Hi coreboot folks,

I'd like to propose adding optional compile-time console log tokenization
to coreboot, libpayload, and cbmem.

Verbose debug logging gets expensive on constrained targets.  Format
strings make up over half of .rodata in some stages, a BIOS_SPEW boot
overflows the 128 KiB cbmem console and overwrites the early boot log, and
at 115200 baud coreboot spends close to 18 seconds just printing to serial.

The proposal adds a CONFIG_CONSOLE_TOKENIZED option.  When it's enabled,
eligible printk() format strings are replaced at compile time with a 32-bit
hash, and the strings move into a non-loaded ELF section that stays in the
.debug and .elf files.  The console prints a short printable token instead,
so vendor blob output and the existing cbmem log format keep working, and
cbmem -d turns the tokens back into text from the build's ELF files.  It's
off by default and costs nothing when disabled.

On google/skyrim the BIOS_SPEW cbmem console drops from about 183 KiB to
117 KiB, so the whole boot fits in the buffer again.  Serial debug boot
time drops by about 7.5 seconds (38.4 s to 30.8 s), and each RW slot gets
about 25 KiB smaller in flash.

Design doc: Compile-Time Console Log Tokenization
<https://docs.google.com/document/d/1tN19v0Cj9Vq7ixURlXUGDC1sudFuQpi1IYLzOMkQ4yU/preview?usp=sharing>

Patches: coreboot, libpayload, and cbmem series (97306-97312)
<https://review.coreboot.org/q/topic:%22ap-fw-log-tokenization%22>

The design doc covers the details and the alternatives considered, and the
patches are up for review on Gerrit.

Thanks,

Jon
_______________________________________________
coreboot mailing list -- [email protected]
To unsubscribe send an email to [email protected]

Reply via email to