arm-trusted-firmware/docs/plat/rpi3.rst

15 KiB

<html xmlns="http://www.w3.org/1999/xhtml" xml:lang="en" lang="en"> <head> </head>

Trusted Firmware-A for Raspberry Pi 3

The Raspberry Pi 3 is an inexpensive single-board computer that contains four Arm Cortex-A53 cores, which makes it possible to have a port of Trusted Firmware-A (TF-A).

The following instructions explain how to use this port of the TF-A with the default distribution of Raspbian because that's the distribution officially supported by the Raspberry Pi Foundation. At the moment of writing this, the officially supported kernel is a AArch32 kernel. This doesn't mean that this port of TF-A can't boot a AArch64 kernel. The Linux tree fork maintained by the Foundation can be compiled for AArch64 by following the steps in AArch64 kernel build instructions.

IMPORTANT NOTE: This port isn't secure. All of the memory used is DRAM, which is available from both the Non-secure and Secure worlds. This port shouldn't be considered more than a prototype to play with and implement elements like PSCI to support the Linux kernel.

1.   Design

The SoC used by the Raspberry Pi 3 is the Broadcom BCM2837. It is a SoC with a VideoCore IV that acts as primary processor (and loads everything from the SD card) and is located between all Arm cores and the DRAM. Check the Raspberry Pi 3 documentation for more information.

This explains why it is possible to change the execution state (AArch64/AArch32) depending on a few files on the SD card. We only care about the cases in which the cores boot in AArch64 mode.

The rules are simple:

  • If a file called kernel8.img is located on the boot partition of the SD card, it will load it and execute in EL2 in AArch64. Basically, it executes a default AArch64 stub at address 0x0 that jumps to the kernel.
  • If there is also a file called armstub8.bin, it will load it at address 0x0 (instead of the default stub) and execute it in EL3 in AArch64. All the cores are powered on at the same time and start at address 0x0.

This means that we can use the default AArch32 kernel provided in the official Raspbian distribution by renaming it to kernel8.img, while TF-A and anything else we need is in armstub8.bin. This way we can forget about the default bootstrap code. When using a AArch64 kernel, it is only needed to make sure that the name on the SD card is kernel8.img.

Ideally, we want to load the kernel and have all cores available, which means that we need to make the secondary cores work in the way the kernel expects, as explained in Secondary cores. In practice, a small bootstrap is needed between TF-A and the kernel.

To get the most out of a AArch32 kernel, we want to boot it in Hypervisor mode in AArch32. This means that BL33 can't be in EL2 in AArch64 mode. The architecture specifies that AArch32 Hypervisor mode isn't present when AArch64 is used for EL2. When using a AArch64 kernel, it should simply start in EL2.

1.1.   Placement of images

The file armstub8.bin contains BL1 and the FIP. It is needed to add padding between them so that the addresses they are loaded to match the ones specified when compiling TF-A.

The device tree block is loaded by the VideoCore loader from an appropriate file, but we can specify the address it is loaded to in config.txt.

The file kernel8.img contains a kernel image that is loaded to the address specified in config.txt. The Linux kernel tree has information about how a AArch32 Linux kernel image is loaded in Documentation/arm/Booting:

The zImage may also be placed in system RAM and called there.  The
kernel should be placed in the first 128MiB of RAM.  It is recommended
that it is loaded above 32MiB in order to avoid the need to relocate
prior to decompression, which will make the boot process slightly
faster.

There are no similar restrictions for AArch64 kernels, as specified in the file Documentation/arm64/booting.txt.

This means that we need to avoid the first 128 MiB of RAM when placing the TF-A images (and specially the first 32 MiB, as they are directly used to place the uncompressed AArch32 kernel image. This way, both AArch32 and AArch64 kernels can be placed at the same address.

In the end, the images look like the following diagram when placed in memory. All addresses are Physical Addresses from the point of view of the Arm cores. Again, note that this is all just part of the same DRAM that goes from 0x00000000 to 0x3F000000, it just has different names to simulate a real secure platform!

0x00000000 +-----------------+
           |       ROM       | BL1
0x00010000 +-----------------+
           |       FIP       |
0x00200000 +-----------------+
           |                 |
           |       ...       |
           |                 |
0x01000000 +-----------------+
           |     Kernel      |
           +-----------------+
           |                 |
           |       ...       |
           |                 |
0x02000000 +-----------------+
           |       DTB       |
           +-----------------+
           |                 |
           |       ...       |
           |                 |
0x10000000 +-----------------+
           |   Secure SRAM   | BL2, BL31
0x10100000 +-----------------+
           |   Secure DRAM   |
0x10300000 +-----------------+
           | Non-secure DRAM | BL33
0x11000000 +-----------------+
           |                 |
           |       ...       |
           |                 |
0x3F000000 +-----------------+
           |       I/O       |
0x40000000 +-----------------+

The area between 0x10000000 and 0x11000000 has to be protected so that the kernel doesn't use it. That is done by adding memmap=256M$16M to the command line passed to the kernel. See the Setup SD card instructions to see how to do it.

The last 16 MiB of DRAM can only be accessed by the VideoCore, that has different mappings than the Arm cores in which the I/O addresses don't overlap the DRAM. The memory reserved to be used by the VideoCore is always placed at the end of the DRAM, so this space isn't wasted.

Considering the 128 MiB allocated to the GPU and the 16 MiB allocated for TF-A, there are 880 MiB available for Linux.

1.2.   Boot sequence

The boot sequence of TF-A is the usual one except when booting an AArch32 kernel. In that case, BL33 is booted in AArch32 Hypervisor mode so that it can jump to the kernel in the same mode and let it take over that privilege level. If BL33 was running in EL2 in AArch64 (as in the default bootflow of TF-A) it could only jump to the kernel in AArch32 in Supervisor mode.

The Linux kernel tree has instructions on how to jump to the Linux kernel in Documentation/arm/Booting and Documentation/arm64/booting.txt. The bootstrap should take care of this.

1.3.   Secondary cores

The kernel used by Raspbian doesn't have support for PSCI, so it is needed to use mailboxes to trap the secondary cores until they are ready to jump to the kernel. This mailbox is located at a different address in the AArch32 default kernel than in the AArch64 kernel.

Also, this port of TF-A has another Trusted Mailbox in Shared BL RAM. During cold boot, all secondary cores wait in a loop until they are given given an address to jump to in this Mailbox (bl31_warm_entrypoint).

Once BL31 has finished and the primary core has jumped to the BL33 payload, it has to call PSCI_CPU_ON to release the secondary CPUs from the wait loop. The payload then makes them wait in another waitloop listening from messages from the kernel. When the primary CPU jumps into the kernel, it will send an address to the mailbox so that the secondary CPUs jump to it and are recognised by the kernel.

2.   Build Instructions

To boot a AArch64 kernel, only the AArch64 toolchain is required.

To boot a AArch32 kernel, both AArch64 and AArch32 toolchains are required. The AArch32 toolchain is needed for the AArch32 bootstrap needed to load a 32-bit kernel.

First, clone and compile Raspberry Pi 3 TF-A bootstrap. Choose the one needed for the architecture of your kernel.

Then compile TF-A. For a AArch32 kernel, use the following command line:

CROSS_COMPILE=aarch64-linux-gnu- make PLAT=rpi3             \
RPI3_BL33_IN_AARCH32=1                                      \
BL33=../rpi3-arm-tf-bootstrap/aarch32/el2-bootstrap.bin     \
all fip

For a AArch64 kernel, use this other command line:

CROSS_COMPILE=aarch64-linux-gnu- make PLAT=rpi3             \
BL33=../rpi3-arm-tf-bootstrap/aarch64/el2-bootstrap.bin     \
all fip

Then, join BL1 and the FIP with the following instructions (replace release by debug if you set the build option DEBUG=1):

cp build/rpi3/release/bl1.bin bl1.pad.bin
truncate --size=65536 bl1.pad.bin
cat bl1.pad.bin build/rpi3/release/fip.bin > armstub8.bin

The resulting file, armstub8.bin, contains BL1 and the FIP in the place they need to be for TF-A to boot correctly. Now, follow the instructions in Setup SD card.

The following build options are supported:

  • PRELOADED_BL33_BASE: Specially useful because the file kernel8.img can be loaded anywhere by modifying the file config.txt. It doesn't have to contain a kernel, it could have any arbitrary payload.
  • RESET_TO_BL31: Set to 1 by default. If using a 32-bit kernel like Raspbian, the space used by BL1 can overwritten by the kernel when it is being loaded. Even when using a AArch64 kernel the region used by BL1 isn't protected and the kernel could overwrite it. The space used by BL31 is reserved by the command line passed to the kernel.
  • RPI3_BL33_IN_AARCH32: This port can load a AArch64 or AArch32 BL33 image. By default this option is 0, which means that TF-A will jump to BL33 in EL2 in AArch64 mode. If set to 1, it will jump to BL33 in Hypervisor in AArch32 mode.

