.
> > a SPARC machine).  But that is a separate issue and can only be
> > addressed once target-specific files have been eliminated.
> > 
> > I also want to point out that a monolithic binary is totally orthogonal
> > to modularity (reducing attack surface, reducing dependencies).  These
> > two issues do not conflict with each other.  We could have a single
> > "qemu-softmmu" binary that dynamically loads needed machine types, CPU,
> > and emulated devices.  That way the monolithic binary can do everything
> > but is still minimal.
> > 
> > So to clarify, three separate steps:
> > 
> > 1. Get rid of target-specific #ifdefs
> > 2a. Modular QEMU, single binary
> > 2b. Heterogenous QEMU
> > 
> > 2a and 2b are independent but both depend on 1.
> 
> (1) may not be required, if those ifdef's are in target-specific
> builds; but that does require that any target-specific stuff goes
> through a well defined interface to be a loadable module and not
> have a zillion overlapping symbols.

Right, there are many cases that I omitted and the arch_init.c example
was the simplest case.  Each instance needs to be handled on a
case-by-case basis.

We may need to keep multiple copies of object code.  Byteswapping is an
example of this:

  #if defined(HOST_WORDS_BIGENDIAN) != defined(TARGET_WORDS_BIGENDIAN)
  #define BSWAP_NEEDED
  #endif

Adding runtime endianness checks to all guest memory accesses could
impact performance.  This is why everything that includes cpu.h is
currently per-target.

This work is no small task but I think it's necessary.  The starting
point for anyone who'd like to contribute:

Look at obj-y files in the makefiles.  These are the target-specific
object files that need to be investigated.  Everything already in
common-obj-y is only built once and can be ignored.

Stefan