[Bug nptl/17165] New: [alpha] Saved stack pointer is not mangled in unwind_stop jumpbuffer

["ubizjak at gmail dot com" <sourceware-bugzilla at sourceware dot org>]

https://sourceware.org/bugzilla/show_bug.cgi?id=17165

            Bug ID: 17165
           Summary: [alpha] Saved stack pointer is not mangled in
                    unwind_stop jumpbuffer
           Product: glibc
           Version: 2.17
            Status: NEW
          Severity: normal
          Priority: P2
         Component: nptl
          Assignee: unassigned at sourceware dot org
          Reporter: ubizjak at gmail dot com
                CC: drepper.fsp at gmail dot com

Following testcase:

--cut here--
#include <bits/cxxabi_forced.h>
#include <pthread.h>

void* f(void*) {
 try
   {
     pthread_exit(0);
   }
 catch (__cxxabiv1::__forced_unwind const&)
   {
     __builtin_puts("unwind");
     throw;
   }
 catch (...)
   {
     __builtin_puts("something else");
     throw;
   }
}

int main()
{
  pthread_t t;
  pthread_create(&t, 0, f, 0);
  pthread_join(t, 0);

}
--cut here--

executes on alpha-linux-gnu without printing anything when compiled with
"gcc-4.9 -pthread". The problem was tracked to invalid execution of unwind_stop
from nptl/unwind.c.

The referred function calls _jmpbuf_sp:

0033 static inline uintptr_t __attribute__ ((unused))
0034 _jmpbuf_sp (__jmp_buf regs)
0035 {
0036   uintptr_t sp = regs[JB_SP];
0037 #ifdef PTR_DEMANGLE
0038   PTR_DEMANGLE (sp);
0039 #endif
0040   return sp;

through _JMPBUF_CFA_UNWINDS_ADJ/ _JMPBUF_UNWINDS_ADJ macros, defined in
âsysdeps/âalpha/âjmpbuf-unwind.h. However, regs[JB_SP] is not mangled at all at
function entry, as confirmed by following gdb session:

(gdb) b jmpbuf-unwind.h:38
No source file named jmpbuf-unwind.h.
Make breakpoint pending on future shared library load? (y or [n]) y
Breakpoint 1 (jmpbuf-unwind.h:38) pending.
(gdb) r
Starting program: /space/homedirs/uros/test/a.out 
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib/libthread_db.so.1".
[New Thread 0x20000c391f0 (LWP 11609)]
[Switching to Thread 0x20000c391f0 (LWP 11609)]

Breakpoint 1, unwind_stop (version=<optimized out>, actions=<optimized out>,
exc_class=<optimized out>, exc_obj=<optimized out>, context=<optimized out>, 
    stop_parameter=<optimized out>) at unwind.c:61
61            || ! _JMPBUF_CFA_UNWINDS_ADJ (buf->cancel_jmp_buf[0].jmp_buf,
context,
(gdb) s
_jmpbuf_sp (regs=0x20000c38a40) at ../ports/sysdeps/alpha/jmpbuf-unwind.h:38
38        PTR_DEMANGLE (sp);
(gdb) p/x regs[8]
$1 = 0x20000c38a10

Following detail can be obtained from dissasembly:

   0x000002000028774c <+188>:   ldq     s1,64(a5)  <-- load SP value
   0x0000020000287750 <+192>:   ldq     t0,0(t0)   <-- load ptr chk guard
=> 0x0000020000287754 <+196>:   xor     s1,t0,s1   <-- mangling function

just before mangling function, we have:

(gdb) i r s1 t0
s1             0x20000c38a10    2199036070416
t0             0x33608b9cba2ff160       3702112398953410912

The consequence of the wrong value is, that do_longjump flag gets set in
unwind_stop function. This caused errorneous call to __libc_unwind_longjmp at
the end of the function, which silently terminates the execution.

I suspect a missing call to PTR_MANGLE, or one of PTR_MANGLE* macros in
sysdeps/unix/alpha/sysdep.h is not defined correctly, so saved SP value in
jmp_buf isn't mangled as expected by unwind_stop function.

-- 
You are receiving this mail because:
You are on the CC list for the bug.