__pthread_setcancelstate called unconditionally, crashes at 0

[Sergey Bugaev]

Hello,

I'm hitting a crash with the following backtrace:

#0  0x0000000000000000 in ?? ()
#1  0x00000000004660dd in __error_internal (status=1,
errnum=1073741826, message=0x9adcef1c, args=0x9adcef18,
args@entry=0x156aa48, mode_flags=2598170400, mode_flags@entry=0) at
error.c:243
#2  0x00000000004661d5 in __error (status=status@entry=0,
errnum=<optimized out>, message=<optimized out>) at error.c:274
#3  0x0000000000401497 in error (__format=<optimized out>,
__errnum=<optimized out>, __status=<optimized out>) at error.h:42

Note that PC is 0 in the top frame. Here's the relevant listing and
backtrace of the frame #1:

(gdb) l
238 {
239 #if defined _LIBC
240  /* We do not want this call to be cut short by a thread
241     cancellation.  Therefore disable cancellation for now.  */
242  int state = PTHREAD_CANCEL_ENABLE;
243  __pthread_setcancelstate (PTHREAD_CANCEL_DISABLE, &state);
244 #endif
245
246  flush_stdout ();
247 #ifdef _LIBC

(gdb) disas
Dump of assembler code for function __error_internal:
   0x00000000004660b0 <+0>: push   %r14
   0x00000000004660b2 <+2>: mov    %r8d,%r14d
   0x00000000004660b5 <+5>: push   %r13
   0x00000000004660b7 <+7>: mov    %rcx,%r13
   0x00000000004660ba <+10>: push   %r12
   0x00000000004660bc <+12>: mov    %rdx,%r12
   0x00000000004660bf <+15>: push   %rbp
   0x00000000004660c0 <+16>: mov    %esi,%ebp
   0x00000000004660c2 <+18>: push   %rbx
   0x00000000004660c3 <+19>: mov    %edi,%ebx
   0x00000000004660c5 <+21>: xor    %edi,%edi
   0x00000000004660c7 <+23>: sub    $0x10,%rsp
   0x00000000004660cb <+27>: lea    0xc(%rsp),%rsi
   0x00000000004660d0 <+32>: movl   $0x1,0xc(%rsp)
   0x00000000004660d8 <+40>: call   0x0
=> 0x00000000004660dd <+45>: mov    $0x569b10,%rax
   0x00000000004660e4 <+52>: mov    (%rax),%rdi
   0x00000000004660e7 <+55>: call   0x433900 <_IO_fflush>

"call 0x0", ouch!

Clearly __pthread_setcancelstate has been pragma weak'd, and used here
without a check. This is a statically linked x86_64-gnu (so, Hurd and
HTL) executable. Commit 93d78ec1cba68184931b75bef29afd3aed30f43a
"nptl: Move pthread_setcancelstate into libc" seems to be the culprit:
that commit only moved the NPTL symbol into libc, yet changed the
original __libc_ptf_call (__pthread_setcancelstate) calls to direct
__pthread_setcancelstate calls, in this and many other places.

This likely hasn't been noticed in the past because the only
statically linked executables typically used on Hurd systems are the
few bootstrap servers, and they're (presumably) all multithreaded.

What would be the best way to get this fixed? (other than eventually
moving htl into libc) Are the other pthread symbols also used
unconditionally? Is there, or should there be, a test for this?

Sergey