On 09/04/2013 12:00 AM, Rob Herring wrote:

On Tue, Sep 3, 2013 at 10:53 AM, Andre Przywara

Hi Rob,

...

a normal Linux kernel currently supports reading the start and end address of a single binary blob via the FDT's /chosen node. This will be interpreted as the location of an initial RAM disk.

The Xen hypervisor itself is a kernel, but needs up to _two_ binaries for proper operation: a Dom0 Linux kernel and it's associated initrd. On x86 this is solved via the multiboot protocol used by the Grub bootloader, which supports to pass an arbitrary number of binary modules to any kernel.

Since in the ARM world we have the versatile device tree, we don't need to implement the mulitboot protocol.

But surely there would be some advantage of reuse by using the multi-boot protocol since Xen, grub, and OS tools already support it for x86.

Yes, but that is x86 only and multiboot is it's nature quite architecture specific. The current(?) multiboot v2 spec has no official ARM support (only x86 and MIPS), so this would need to be "invented" first. While this is technically easy, ARM software currently has no support for multiboot at all: not in u-boot and not in Xen. Multiboot support in Xen lives entirely in the x86 directory, and big parts of it are even in assembly.

I am about to write up a more elaborate technical rationale describing the problems with multiboot on ARM:

https://wiki.linaro.org/AndrePrzywara/Multiboot

...

So I'd like to propose a new binding which denotes binary modules a kernel can use at it's own discretion. The need is triggered by the Xen hypervisor (which already uses a very similar scheme), but the approach is deliberately chosen to be as generic as possible to allow future uses (like passing firmware blobs for devices or the like). Credits for this go to Ian Campbell, who started something very similar [1] for the Xen hypervisor. The intention of this proposal is to make this generic and publicly documented.

Can you describe how you see the boot flow working starting with OS installer writes kernel, initrd, xen and ??? to disk. How does the bootloader know what to load? The OS may not have access to the dtb, so this has to be described to the bootloader as well.

The idea is to use bootscripts (for instance in u-boot) to tackle this. See for an example below.

I don't see how the process would be differ significantly from the current process, where you have to load mostly two images, get hold of the DTB, enter image data into the DTB and launch the kernel. Now you just need to load an additional image and enter it's properties into the DTB, actually in a pretty generic way.

...

Looking forward to any comments!

Thanks, Andre.

[1] http://xenbits.xen.org/gitweb/?p=xen.git%3Ba=blob%3Bf=docs/misc/arm/device-t...

---

• Multiple boot modules device tree bindings

Boot loaders wanting to pass multiple additional binaries to a kernel shall add a node "module" for each binary blob under the /chosen node with the following properties:

• compatible: compatible = "boot,module"; A bootloader may add names to more specifically describe the module, e.g. Xen may use "xen,dom0-kernel" or "xen,dom0-ramdisk". If possible a kernel should be able to use modules even without a descriptive naming, by enumerating them in order and using hard-coded meanings for each module (e.g. first is kernel, second is initrd).

• reg: specifies the base physical address and size of a region in memory where the bootloader loaded the respective binary data to.

• bootargs: An optional property describing arguments to use for this module. Could be a command line or configuration data.

...

Example: /chosen { #size-cells = <0x1>; #address-cells = <0x1>; module@0 { compatible = "xen,linux-zimage", "xen,multiboot-module", "boot,module"; reg = <0x80000000 0x003dcff8>; bootargs = "console=hvc0 earlyprintk ro root=/dev/sda1 nosmp"; }; module@1 { compatible = "xen,linux-initrd", "xen,multiboot-module", "boot,module"; reg = <0x08000000 0x00123456>; };

This has to be created and parsed typically in FDT format by early boot code, and I worry about the complexity this has. Being future proof and extensible is good, but we could meet today's needs with something simple like this:

Parsing this is already done in Xen, for instance. In fact we just look for nodes matching "boot,module" and then check for other names to determine it's type (which is a few-liner patch in Xen which I will post later today). And I don't see a need to load modules that early that we don't have an un-flattened tree available.

