[GHC] #8935: Obscure linker bug leads to crash in GHCi

[GHC]

#8935: Obscure linker bug leads to crash in GHCi
------------------------------------+-------------------------------------
       Reporter:  simonmar          |             Owner:  simonmar
           Type:  bug               |            Status:  new
       Priority:  high              |         Milestone:  7.8.2
      Component:  Runtime System    |           Version:  7.6.3
       Keywords:                    |  Operating System:  Unknown/Multiple
   Architecture:  Unknown/Multiple  |   Type of failure:  GHCi crash
     Difficulty:  Rocket Science    |         Test Case:
     Blocked By:                    |          Blocking:
Related Tickets:                    |
------------------------------------+-------------------------------------
 I have a build of GHC (with `DYNAMIC_GHC_PROGRAMS=NO`) that exhibits the
 following crash:

 {{{
   $ ghc -e 'System.Environment.getEnvironment'
   <segfault>
 }}}

 I tracked it down, eventually, to a bad reference to the symbol `environ`
 from `__hscore_environ` in `libraries/base/includes/HsBase.h`.  Somehow,
 `environ` had got linked to the wrong address.

 Lots more investigation lead me to discover this: `internal_dlsym()` in
 `Linker.c` tries to look up a symbol in all the different shared libraries
 we have loaded so far, one by one.  (see
 be497c202b790999c3fd0ddc4a4176b8cf6acf7e).  Unfortunately, this seems to
 break things in my case.  Here's a simple test program that works on
 Ubuntu 12.04:

 {{{
 #include <dlfcn.h>
 #include <stdio.h>

 char *so = "/usr/lib/x86_64-linux-gnu/libgmp.so";
 char *so2 = "/usr/lib/x86_64-linux-gnu/libpthread.so";

 extern char**environ;

 int main(int argc, char *argv[])
 {
   void *deflt, *hdl;
   deflt = dlopen(NULL, RTLD_LAZY | RTLD_GLOBAL);
   printf("environ = %p\n", &environ);
   printf("dlsym(deflt, \"environ\") = %p\n", dlsym(deflt,"environ"));
   hdl = dlopen(so, RTLD_LAZY);
   printf("dlsym(\"libgmp\", \"environ\") = %p\n", dlsym(hdl,"environ"));
   hdl = dlopen(so2, RTLD_LAZY);
   printf("dlsym(\"libpthread\", \"environ\") = %p\n",
 dlsym(hdl,"environ"));
 }
 }}}

 And the output:

 {{{
 environ = 0x601040
 dlsym(deflt, "environ") = 0x601040
 dlsym("libgmp", "environ") = 0x2aaaab290568
 dlsym("libpthread", "environ") = 0x601040
 }}}

 Note that the value we get from looking up `environ` in `libgmp` is
 different to the others.  The correct value is `0x601040`.  gdb thinks
 that `0x2aaaab290568` is also `environ`:

 {{{
 (gdb) p4 0x2aaaab290568
 0x2aaaab290580 <buflen.9817>:   0x0
 0x2aaaab290578: 0x0
 0x2aaaab290570 <miss_F_GETOWN_EX>:      0x0
 0x2aaaab290568 <environ>:       0x0
 }}}

 but note that it contains zero.  The real one is:

 {{{
 (gdb) p4 0x601040
 0x601058:       0x0
 0x601050 <dtor_idx.6533>:       0x0
 0x601048 <completed.6531>:      0x0
 0x601040 <environ@@GLIBC_2.2.5>:        0x7fffffffe268
 }}}

 In GHC we're loading `libgmp` when we load the `integer-gmp` package, and
 this causes future references to `environ` to go wrong.

 I've locally fixed this by changing `internal_dlsym` to `dlsym`, but since
 there was a reason to make this change in the first place I haven't pushed
 it to master.  Suggestions welcome.

--
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/8935>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler