Booting U-Boot SPL on Qualcomm SoCs

Overview

The boot process on sdm845 (and some other Qualcomm SoCs) starts with the bootrom of the Application Processor, which executes XBL_SEC, which jumps to “OEM” code in EL1. Production devices are typically “fused”, with a hash of the OEM’s signing key burnt into one of the “QFUSE” banks on the SoC making it impossible to run custom bootloader code. As a result U-Boot SPL is only supported on unfused (“secureboot off”) devices. XBL_SEC is always signed by Qualcomm, and the fuses to disable turning off signature verification for it are always burnt at the factory, so replacing XBL_SEC is impossible without using JTAG. Of course JTAG is typically disabled on devices that have secure boot enabled, or at minimum greatly neutered.

U-Boot SPL for Qualcomm platforms uses a custom linker script (per SoC) to build a bootable ELF. For sdm845 (and some other platforms) this has two sections, u-boot code and an embedded xbl_sec elf (signed by Qualcomm). To boot on an unfused SoC, the elf additionally needs to have hash sections added, which can be accomplished with qtestsign.

Currently, sdm845 is supported. You need a device with secure boot disabled (or with secure boot enabled if you enabled it yourself and have the private key, though for full security you’d also want to disable JTAG which will remove your ability to mess with the control flow in the bootrom (immutable) and in XBL_SEC (signed)).

Building

First, obtain an xbl_sec that includes the EL3 privilege escalation feature and place it at .output/xbl_sec.elf. You can extract it from an xbl elf. If you’re unable to find one, you can also use JTAG/SWD to break at the SMC entry and use gdb to jump to the u-boot entry point in EL3.

To build a bootable image, you need to use a defconfig specific to your SoC. This is because the ELF has to specify where in the address space to put u-boot SPL, and this may differ per SoC. There may be other SoC-dependent build time choices, though in principle those could be made at runtime.

First run make qcom_sdm845_spl_defconfig:

make CROSS_COMPILE=aarch64-suse-linux- O=.output DEVICE_TREE=qcom/sdm845-shift-axolotl qcom_sdm845_spl_defconfig

Then compile u-boot and specify the dts for your board (technically nothing about the resulting SPL image should be board-specific, but there are no non-board-specific device trees in Linux):

make CROSS_COMPILE=aarch64-suse-linux- O=.output DEVICE_TREE=qcom/sdm845-shift-axolotl

Finally, use qtestsign to add the hash segments required by PBL:

qtestsign -v 5 -o .output/spl/u-boot-spl_signed.elf prog .output/spl/u-boot-spl.elf

Running

Currently, U-Boot SPL for Qualcomm platforms expects to be booted via EDL:

edl.py --loader=$PWD/.output/spl/u-boot-spl_signed.elf

SPL will then launch the DFU gadget and wait for you to upload u-boot proper:

dfu-util -RD .output/u-boot.img

u-boot proper will then likely crash, since SPL currently doesn’t init DRAM on Qualcomm platforms and u-boot proper currently doesn’t support running from SRAM. The latter should be an easy fix.

Notes on memory map

There are various banks of SRAM on a Qualcomm SoC that we can use prior to DRAM init.
For example:
msm8916 - 512K L2-as-TCM (at 0x08000000), 16K OCIMEM (at 0x08600000)
msm8998 - 1M L2-as-TCM (at 0x14000000), 256K OCIMEM (at 0x14680000)
sdm845 - 1.5M BOOT_IMEM (at 0x14800000), 256K OCIMEM (at 0x14680000)

There’s also RPM code/data RAM and hexagon TCMs, but unless we want to boot dram-less Linux we can probably safely ignore those. On msm8916 they may come in handy though.

sdm845 can also have 8M LLCC-as-TCM in theory, but this appears to be broken. L2-as-TCM is no longer present.

Since a limited amount of not necessarily continuous SRAM is available, we need to manually specify where .text, .bss, the malloc pool and the stack go. The Kconfig contains reasonable defaults per SoC.

On sdm845, we by default put U-Boot SPL in BOOT_IMEM, with .bss, malloc pool and the stack filling OCIMEM. We can also fit U-Boot proper in BOOT_IMEM, for dram-less DFU or peek/poke with a shell. To that end, we set CONFIG_TEXT_BASE at 512K into BOOT_IMEM, and set CONFIG_SPL_MAX_SIZE to 512K - 64. We also configure dfu to load U-Boot proper to CONFIG_TEXT_BASE - 64. (64 bytes is the size of u-boot legacy header)