virtio: move bi-endian target support to a single location

Commit Message

Greg Kurz May 31, 2016, 2:10 p.m. UTC

On Tue, 31 May 2016 15:15:21 +0200
Paolo Bonzini <[e-mail]> wrote:

> On 31/05/2016 15:10, Greg Kurz wrote:
> >>> > > +#if defined(TARGET_PPC64) || defined(TARGET_ARM)
> >>> > > +#define LEGACY_VIRTIO_IS_BIENDIAN 1
> >>> > > +#endif    
> >> > 
> >> > These will only be correct if something else includes cpu.h.  Instead of  
> > Unless I missed something, the TARGET_* macros come from the generated
> > config-target.h header, which is in turn included by qemu/osdep.h and
> > thus included by most of the code.  
> 
> You're right.  Problems _could_ happen if virtio-access.h is included in
> a file compiled without -DNEED_CPU_H (i.e. with common-obj-y instead of
> obj-y) but include/exec/poison.h should take care of that.
> 

Exactly !

> >> > defining this, you should add
> >> > 
> >> > #include "cpu.h"
> >> > 
> >> > at the top of include/hw/virtio-access.h and leave the definitions in
> >> > target-*/cpu.h.
> >> >   
> > All this bi-endian stuff is really an old-virtio-only thing... it is
> > only to be used by virtio_access_is_big_endian(). The fact that it
> > broke silently with your cleanup series is yet another proof that
> > this workaround is fragile.  
> 
> It is not fragile actually.  cpu.h doesn't exist in common-obj-y, so the
> TARGET_IS_BIENDIAN define can be safely taken from cpu.h.
> 

I was referring to the current situation where virtio_access_is_big_endian()
silently lost its ability to support endian changing guests with:

[    8.959600] blk-mq: failed to allocate request map
[    8.960171] virtio_blk: probe of virtio0 failed with error -12

> Anyway because of poison.h your solution isn't fragile either, so
> 

At least, if someone tries to include virtio-access.h in common-obj-y, she
gets a build break.

> Reviewed-by: Paolo Bonzini <[e-mail]>
> 
> Paolo
> 

Thanks !

--
Greg

Patch

--- a/include/qemu-common.h
+++ b/include/qemu-common.h
@@ -22,11 +22,6 @@ 
 
 #include "qemu/option.h"
 
-/* FIXME: Remove NEED_CPU_H.  */
-#ifdef NEED_CPU_H
-#include "cpu.h"
-#endif /* !defined(NEED_CPU_H) */
-

which then leads to the usual ENOMEM errors in the guest: