Re: [OpenBIOS] next-property

[Olivier Danet]

On 25/01/2014 13:58, Mark Cave-Ayland wrote:
> On 03/01/14 19:16, Olivier Danet wrote:
> > On 03/01/2014 17:09, Artyom Tarasenko wrote:
> > > On Thu, Jan 2, 2014 at 2:43 AM, Olivier Danet 
> > > wrote:
> > > > I may have found a bug in OpenBIOS.
> > > >
> > > > An OS uses the "nextprop" iterator to scan device properties, and
> > > > eventually
> > > > asks about nonexistant properties.
> > > > "const char *obp_nextprop(int node, const char *name)"
> > > > (arch/sparc32/romvec.c)
> > > > calls the "next-property" forth code, then pops the result.
> > > > According to the standard, next-property should return 0 if the 
> > > > property
> > > > selected
> > > > is the last _OR_ if the property does not exist.
> > > > The current code in forth/admin/devices.fs returns nothing when the
> > > > property
> > > > does
> > > > not exit.
> > > >
> > > > For example (SPARC 32) :
> > > >
> > > > cd screen
> > > > " name" ?active-package next-property drop cr type
> > > > --> returns "device_type"
> > > >
> > > > " interrupts" ?active-package next-property
> > > > --> returns 0, this is the last property.
> > > >
> > > > " bozo" ?active-package next-property
> > > > --> returns nothing. Which, imho, is wrong.
> > >
> > > Actually this matches pretty much with what OBP does:
> > >
> > > ok showstack
> > > ok " model" current-device next-property
> > > fffe61e8 4 ffffffffffffffff ok drop type
> > > name
> > > ok " bozo" current-device next-property
> > > ok
> > >
> > > Artyom
> > > > Here is a patch, I am not quite sure about it (debuting Forth)
> > > > but it helps the target OS :
> > > > Index: forth/device/property.fs
> > > >
> > > > --- forth/device/property.fs (révision 1246)
> > > > +++ forth/device/property.fs (copie de travail)
> > > > @@ -70,7 +70,7 @@
> > > > 2drop r> >dn.properties @
> > > > else
> > > > r> find-property dup if @ then
> > > > ?dup if >prop.next @ then
> > > > dup if >prop.next @ then
> > > >
> > > > then
> > > > ?dup if
> > > >
> > > > The standard asks for a pointer to a zero lenght string instead of 
> > > > zero,
> > > > the C code obp_nextprop() expects a zero. I don't know if it makes a
> > > > difference.
> > > >
> > > > Olivier
> > > > (Btw, the OS is NetBSD :-)
> >
> > You are right.
> > I tried on an Ultra 10 with OpenBoot 3.12.
> >
> > next-property returns nothing when the property is unknown and also when
> > the property is the last one.
> > This contradicts Sun own documentation "Writing FCode 3.x Programs".
> >
> > Puzzled, I tried then on a Apple iBook G3 :
> > dev screen
> > " <something>" active-package next-property
> >
> > This laptop returns :
> > true and a forth string when the property is not the last one
> > true and "0 0" when the property is the last one
> > false when the property does not exist.
> >
> > Anyway, whatever method is chosen (return nothing, return zero...),
> > OpenBIOS forth and/or C code ("opb_nextprop()") should be modifed to
> > avoid page faults.
>
> Hi Olivier,
>
> Do you have a revised version of this patch based upon the discussion 
> in this thread?
>
> I'm wondering whether we should simply use a [IFDEF] SPARC32 ... 
> [ENDIF] block so that your change only takes effect on SPARC32?
>
> ATB,
>
> Mark.

When starting X under NetBSD 6.x, it enumerates devices and
eventually asks next-property of "name" and "device_type"
  to many nodes, including "/chosen", "/builtin"...

I did not look into NetBSD to determine whether this behaviour is sane.
I have found several 'almost bugs' in NetBSD. For example the TCX 
framebuffer
should not have an address property.

Sun's OpenBOOT does not define "next-property" on 32bits, any
behavior is acceptable on that platform.

Should we make Sparc64's behaviour (silence) the default one? Answering zero
seemed more polite and simpler for the C obp_ wrapper.

I have a dozen patches waiting for OpenBIOS and QEMU.
I hope I will be able to post them next week.

Regards
Olivier