Re: using valgrind on a loadable builtin

[Robert E. Griffith]

In case someone reads the archive in search for a similar situation, 
this is what ultimately worked for me to get valgrind analysis on my 
loadable builtin runs.

I rebuilt bash to use glibc's malloc. Passing --without-bash-malloc to 
configure did not seem to work for me -- maybe I misunderstood where to 
use it.    I edited the configure script manually to change...

   $ grep "^opt_bash_malloc=" configure
   opt_bash_malloc=no  # changed yes to no
   $ ./configure; make clean; make CFLAGS="-g -O0"

Confirmed that worked by observing 'objdump -t bash | grep malloc' 
before contained sh_malloc and after did not (amung others)

For some reason I dont understand, running "valgrind bash -c 
"myscript.sh"' valgrind would only print its starting comments and 
nothing else.  Maybe because myscript.sh was crashing?

What worked was this...

   $ valgrind  --log-file=valgrind.log --leak-check=full    
../../bashParse/bash # your options and path to bash may vary
   (new bash prompt)$ echo $$
   <thePID>
   (new bash prompt)$ source myscript.sh
   (new bash prompt)$ exit # only if myscript.sh did not end the bash process
   $ grep <thePID> valgrind.log valgrind.log.myscript

valgrind.log contained a lot of information about each subshell bash 
created along the way so that is why I filtered for the top level bash pid.

Once I got this working, valgrind made it clear where the bug was (I had 
accidentally misplaced the +1 for the terminator in the expression --  
malloc(strlen(s1)+strlen(s2+1))

--BobG


On 5/25/22 21:52, Robert E. Griffith wrote:
> Oh! That makes sense now. Thanks. I was inspecting the memory and was 
> confused that it did not seem to use the same chunk structure of the 
> glibc malloc.
>
> --BobG
>
> On 5/25/22 18:19, Eduardo Bustamante wrote:
> > On Wed, May 25, 2022 at 10:44 AM Robert E. Griffith <bobg@junga.com> 
> > wrote:
> >
> >     I am trying to to use valgrind memcheck on my loadable builtin
> >     which is
> >     getting a "corrupted size vs. prev_size" error but its output is
> >     suspiciously clean and when I use the -v (verbose) switch I do not
> >     see
> >     any msg about my builtin's .so being loaded even though I do see
> >     confirmation that the builtin gets loaded and ran in the output. 
> > So I
> >     suspect that it is not monitoring my .so.
> >
> >     I am running it like...
> >
> >         valgrind -v  ../../bashParse/bash -c  'bg-dev tests run
> >     bg_objects.sh:'
> >
> >     The script enables my builtin.
> >
> >     bash and my builtin are being built with -g and no -O<n>
> >
> >     Anyone have any tips on getting meaningful output from valgrind 
> > for a
> >     loadable builtin?
> >
> >
> > Bash provides its own (lib/malloc/) malloc implementation.  When 
> > building bash, you can provide --without-bash-malloc to the configure 
> > script to instead link against the libc malloc. You can search for 
> > see opt_bash_malloc in "configure", or the USING_BASH_MALLOC macro.
> >
> > You can review these files to see how it's all hooked 
> > up: builtins.h, general.h, xmalloc.h
> >
> > Valgrind is not capable of tracing the malloc that Bash provides.