Generating is also part of libfdt, in fact this whole subtree above has been generated on the command line of a stock Calxeda U-Boot: dom0kernel=fdt addr ${fdt_addr}; fdt resize; fdt mknod /chosen module@0; fdt set /chosen/module@0 compatible "xen,linux-zimage" "xen,multiboot-module" "boot,module"; fdt set /chosen/module@0 reg <${dom0_addr_r} 0x${filesize}>; fdt set /chosen/module@0 bootargs "console=hvc0 earlyprintk ro root=/dev/sda1 nosmp"

With this you load the Dom0 kernel (via TFTP or ext2load) and do "run dom0kernel" and are done. I have also patches which add an u-boot command called "module" which automates this, but this is mostly syntactic sugar (though may be useful for future abstraction, for instance to support the x86 (or even ARM) "real" multiboot protocol).

bootargs = "xen args --- linux args"; xen,linux-image = <start size>;

So, is having a more generic solution really needed?

Not necessarily needed, but useful, I think. As described above I don't see any technical obstacles of doing it in a more generic way, so we could as well go ahead with this. On x86 from time to time the need for additional binaries pops up (early microcode loading, for instance), so why not be be prepared. Also this approach avoids hard-coding the Xen name into the bootloader, as said in the proposal the meaning could be derived from the order of the modules (as on x86), so a bootloader does not need to know anything about Xen at all.

Regards, Andre.

On 09/05/2013 03:23 PM, Rob Herring wrote:

On Wed, Sep 4, 2013 at 3:43 AM, Andre Przywara andre.przywara@linaro.org wrote:

...

On 09/04/2013 12:00 AM, Rob Herring wrote:

...

On Tue, Sep 3, 2013 at 10:53 AM, Andre Przywara

...

...

But surely there would be some advantage of reuse by using the multi-boot protocol since Xen, grub, and OS tools already support it for x86.

Yes, but that is x86 only and multiboot is it's nature quite architecture specific. The current(?) multiboot v2 spec has no official ARM support (only x86 and MIPS), so this would need to be "invented" first. While this is technically easy, ARM software currently has no support for multiboot at all: not in u-boot and not in Xen. Multiboot support in Xen lives entirely in the x86 directory, and big parts of it are even in assembly.

If so, I cannot find the assembly and most of it seems to be moved out of x86 code now.

There is some code that relocates and maps the multiboot structure: in files in the xen/arch/x86/boot directory, including head.S, reloc.c and setup.c. That would need to be completely rewritten, not only because it deals with multiboot v1, but also because it is scattered throughout the x86 setup code and the initial page setup is different from ARM.

All these things are also done on ARM - but for the device tree already. Of course proper multiboot v2 support is doable - but please compare the effort needed for this against this one: http://lists.xen.org/archives/html/xen-devel/2013-09/msg00359.html

Regards, Andre.

On 09/15/2013 02:37 PM, Grant Likely wrote:

On Tue, 03 Sep 2013 17:53:44 +0200, Andre Przywara andre.przywara@linaro.org wrote:

...

Hi,

a normal Linux kernel currently supports reading the start and end address of a single binary blob via the FDT's /chosen node. This will be interpreted as the location of an initial RAM disk.

The Xen hypervisor itself is a kernel, but needs up to _two_ binaries for proper operation: a Dom0 Linux kernel and it's associated initrd. On x86 this is solved via the multiboot protocol used by the Grub bootloader, which supports to pass an arbitrary number of binary modules to any kernel.

Since in the ARM world we have the versatile device tree, we don't need to implement the mulitboot protocol.

So I'd like to propose a new binding which denotes binary modules a kernel can use at it's own discretion. The need is triggered by the Xen hypervisor (which already uses a very similar scheme), but the approach is deliberately chosen to be as generic as possible to allow future uses (like passing firmware blobs for devices or the like). Credits for this go to Ian Campbell, who started something very similar [1] for the Xen hypervisor. The intention of this proposal is to make this generic and publicly documented.

