Apparently calling printf() from more than one thread triggers a race condition on _IO_2_1_stdout_.file._mode. This race is triggered by line 1270 of file vfprintf.c, which contains the following statement: "ORIENT;". This statement tests some bits in _IO_2_1_stdout_.file._mode and takes a decision. Later on vfprintf() calls _IO_flockfile(). This is a race condition because other threads may change these bits after vfprintf() tested these bits and before vfprintf() calls _IO_flockfile().

I discovered this issue through verifying a multithreaded program with DRD (a data race detection tool). The first error message printed by DRD for a program that calls printf() from more than one thread is as follows:

$ svn info .
URL: svn://svn.valgrind.org/valgrind/branches/DRDDEV
...
Revision: 8895
...
$ ./vg-in-place --var-info=yes --tool=drd drd/tests/circular_buffer
...
==9097== Thread 3:
==9097== Conflicting load by thread 3/3 at 0x0539d840 size 4
==9097==    at 0x508D460: vfprintf (vfprintf.c:1270)
==9097==    by 0x5097599: printf (printf.c:35)
==9097==    by 0x400D99: buffer_send (circular_buffer.c:105)
==9097==    by 0x400E08: producer (circular_buffer.c:124)
==9097==    by 0x4C2A047: vg_thread_wrapper (drd_pthread_intercepts.c:184)
==9097==    by 0x4E3506F: start_thread (pthread_create.c:297)
==9097==    by 0x511D0EC: clone (clone.S:112)
==9097== Location 0x539d840 is 0 bytes inside _IO_2_1_stdout_.file._mode,
==9097== a global variable declared at stdfiles.c:70
...