On Wed, Feb 8, 2012 at 10:41 PM, Alexander Graf <agraf@suse.de> wrote:

>
> On 08.02.2012, at 13:27, Paul Brook wrote:
>
> >> 2012/2/8 Paul Brook <paul@codesourcery.com>
> >>
> >>>>> I suspect we want to replace the arm_load_kernel call with an
> >>>>> arm_linux_loader device with appropriate properties.
> >>>>
> >>>> Ok, so does this mean the machine model would still explicitly
> >>>> instantiate the bootloader device?
> >>>
> >>> Yes.  Bootloaders inherently have machine specific knowledge.  They
> need
> >>> to know ram location, board ID, secondary CPU boot protocols, etc.
> >>> Requiring the user specify all these things separately from the rest of
> >>> the machine description is IMO not acceptable.
> >>
> >> So what im suggesting here is that machines export these properties to a
> >> globally accessible location. Perhaps via the machine opts mechanism?
> Then
> >> we are in a best of both worls situation where machine models do not
> need
> >> bootloader awareness yet bootloaders can still query qemu for ram_size,
> >> smp#, board_id and friends.
> >
> > Hmm, I suppose this might work.  I'm not sure what you think the benefit
> of
> > this is though.  Fact is the machine needs to have bootloader awareness,
> > whether it be instantating an object or setting magic variables.
> > Having devices rummage around in global state feels messy.  I'd much
> rather
> > use actual properties on the device.  IMO changing the bootloader is
> similar
> > complexity to (say) changing a UART. i.e. it's a board-level change not
> an
> > end-user level change.  Board-level changes are something that will
> happen
> > after QOM conversion, i.e. when we replace machine->init with a board
> config
> > file.
>
>
> Yeah, basically the variable flow goes:
>
>  vl.c -> machine_opts -> machine_init() -> device properties ->
> device_init()
>

So that the machine init function that creates the bootloader device
> enumerates the machine_opts (just like is done in Peter's patches) and then
> passes those on to the bootloader device as device properties.
>
>
So in patch 4/4 in Peters series where he adds a new bootloader feature
(the -dtb switch) its done slightly differently, the machine model does not
handle the machine_opts at all, i.e. The machine model has no awareness of
this dtb argument. Instead the arm boot loader directly queries the
machine_opts API itself:

@@ -251,6 +317,9 @@ void arm_load_kernel(CPUState *env, struct
arm_boot_info *info)
        exit(1);
    }

+    info->dtb_filename = qemu_opt_get(qemu_opts_find(
qemu_find_opts("machine"),
+                                                     0), "dtb");
+

There is no path through the machine_init for this particular property.
What I am suggesting is that a similar approach is take for machine model
set properties (such as ram_size), but instead of the command line setting
the props its done by the machine model. The machine model qemu_opt_set()
the ram_size = whatever. Then the bootloader qemu_opt_get()s it. If you did
this for the key properties related to boot then you would remove the need
for machines to have awareness of their boot process.


> The rationale behind machine opts is that they're basically a dynamic
> number of properties for the not-yet-existing machine object.
>
>
> Alex
>
>