The following is not currently supported:

  • AArch32 for TF-A itself.
  • EL3_PAYLOAD_BASE: The reason is that you can already load anything to any address by changing the file armstub8.bin, so there's no point in using TF-A in this case.
  • LOAD_IMAGE_V2=0: Only version 2 is supported.

3.   AArch64 kernel build instructions

The following instructions show how to install and run a AArch64 kernel by using a SD card with the default Raspbian install as base. Skip them if you want to use the default 32-bit kernel.

Note that this system won't be fully 64-bit because all the tools in the filesystem are 32-bit binaries, but it's a quick way to get it working, and it allows the user to run 64-bit binaries in addition to 32-bit binaries.

  1. Clone the Linux tree fork maintained by the Raspberry Pi Foundation. To speed things up, do a shallow clone of the desired branch.
git clone --depth=1 -b rpi-4.14.y https://github.com/raspberrypi/linux
cd linux
  1. Configure and compile the kernel. Adapt the number after -j so that it is 1.5 times the number of CPUs in your computer. This may take some time to finish.
make ARCH=arm64 CROSS_COMPILE=aarch64-linux-gnu- bcmrpi3_defconfig
make -j 6 ARCH=arm64 CROSS_COMPILE=aarch64-linux-gnu-
  1. Copy the kernel image and the device tree to the SD card. Replace the path by the corresponding path in your computers to the boot partition of the SD card.
cp arch/arm64/boot/Image /path/to/boot/kernel8.img
cp arch/arm64/boot/dts/broadcom/bcm2710-rpi-3-b.dtb /path/to/boot/
  1. Install the kernel modules. Replace the path by the corresponding path to the filesystem partition of the SD card on your computer.
make ARCH=arm64 CROSS_COMPILE=aarch64-linux-gnu- \
INSTALL_MOD_PATH=/path/to/filesystem modules_install
  1. Follow the instructions in Setup SD card except for the step of renaming the existing kernel7.img (we have already copied a AArch64 kernel).

4.   Setup SD card

The instructions assume that you have an SD card with a fresh install of Raspbian (or that, at least, the boot partition is untouched, or nearly untouched). They have been tested with the image available in 2017-09-07.

  1. Insert the SD card and open the boot partition.

  2. Rename kernel7.img to kernel8.img. This tricks the VideoCore bootloader into booting the Arm cores in AArch64 mode, like TF-A needs, even though the kernel is not compiled for AArch64.

  3. Copy armstub8.bin here. When kernel8.img is available, The VideoCore bootloader will look for a file called armstub8.bin and load it at address 0x0 instead of a predefined one.

  4. Open cmdline.txt and add memmap=256M$16M to prevent the kernel from using the memory needed by TF-A. If you want to enable the serial port "Mini UART", make sure that this file also contains console=serial0,115200 console=tty1.

    Note that the 16 MiB reserved this way won't be available for Linux, the same way as the memory reserved in DRAM for the GPU isn't available.

  5. Open config.txt and add the following lines at the end (enable_uart=1 is only needed to enable debugging through the Mini UART):

enable_uart=1
kernel_address=0x01000000
device_tree_address=0x02000000

If you connect a serial cable to the Mini UART and your computer, and connect to it (for example, with screen /dev/ttyUSB0 115200) you should see some text. In the case of an AArch32 kernel, you should see something like this:

NOTICE:  Booting Trusted Firmware
NOTICE:  BL1: v1.4(release):v1.4-329-g61e94684-dirty
NOTICE:  BL1: Built : 00:09:25, Nov  6 2017
NOTICE:  BL1: Booting BL2
NOTICE:  BL2: v1.4(release):v1.4-329-g61e94684-dirty
NOTICE:  BL2: Built : 00:09:25, Nov  6 2017
NOTICE:  BL1: Booting BL31
NOTICE:  BL31: v1.4(release):v1.4-329-g61e94684-dirty
NOTICE:  BL31: Built : 00:09:25, Nov  6 2017
[    0.266484] bcm2835-aux-uart 3f215040.serial: could not get clk: -517

Raspbian GNU/Linux 9 raspberrypi ttyS0
raspberrypi login:

Just enter your credentials, everything should work as expected. Note that the HDMI output won't show any text during boot.

</html>