[Bug dynamic-link/29740] New: Race condition between pthread_exit and fork

[Bug dynamic-link/29740] New: Race condition between pthread_exit and fork

[prekgeo at yahoo dot com]

https://sourceware.org/bugzilla/show_bug.cgi?id=29740

            Bug ID: 29740
           Summary: Race condition between pthread_exit and fork
           Product: glibc
           Version: unspecified
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: dynamic-link
          Assignee: unassigned at sourceware dot org
          Reporter: prekgeo at yahoo dot com
  Target Milestone: ---

Created attachment 14427
  --> https://sourceware.org/bugzilla/attachment.cgi?id=14427&action=edit
glibc patch to add delay in do_lookup_x

I have observed a crash in a Python program that uses threads and fork. Below,
I attach the following:

* a patch for the latest glibc master (commit
b4174c28d21e1672ef3cc15a058558e97b8471c6) that introduces a delay to force the
race condition.
* a minimal C reproduction.

My understanding of the problem is that the following sequence of events takes
place:

parent: main-thread: starts thread-1
parent: thread-1: exits
parent: thread-1: tries to call _Unwind_Find_FDE
parent: thread-1: ld.so loads libgcc_s.so in memory
parent: main-thread: fork()
parent: thread-1: resolves _Unwind_Find_FDE@got.plt
child: main-thread: starts thread-1
child: thread-1: exits
child: thread-1: calls _Unwind_Find_FDE using the unresolved
_Unwind_Find_FDE@got.plt

The child process crashes with a SIGSEGV with the following backtrace:

#0  0x0000000000003230 in ?? ()
#1  0x00007ffff75c2868 in uw_frame_state_for (context=0x7ffff7dcdc70,
fs=0x7ffff7dcdab0) at ../../../src/libgcc/unwind-dw2.c:1263
#2  0x00007ffff75c3a20 in uw_init_context_1 (context=0x7ffff7dcdc70,
outer_cfa=0x7ffff7dcdea0,
    outer_ra=0x7ffff7e64e36 <__GI___pthread_unwind+70>) at
../../../src/libgcc/unwind-dw2.c:1592
#3  0x00007ffff75c431a in _Unwind_ForcedUnwind (exc=0x7ffff7dced30,
stop=stop@entry=0x7ffff7e64ca0 <unwind_stop>,
    stop_argument=0x7ffff7dcdef0) at ../../../src/libgcc/unwind.inc:211
#4  0x00007ffff7e64e36 in __GI___pthread_unwind (buf=<optimized out>) at
unwind.c:130
#5  0x00007ffff7e5ddba in __do_cancel () at ../sysdeps/nptl/pthreadP.h:276
#6  __GI___pthread_exit (value=0x0) at pthread_exit.c:36
#7  0x00007ffff7fbf283 in ?? ()
#8  0x0000000000000000 in ?? ()

0x3230 is the value of _Unwind_Find_FDE@got.plt in the binary:

$ gdb -q /usr/lib/x86_64-linux-gnu/libgcc_s.so.1
(gdb) p '_Unwind_Find_FDE@got.plt'
$1 = (<text from jump slot in .got.plt, no debug info>) 0x3230

I have also a Python reproduction that fails without the forced delay and by
just enabling LD_DEBUG=all (which seems to add the needed delays). Let me know
if it's valuable to attach in this bug.

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

[Bug dynamic-link/29740] Race condition between pthread_exit and fork

[prekgeo at yahoo dot com]

https://sourceware.org/bugzilla/show_bug.cgi?id=29740

--- Comment #1 from George Prekas <prekgeo at yahoo dot com> ---
Created attachment 14428
  --> https://sourceware.org/bugzilla/attachment.cgi?id=14428&action=edit
Reproduction

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

[Bug dynamic-link/29740] Race condition between pthread_exit and fork

[prekgeo at yahoo dot com]

https://sourceware.org/bugzilla/show_bug.cgi?id=29740

George Prekas <prekgeo at yahoo dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |prekgeo at yahoo dot com

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

[Bug dynamic-link/29740] Race condition between pthread_exit and fork

[schwab@linux-m68k.org]

https://sourceware.org/bugzilla/show_bug.cgi?id=29740

Andreas Schwab <schwab@linux-m68k.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|---                         |INVALID
             Status|UNCONFIRMED                 |RESOLVED

--- Comment #2 from Andreas Schwab <schwab@linux-m68k.org> ---
The child may only call async-signal-safe functions, which pthread_create
isn't.

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

[Bug dynamic-link/29740] Race condition between pthread_exit and fork

[fw at deneb dot enyo.de]

https://sourceware.org/bugzilla/show_bug.cgi?id=29740

Florian Weimer <fw at deneb dot enyo.de> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |fw at deneb dot enyo.de

--- Comment #3 from Florian Weimer <fw at deneb dot enyo.de> ---
What's your glibc version? I don't expect unwind-link to have this problem. See
__libc_unwind_link_after_fork. The unwind-link facility was added in glibc
2.34.

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

[Bug dynamic-link/29740] Race condition between pthread_exit and fork

[prekgeo at yahoo dot com]

https://sourceware.org/bugzilla/show_bug.cgi?id=29740

--- Comment #4 from George Prekas <prekgeo at yahoo dot com> ---
This was tested and reproduces on the latest master (commit
b4174c28d21e1672ef3cc15a058558e97b8471c6).

After reading Andreas' response, I found this snippet in fork's man page which
renders the reproduction code invalid:

       *  After a fork() in a multithreaded program, the child can
          safely call only async-signal-safe functions (see
          signal-safety(7)) until such time as it calls execve(2).

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

[Bug dynamic-link/29740] Race condition between pthread_exit and fork

[fw at deneb dot enyo.de]

https://sourceware.org/bugzilla/show_bug.cgi?id=29740

--- Comment #5 from Florian Weimer <fw at deneb dot enyo.de> ---
I see. So it's probably the half-completed dlopen that confuses the subprocess.
We do this after fork:

      /* Reset the lock the dynamic loader uses to protect its data.  */
      __rtld_lock_initialize (GL(dl_load_lock));

      /* Reset the lock protecting dynamic TLS related data.  */
      __rtld_lock_initialize (GL(dl_load_tls_lock));

That's unfortunately not particularly safe.

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