On Mon, Jan 30, 2012 at 6:42 AM, Edgar E. Iglesias
<address@hidden> wrote:
On Sat, Jan 28, 2012 at 11:48:37AM -0700, Grant Likely wrote:
> On Fri, Jan 27, 2012 at 10:34:01PM +0000, Paul Brook wrote:
> > > If compiled with CONFIG_FDT, allow user to specify a device tree file using
> > > the -dtb argument. If the machine supports it then the dtb will be loaded
> > > into memory and passed to the kernel on boot.
> >
> > Adding annother machine feels wrong. Why does the board specific code need to
> > know about this at all? You already going it via a global variable, so can't
> > this be entirely contained within arm_boot.c?
>
> Mostly because the infrastructure isn't yet in place to pass the .dtb file
> through to the arm_boot code (or maybe it is; what is the best way to pass
> command line data through to the arm_boot.c code (or similar for other
> architectures)?
>
> > If the board file is involved, why is it asking the user?
>
> There is a lot of configuration in the .dts file that the QEMU user may want
> to manipulate; particularly when using QEMU for testing embedded platforms.
> The direction I want to go is to select the machine model based on the top
> level DT compatible property (making -M optional), and then also allow a lot
> of the machine layout to be driven by DT data. ie. populate emulated devices
> from DT data.
>
> I believe this is how Edgar is using the microblaze model, but I don't
> think those patches have been upstreamed yet. I hope that microblaze,
> ARM and powerpc can all use the same model here.
Yes, that's right.
> >
> > > + versatile_init(ram_size,
> > > + boot_device,
> > > + kernel_filename, kernel_cmdline,
> > > + initrd_filename, cpu_model, 0xffffffff);
> >
> > This only works because we're currently too dumb to emulate the differences
> > between the two board variants.
>
> Yeah, this is a hack so I could play with forcing the machine id.
> I'll remove it.
>
> > What we probably want to be doing is shipping/constructing device trees for
> > the boards we implement, with an option to turn this on/off. Requiring a user
> > to invent their own seems deeply sub-optimal given we know exactly what
> > hardware we're emulating. A user that needs to provide their own FDT seems
> > like a fairly rare corner case.
>
> I disagree. QEMU may want to ship stock .dts files, but it will
> absolutely be a common use case for embedded developers to pass in
> their own .dtb file.
>
> > This gets slightly more interesting when you have custom machine variants
> > (i.e. once we fix the object model, and have proper dynamic machine
> > construction). Even then I'd expect the FDT to be derived from/specificed by
> > the machine description, not a separate option.
>
> I started with going down that route, but switched to this model after
> playing with it and noticing that it doesn't seem to fit as well for
> embedded development as providing a .dtb file and having QEMU
> construct a machine that matches the data.
Another major issue with that is DTBs have little bits of software in them as well, such as the "chosen" node, which would need to be populated before passing off to software. Coding these dtb properties into machine models would be sub-optimal, you would need some way of parameterizing certain dtb properties at boot time. The easiest approach to this is just passing in your own dtb file.
That's the way I see it too. Our use-case is to build the virtual machine
from the dtb. We then run the exact same software that runs on real hw.
I don't know the details on how practical dtb's are for describing more
complex bus hierarchies, maybe Peter has more info on that.
Dtbs are capable of describing pretty much anything you want, the main issue is ambiguity on how the dts data structure is used to describe complex or irregular bus heirachies. Im our approach, device models themselves describe how dts syntax is interpreted wrt to machine creation. This means different device models can interpret dts node handles or subnode relationships in any arbitrary way, which would proivde you a backend with which you could implement complex bus heirachies. For example, a device tree node with compatibility "simple-bus" will instantiate all its subnodes as elements on the parent bus (or system bus in the case of the dtb root node). There is however a software hook which would allow you to implement more complex behaviours such as address range remapping for bus bridges (based on the ranges property). We didnt do this, as the memory API wasnt adequate at the time, but recent updates would now make this possible. As another example, a USB controller could interpret its children as USB devices and create the corresponding USB device model interconnects. We successfully prototyped this idea out of tree for SPI busses, but the mechanism is there to do USB, PCI, SD and the other usual suspects. Just needs to be implemented and the framework reworked for QOM.
Cheers,
Edgar