arch/arm/cpu/armv8/fsl-layerscape/doc/README.falcon

Falcon boot option
------------------
Falcon boot is a short cut boot method for SD/eMMC targets. It skips loading the
RAM version U-Boot. Instead, it loads FIT image and boot directly to Linux.
CONFIG_SPL_OS_BOOT enables falcon boot. CONFIG_SPL_LOAD_FIT enables the FIT
image support (also need CONFIG_SPL_OF_LIBFDT, CONFIG_SPL_FIT and optionally
CONFIG_SPL_GZIP).

To enable falcon boot, a hook function spl_start_uboot() returns 0 to indicate
booting U-Boot is not the first choice. The kernel FIT image needs to be put
at CONFIG_SYS_MMCSD_RAW_MODE_KERNEL_SECTOR. SPL mmc driver reads the header to
determine if this is a FIT image. If true, FIT image components are parsed and
copied or decompressed (if applicable) to their destinations. If FIT image is
not found, normal U-Boot flow will follow.

An important part of falcon boot is to prepare the device tree. A normal U-Boot
does FDT fixups when booting Linux. For falcon boot, Linux boots directly from
SPL, skipping the normal U-Boot. The device tree has to be prepared in advance.
A command "spl export" should be called under the normal RAM version U-Boot.
It is equivalent to go through "bootm" step-by-step until device tree fixup is
done. The device tree in memory is the one needed for falcon boot. Falcon boot
flow suggests to save this image to SD/eMMC at the location pointed by macro
CONFIG_SYS_MMCSD_RAW_MODE_ARGS_SECTOR, with maximum size specified by macro
CONFIG_SYS_MMCSD_RAW_MODE_ARGS_SECTORS. However, when FIT image is used for
Linux, the device tree stored in FIT image overwrites the memory loaded by spl
driver from these sectors. We could change this loading order to favor the
stored sectors. But when secure boot is enabled, these sectors are used for
signature header and needs to be loaded before the FIT image. So it is important
to understand the device tree in FIT image should be the one actually used, or
leave it absent to favor the stored sectors. It is easier to deploy the FIT
image with embedded static device tree to multiple boards.

Macro CONFIG_SYS_SPL_ARGS_ADDR serves two purposes. One is the pointer to load
the stored sectors to. Normally this is the static device tree. The second
purpose is the memory location of signature header for secure boot. After the
FIT image is loaded into memory, it is validated against the signature header
before individual components are extracted (and optionally decompressed) into
their final memory locations, respectively. After the validation, the header
is no longer used. The static device tree is copied into this location. So
this macro is passed as the location of device tree when booting Linux.

Steps to prepare static device tree
-----------------------------------
To prepare the static device tree for Layerscape boards, it is important to
understand the fixups in U-Boot. Memory size and location, as well as reserved
memory blocks are added/updated. Ethernet MAC addressed are updated. FMan
microcode (if used) is embedded in the device tree. Kernel command line and
initrd information are embedded. Others including CPU status, boot method,
Ethernet port status, etc. are also updated.

Following normal booting process, all variables are set, all images are loaded
before "bootm" command would be issued to boot, run command

spl export fdt <address>

where the address is the location of FIT image. U-Boot goes through the booting
process as if "bootm start", "bootm loados", "bootm ramdisk"... commands but
stops before "bootm go". There we have the fixed-up device tree in memory.
We can check the device tree header by these commands

fdt addr <fdt address>
fdt header

Where the fdt address is the device tree in memory. It is printed by U-Boot.
It is useful to know the exact size. One way to extract this static device
tree is to save it to eMMC/SD using command in U-Boot, and extract under Linux
with these commands, repectively

mmc write <address> <sector> <sectors>
dd if=/dev/mmcblk0 of=<filename> bs=512 skip=<sector> count=<sectors>

Note, U-Boot takes values as hexadecimals while Linux takes them as decimals by
default. If using NAND or other storage, the commands are slightly different.
When we have the static device tree image, we can re-make the FIT image with
it. It is important to specify the load addresses in FIT image for every
components. Otherwise U-Boot cannot load them correctly.

Generate FIT image with static device tree
------------------------------------------
Example:

/dts-v1/;

/ {
	description = "Image file for the LS1043A Linux Kernel";
	#address-cells = <1>;

	images {
		kernel {
			description = "ARM64 Linux kernel";
			data = /incbin/("./arch/arm64/boot/Image.gz");
			type = "kernel";
			arch = "arm64";
			os = "linux";
			compression = "gzip";
			load = <0x80080000>;
			entry = <0x80080000>;
		};
		fdt-1 {
			description = "Flattened Device Tree blob";
			data = /incbin/("./fsl-ls1043ardb-static.dtb");
			type = "flat_dt";
			arch = "arm64";
			compression = "none";
			load = <0x90000000>;
		};
		ramdisk {
			description = "LS1043 Ramdisk";
                        data = /incbin/("./rootfs.cpio.gz");
			type = "ramdisk";
			arch = "arm64";
			os = "linux";
			compression = "gzip";
			load = <0xa0000000>;
		};
	};

	configurations {
		default = "config-1";
		config-1 {
			description = "Boot Linux kernel";
			kernel = "kernel";
			fdt = "fdt-1";
			ramdisk = "ramdisk";
			loadables = "fdt", "ramdisk";
		};
	};
};

The "loadables" is not optional. It tells SPL which images to load into memory.

Other things to consider
-----------------------
Falcon boot skips a lot of initialization in U-Boot. If Linux expects the
hardware to be initialized by U-Boot, the related code should be ported to SPL
build. For example, if Linux expect Ethernet PHY to be initialized in U-Boot
(which is not a common case), the PHY initialization has to be included in
falcon boot. This increases the SPL image size and should be handled carefully.
If Linux has PHY driver enabled, it still depends on the correct MDIO bus setup
in U-Boot. Normal U-Boot sets the MDC ratio to generate a proper clock signal.