On Mon, 2014-11-10 at 17:08 +0100, Mark Wielaard wrote:
> We had a regression on i686 where we failed to find the vdso symbol table
> and so couldn't resolve any addresses pointing into the vdso to function
> names. This was because at least on i686 we have to rely on the phdrs and
> the DYNAMIC segment to locate the symbol table.
> 
> The fix look large, but ignoring whitespace it is actually pretty small:
> 
> diff --git a/libdwfl/dwfl_module_getdwarf.c b/libdwfl/dwfl_module_getdwarf.c
> index e705f57..adb94b4 100644
> --- a/libdwfl/dwfl_module_getdwarf.c
> +++ b/libdwfl/dwfl_module_getdwarf.c
> @@ -718,10 +718,12 @@ find_dynsym (Dwfl_Module *mod)
>               break;
>             }
>  
> -         /* Translate pointers into file offsets.  */
> +         /* Translate pointers into file offsets.  ADJUST is either zero
> +            in case the dynamic segment wasn't adjusted or mod->main_bias.  */
> +         void translate_offs (GElf_Addr adjust)
> +         {
>             GElf_Off offs[i_max] = { 0, };
> -         find_offsets (mod->main.elf, mod->main_bias, phnum, i_max, addrs,
> -                       offs);
> +           find_offsets (mod->main.elf, adjust, phnum, i_max, addrs, offs);
>  
>             /* Figure out the size of the symbol table.  */
>             if (offs[i_hash] != 0)
> @@ -824,9 +826,18 @@ find_dynsym (Dwfl_Module *mod)
>                     mod->symfile = &mod->main;
>                     mod->symerr = DWFL_E_NOERROR;
>                   }
> -             return;
>               }
>           }
> +
> +         /* First try unadjusted, like ELF files from disk, vdso.
> +            Then try for already adjusted dynamic section, like ELF
> +            from remote memory.  */
> +         translate_offs (0);
> +         if (mod->symfile == NULL)
> +           translate_offs (mod->main_bias);
> +
> +         return;
> +       }
>      }
>  }
> 
> 
> <--->
> 
> commit 037505 "Fix resolving ELF symbols for live PIDs with deleted files"
> changed find_dynsym to assume the PT_DYNAMIC entries had been adjusted by
> the dynamic linker. That is often a correct assumption when the ELF image
> comes from remote memory. But we cannot rely on that. In the case of the
> vdso image the DYNAMIC segment has not been adjusted for example.
> 
> There is no good way to determine whether the DYNAMIC segment has or
> hasn't been adjusted already to the load address by the dynamic linker.
> So we just try twice. Once without and if the fails again with assuming
> adjustments being applied.
> 
> Includes a new vdsosyms testcase that fails on i686 before and succeeds
> after the fix.

After some more testing I pushed this to master now.