Jason Xu 9c5c2813a0 Update QEMU command in all README and Makefile | 2 years ago | |
---|---|---|
.. | ||
Makefile | 5 years ago | |
Makefile.clang | 2 years ago | |
Makefile.gcc | 2 years ago | |
OLVASSEL.md | 4 years ago | |
README.md | 4 years ago | |
gpio.h | 6 years ago | |
kernel8.img | 4 years ago | |
link.ld | 6 years ago | |
main.c | 6 years ago | |
mbox.c | 6 years ago | |
mbox.h | 6 years ago | |
start.S | 3 years ago | |
uart.c | 3 years ago | |
uart.h | 6 years ago |
Because changing SD card is boring and also to avoid potential SD card damage, we create a kernel8.img that will load the real kernel8.img over serial.
This tutorial is a rewrite of the well known serial boot loader, raspbootin in 64 bit.
I only provide one part of the loader, the kernel receiver, which runs on the RPi. For the other
part, the sender, which runs on your PC see the original raspbootcom utility.
If you want to send kernels from Windows or MacOSX machines, or you just prefer GUI application, then I suggest to use
USBImager application with the -S
(uppercase) flag. Even more, @milanvidakovic
was kind to share a Java version of the kernel sender with you
(don't let the name fool you, this is actually the server part, never mind the name).
In order to load the new kernel to the same address, we have to move ourself out of the way. It's called chain loading: one code loads the next code to the same position in memory, therefore the latter thinks it was loaded by the firmware. To implement that we use a different linking address this time, and since GPU loads us to 0x80000 regardless, we have to copy our code to that link address. When we're done, the memory at 0x80000 must be free to use.
We also should minimize the size of the loader, since it will be regarded by the newly loaded code anyway.
By removing uart_puts()
and other functions, I've managed to shrink the loader's size below 1024 bytes. This
way I can assure that link address 0x80000 - 1024 will be suitable, and our loader code won't overlap with
the load address. You can check that with:
$ aarch64-elf-readelf -s kernel8.elf | grep __bss_end
27: 000000000007ffb0 0 NOTYPE GLOBAL DEFAULT 4 __bss_end
We have to save the arguments in registers passed by the firmware. Added a loop to relocate our code to the address it should have been loaded to. And last, since gcc generates RIP-relative jumps, we must adjust the branch instruction to jump to the relocated C code.
Thanks to @mrvn for noticing running a non-relocated spin-loop would be very bad. Because of a change in the firmware, this code runs only on the BSP, that's why it didn't caused trouble and that's why nobody noticed.
We use a different linking address this time. Similarly to bss size calculation, we calculate our code's size to know how many bytes we have to copy.
We print 'RBIN64', receive the new kernel over serial and save it at the memory address where the start.elf would have loaded it. When finished, we restore the arguments and jump to the new kernel using an absolute address.