Thread

1. BUG #18899: FreeBSD, assembly by means of GCC with ASAN ends with error: undef reference to backtrace_symbols_fd

   PG Bug reporting form <noreply@postgresql.org> — 2025-04-18T12:38:08Z

   The following bug has been logged on the website:
   
   Bug reference:      18899
   Logged by:          Alexander Mitrokhin
   Email address:      a.mitrokhin@postgrespro.ru
   PostgreSQL version: 17.4
   Operating system:   FreeBSD 14.2
   Description:        
   
   Hello.
   
   After `configure' we have $BUILDDIR/src/pg_config.h:
   
   /* Define to 1 if you have the `backtrace_symbols' function. */
   #define HAVE_BACKTRACE_SYMBOLS 1
   
   When checking, such configtest.c is used:
   
   extern char backtrace_symbols ();
   int main () {
       return backtrace_symbols ();
   }
   
   With ASAN, configtest.c is built without the need for some additional
   libraries, because `backtrace_symbols' defined in
   /usr/local/lib/gcc13/libasan.so.8.
   
   $ nm -D /usr/local/lib/gcc13/libasan.so.8 | grep backtrace_symbols
   000000000007e6c0 T __interceptor_backtrace_symbols
   000000000007e6c0 T backtrace_symbols
   
   But when linking src/backend/postgres we will get:
   
   src/backend/utils/error/assert.c:53:(.text+0x19a): undefined reference to
   `backtrace_symbols_fd'
   
    46         /* If we have support for it, dump a simple backtrace */
    47 #ifdef HAVE_BACKTRACE_SYMBOLS
    48         {
    49                 void       *buf[100];
    50                 int                     nframes;
    51 
    52                 nframes = backtrace(buf, lengthof(buf));
    53                 backtrace_symbols_fd(buf, nframes, fileno(stderr));
    54         }
    55 #endif
   
   The `configure' code that checks whether it is necessary to use the
   libexecinfo:
   
   for ac_lib in '' execinfo; do
           if test -z "$ac_lib"; then
                   ac_res="none required"
           else
                   ac_res=-l$ac_lib
                   LIBS="-l$ac_lib $ac_func_search_save_LIBS"
           fi
           if ac_fn_c_try_link "$LINENO"; then
                   ac_cv_search_backtrace_symbols=$ac_res
           fi
           if ${ac_cv_search_backtrace_symbols+:} false; then
                   break
           fi
   done
   ...
   LIBS=$ac_func_search_save_LIBS
   
   With ASAN, the check to -lexecinfo does not reach and this library is not
   added to LIBS.
   
   PS:
   This corrects the problem for me:
   
   diff --git a/configure.ac b/configure.ac
   index 22441728459..fcf9c00c590 100644
   --- a/configure.ac
   +++ b/configure.ac
   @@ -1373,7 +1373,7 @@ AC_SEARCH_LIBS(clock_gettime, rt)
    # Cygwin:
    AC_SEARCH_LIBS(shmget, cygipc)
    # *BSD:
   -AC_SEARCH_LIBS(backtrace_symbols, execinfo)
   +AC_SEARCH_LIBS(backtrace_symbols_fd, execinfo)
    
    AC_SEARCH_LIBS(pthread_barrier_wait, pthread)
   
   /foo
   
   

2. Re: BUG #18899: FreeBSD, assembly by means of GCC with ASAN ends with error: undef reference to backtrace_symbols_fd

   Tom Lane <tgl@sss.pgh.pa.us> — 2025-04-18T17:38:50Z

   PG Bug reporting form <noreply@postgresql.org> writes:
   > With ASAN, configtest.c is built without the need for some additional
   > libraries, because `backtrace_symbols' defined in
   > /usr/local/lib/gcc13/libasan.so.8.
   
   > $ nm -D /usr/local/lib/gcc13/libasan.so.8 | grep backtrace_symbols
   > 000000000007e6c0 T __interceptor_backtrace_symbols
   > 000000000007e6c0 T backtrace_symbols
   
   Isn't this a flat-out bug/oversight in libasan?  It's hard to believe
   that they have a reason for overriding backtrace_symbols that would
   not also apply to backtrace_symbols_fd.  Nor that it's a good idea
   to risk applications seeing different behaviors from those functions.
   
   > This corrects the problem for me:
   
   > -AC_SEARCH_LIBS(backtrace_symbols, execinfo)
   > +AC_SEARCH_LIBS(backtrace_symbols_fd, execinfo)
   
   I do not love this fix.  It's certainly arguable that we shouldn't
   assume that backtrace_symbols_fd is available just because
   backtrace_symbols is, but the reverse assumption is not much better.
   If we were to do anything about this, I think what it ought to be
   is to make separate configure probes for the two functions and then
   have assert.c say "#ifdef HAVE_BACKTRACE_SYMBOLS_FD".  But on the
   whole I think this is libasan's problem not ours.  While you're
   waiting for them to fix it you could just manually add
   LDFLAGS=-lexecinfo in ASAN builds.
   
   			regards, tom lane