Re: Linux 6.1-rc3

["Arnd Bergmann" <arnd@arndb.de>]

On Tue, Nov 1, 2022, at 18:29, Linus Torvalds wrote:

> IOW, I expect it is - once again - some random linker-generated stub
> that ends up causing problems, where the re-link stage ends up being
> unstable because of some subtle alignment issue or other. I think
> zero-sized symbols have often been involved.

Linker-generated symbols are usually the cause, but those tend to
be fixed with the extra pass, and Guenter said that doing an extra
pass does not avoid the issue here.

I have not tried to understand what the 73bbb94466fd patch actually does,
but as the description explains that it uses either 1-byte or
2-byte encodings for some symbols, I suspect this is related to
the new problem here, possibly it gets into an oscillating state
where making the symbol table shorter causes a symbol to use the
longer representation in the next round, which in turn makes the
table longer again.

What I've done in the past to debug this was to change
scripts/link-vmlinux.sh to use a larger number of steps,
mainly to see if it eventually converges, grows indefinitely
or oscillates. After that I would compare the temporary files
from the last two different steps to see which symbols are
actually different. Unfortunately, there is no built-in
debugging mode in kallsyms, so this is a rather manual process.

       Arnd