Testing valgrind in rawhide (separate valgrind-debuginfo package)

[Mark Wielaard]

On Thu, 2013-03-07 at 09:42 +0100, Jan Kratochvil wrote:
> On Wed, 06 Mar 2013 16:13:47 +0100, Mark Wielaard wrote:
> > On Wed, 2013-03-06 at 14:38 +0000, Richard W.M. Jones wrote:
> [...]
> > > ==11843== 56 bytes in 1 blocks are definitely lost in loss record 6 of 6
> > > ==11843==    at 0x4A06409: malloc (in /usr/lib64/valgrind/vgpreload_memcheck-amd64-linux.so)
> > > ==11843==    by 0x38EAC861F9: strdup (strdup.c:42)
> > > ==11843==    by 0x38EC8097C9: setprocattrcon_raw.constprop.3 (procattr.c:241)
> > > ==11843==    by 0x38EC8099B7: setprocattrcon.constprop.2 (procattr.c:274)
> > > ==11843==    by 0x400955: main (in /tmp/test)
> [...]
> > It looks like you will have to use the setprocattrcon[.constprop]
> > function name in your suppression file. I am not exactly sure how the
> > linker ends up pointing directly to that one for setsockcreatecon (),
> > but it apparently is. And so valgrind will only know it by that name.
> 
> Symbol table '.symtab' contains 770 entries:
>    Num:    Value          Size Type    Bind   Vis      Ndx Name
>     70: 0000000000009990    62 FUNC    LOCAL  DEFAULT   11 setprocattrcon.constprop.2
> 
> Contents of the .debug_info section:
>  <1><62c6>: Abbrev Number: 58 (DW_TAG_subprogram)
>     <62c7>   DW_AT_name        : (indirect string, offset: 0x3eb4): setprocattrcon
>     <62d2>   DW_AT_inline      : 1      (inlined)
>  <1><6791>: Abbrev Number: 40 (DW_TAG_subprogram)
>     <6792>   DW_AT_abstract_origin: <0x62c6>
>     <6794>   DW_AT_low_pc      : 0x9990
>     <679c>   DW_AT_high_pc     : 62
> 
> Valgrind apparently used the ELF symbol name.  But in DWARF there is the real
> function name (and it is even marked as 'inlined' - although it is a standalone
> function).
> 
> GDB also knows the real name from DWARF:
> 
> (gdb) disas 'setprocattrcon.constprop.2'
> Dump of assembler code for function setprocattrcon:
>    0x0000000000009990 <+0>:	push   %rbx
> [...]
>    0x00000000000099cd <+61>:	retq   
> End of assembler dump.
> 
> So it is possible to fix it in Valgrind.

Aha, thanks. Yes using DWARF might help getting more user
friendly/recognizable names.

Though note that the name we were really looking for was
setsockcreatecon, since that is what was called from main. I think that
one is doing a tail-call to setprocattrcon.constprop.2 so might not
easily be available in the backtrace. If you compile and run Richard's
reproducer from https://bugzilla.redhat.com/show_bug.cgi?id=918572 and
break at the strdup call, can you get a backtrace from gdb with
setsockcreatecon in it?

Also using DWARF .debug_info will only work if it is available. By
default valgrind doesn't require DWARF, it uses only the symbol table. I
can look in extending valgrind to use the DWARF info when available for
matching suppressions, but that might mean a suppression only works when
the debuginfo is installed (and it might make valgrind even slower).

Cheers,

Mark