Re: resolving kernel addresses to symbol names via /proc/kallsyms

[Mark Wielaard <mjw@redhat.com>]

On Tue, 2017-01-31 at 14:39 +0100, Milian Wolff wrote:
> should I be able to resolve kernel addresses to symbol names with elfutils 
> libdwfl?

Yes. See also eu-addr2line -k -S which I believe does what you are
trying to do below.

> I did something like this:
> 
>     m_dwfl = dwfl_begin(m_callbacks);
>     auto i_k = dwfl_linux_kernel_report_kernel(m_dwfl);
>     auto i_m = dwfl_linux_kernel_report_modules(m_dwfl);
> 
> Both i_k and i_m returned 0. Then, I picked a random address from my /proc/
> kallsyms, e.g.:
> 
>     ffffffff81001000 T hypercall_page
> 
> and try to find that symbol:
> 
>     auto module = dwfl_addrmodule(m_dwfl, 0xffffffff81001000ll);
> 
> The module is non-null and points me at this module (comparing to output I get 
> from dwfl_getmodules):
> 
>     0x61400000fa40 kernel ffffffff81000000

You could also use dwfl_module_info () to get the name and other
information about the module found for that address.

> But when I try to get a symbol, I get a nullptr:
> 
>     auto sym = dwfl_module_addrname(module, 0xffffffff81001000ll);
> 
> So, what am I missing?

I don't immediately know without seeing the full program.
You could try using dwfl_module_addrinfo () to see if you can get more
information about the address in that module.

You could also try calling dwfl_module_getsymtab () on the module and
look at some of the symbols found with dwfl_module_getsym_info () to see
if libdwfl found the symbol table for that module.

> Also, it doesn't seem to be possible to specify the path to the kallsyms file 
> with the current dwfl_linux_kernel_report_kernel API - how is this intended to 
> be handled? Should one instead use dwfl_linux_kernel_report_offline when one 
> wants to support cross-machine analysis?

Yes, one should use dwfl_linux_kernel_report_offline when cross-machine
(or even just cross kernel version) analysis. The /proc/kallsyms file is
only used to estimate some running kernel address boundaries (but isn't
strictly necessary).