Re: [Qemu-devel] [RFC PATCH] arm boot: added QOM device definition

[Peter Crosthwaite]

On Wed, Feb 8, 2012 at 11:35 PM, Alexander Graf <address@hidden> wrote:

On 08.02.2012, at 14:30, Peter Crosthwaite wrote:

On Wed, Feb 8, 2012 at 11:10 PM, Alexander Graf <address@hidden> wrote:

On 08.02.2012, at 14:04, Peter Crosthwaite wrote:

On Wed, Feb 8, 2012 at 10:41 PM, Alexander Graf <address@hidden> wrote:

On 08.02.2012, at 13:27, Paul Brook wrote:

>> 2012/2/8 Paul Brook <address@hidden>
>>
>>>>> 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.

Ah, I see. So he derived from the original proposal, oh well :).

Isn't this the best approach tho? If you changed it over to the other flow, then you would end up with a change pattern where every machine model would need to get and pass the new argument. This way the diff is limited to the command line infrastructure and the bootloader.

If you want the smallest diff, just make things globals and call it a day. This whole thing is a discussion around device architecture, right?

So if we consider this bootloader a device and its -dtb argument a property of that device, then what you are implying is that every device property of every device in a machine must be managed by the machine model? Isn't the dynamic machine model work that is in progress is trying to get away the approach where fixed machine models have to micromanage all their attached devices? with the ultimate goal of -no-machine how will the bootloader get this dtb argument?

 

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.

But that's exactly what we want. We want the machine to be aware of its boot process, because that's the one place that needs to instantiate the boot device, no?

More a case of the reverse isnt it? The bootloader needs to know about the machine its generating code for, but the machine generation process doesnt need to know about the software its going to run. The machine being aware of the bootloader implementation and instantiating it you are putting in place a policy where you forcing a particular bootflow on every user of that machine. The hardware is placing policy on what software its running.

In the case of arm boot you end up with a situation where you are trying to write a bootloader that can handle every possible scenario, what I am suggesting is arm_boot.c as it is becomes a linux specific bootloader and other bootflows such as booting from elfs or raw images (or other unforseen bootflows) are different bootloader devices. The choice of which bootloader you use is driven by which -device you put down on command line.

Hrm. I wouldn't want to have to select different bootloader types using -device. Today, we have autodetect code in almost all targets that checks the binary and figures out what to load from it. So on x86 you just do -kernel foo and if foo is a Linux kernel, it will run it using that loader, while if it's a multiboot image, it will load it through that.

Any reason you can't just upstream your specific bootloader code?

Its currently just a set of hacks to arm_boot.c, but for upstream acceptance you would need to add serveral more options to arm_boot (like -dtb) to be able to do this needed parameteristation. It would be tedious to have to editevery arm platform to pass in a suite of options from the machine model.

 

We can then worry about replacing the loader device at a different point in time and model things easily now.

Alex