Looking forward to any comments!

Thanks, Andre.

[1] http://xenbits.xen.org/gitweb/?p=xen.git%3Ba=blob%3Bf=docs/misc/arm/device-t...

---

• Multiple boot modules device tree bindings

Boot loaders wanting to pass multiple additional binaries to a kernel shall add a node "module" for each binary blob under the /chosen node with the following properties:

• compatible: compatible = "boot,module"; A bootloader may add names to more specifically describe the module, e.g. Xen may use "xen,dom0-kernel" or "xen,dom0-ramdisk". If possible a kernel should be able to use modules even without a descriptive naming, by enumerating them in order and using hard-coded meanings for each module (e.g. first is kernel, second is initrd).

• reg: specifies the base physical address and size of a region in memory where the bootloader loaded the respective binary data to.

• bootargs: An optional property describing arguments to use for this module. Could be a command line or configuration data.

Example: /chosen { #size-cells = <0x1>; #address-cells = <0x1>; module@0 { compatible = "xen,linux-zimage", "xen,multiboot-module", "boot,module"; reg = <0x80000000 0x003dcff8>; bootargs = "console=hvc0 earlyprintk ro root=/dev/sda1 nosmp"; }; module@1 { compatible = "xen,linux-initrd", "xen,multiboot-module", "boot,module"; reg = <0x08000000 0x00123456>; };

Grant,

thanks for looking at this. Appreciate your comments.

The moment you pull in a 'reg' property, your required to property parse #size-cells and #address-cells and also 'ranges' which is missing from the above.

You mean that early boot code has to read #size-cells and #address-cells and act accordingly when parsing the reg property? So just more code in early boot? But that would easily open the door for 64bit addresses for instance, right?

I don't necessarily object to that, but it can be a lot to ask for during early boot.

I don't think Xen in particular has a lot of problems with that. The DT implementation is quite capable (see Julien's latest updates).

It also looks really weird in the chosen node.

TBH this initrd syntax looks a bit weird to me in general, not following the nice and well defined DT way (no compatible, separate properties for start and length, ...)

For boot time artifacts like this, I'd rather see a duplication of the linux,initrd- pattern. Something like: /chosen { xen,linux-zimage-start = <0x80000000>; xen,linux-zimage-end = <0x803dcff8>; linux,initrd-start = <0x8000000>; linux,initrd-end = <0x8123456>; }

To me that sounds like a hack to avoid more coding work. The problem with that is that it would introduce a Xen specific property (actually two). So from a user's perspective you want:

a) generic properties to support future OSes without much hassle (using the DT's compatible semantic) b) specifying load address and length in the normal way c) adding command lines d) potentially adding more arbitrary properties e) allowing an arbitrary number of modules

To just go ahead with another hard-coded property to make life easier for us kernel level developers doesn't sound very convincing to me. And especially supporting multiple modules would be very hackish then.

If have two reason for this; consistency and simplicity. Consistent because it matches what Linux already uses,

But this is actually not consistent with the rest of the DT architecture. Defining an own node for that looks like a perfect match to me in the DT world.

and really easy to parse without mucking with ranges or #address/size-cells. The assumption is that during early boot the system is dealing with a 1:1 mapped physical address and there isn't any bus translations that need to be processed.

I see your point with early boot code constraints. But my idea was to create a flexible and generic solution, so that there will be a framework for bootloaders to specify blobs without needing updates every time someone comes up with a need for one. On x86 there was once such a need for microcode updates which couldn't be easily satisfied because Linux/x86 itself does not support the multiboot protocol.

Question: what is the need for the xen,linux-initrd node? Why can't the existing "linux,initrd-*" properties continue to be used?

The main driver for this was to avoid patching bootloaders just to boot Xen specifically. Rather we want to patch bootloaders to support some kind of "multiboot" protocol once and let Xen piggy-back on this. I think that is somewhat driven by experiences in the x86 world, were especially kernel people are reluctant to support something just for Xen's need.

Regards, Andre.