Re: [PATCH] ARM: Remove the special printing format of pc and lr in __show_regs()

[Leizhen (ThunderTown)]

On 2022/8/8 17:41, Russell King (Oracle) wrote:
> On Mon, Aug 01, 2022 at 11:20:16AM +0800, Zhen Lei wrote:
>> Currently, instruction pointers are printed in [<%08lx>] format to make
>> them more visible. However, it is not necessary in __show_regs() because
>> they have the prefix 'pc :' or 'lr :', and it is also inconsistent with
>> that of other registers, which causes misalignment.
> 
> The formatting is not "to make them more visible" - it was to mark the
> addresses that we wanted the ksymoops utility to translate to kernel
> symbols before we had kallsyms in the kernel. If one disables kallsyms,
> then we still need a way to translate kernel addresses to symbols.

I searched the git log and found that the ksymoops utility is discarded.

See:
073a9ecb3a73401662430bb955aedeac1de643d1
However, a commit in the pre-git era [1] had added the statement,
"ksymoops is useless on 2.6.  Please use the Oops in its original format".

That statement existed until commit 4eb9241127a0 ("Documentation:
admin-guide: update bug-hunting.rst") finally removed the stale
ksymoops information.

4eb9241127a0b5ac3aaaf1b246728009527ebc86
- delete all references to ksymoops since it is no longer applicable;

> 
> I notice there is a script which helps with this that is part of the
> kernel source - scripts/decode_stacktrace.sh. I haven't tried this on
> arm32 since I always use kallsyms - and I suspect that is rather
> universally true as it avoids needing System.map files etc to decode
> the oops. That said, if you're building a kernel for small systems,
> you probably don't want the overhead of kallsyms.

Yes, I read scripts/decode_stacktrace.sh, it requires the format "[<...>]".
But if that's the only concern, maybe we can do the conversion from
"pc: addr" and "lr: addr" to "[<addr>]" first in scripts/decode_stacktrace.sh

I'm usually "objdump -d vmlinux > asm_file", then search "addr:" in asm_file.

Honestly, I think format "[<...>]" is dump_backtrace()'s requirement, not __show_regs()'s.


> 
> So, there's an argument for keeping it - it's an API in that it
> provides hints to scripting to identify which values need to be
> converted to symbols. There's also the argument for getting rid of it,
> which is that hardly anyone does that anymore.
> 
> The question is, which is the more important argument, and I don't
> think there's a definite answer. So I'm inclined to leave this
> as-is.

OK

> 

-- 
Regards,
  Zhen Lei