id,summary,reporter,owner,description,type,status,milestone,component,version,severity,resolution,keywords,cc
5752,boost::call_once() is unreliable on some platforms,Matthew Dempsky <matthew@…>,viboes,"boost::call_once() is an implementation of Mike Burrows's fast_pthread_once() algorithm, as described in http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2008/n2660.htm.  In the correctness proof, there is a requirement that loads and stores of the epoch value are atomic (i.e., do not exhibit word tearing).  In Mike's example implementation, the sig_atomic_t type is used, which is required by the C standard to support atomic loads and stores.

However, in the boost::call_once() implementation, the epoch value is defined as uintmax_t, which has no such guarantee, and in practice is not atomic on some architectures.  E.g., on OpenBSD/i386, uintmax_t is a 64-bit type and assignments to a 64-bit memory address must be split into two (non-atomic) store instructions.

Therefore, thread/pthread/once.hpp should be changed to use a type that is guaranteed to support atomic loads and stores instead of uintmax_t.  Additionally, since once_flag::epoch is accessed by multiple threads without any synchronization, it should be marked volatile.

(Alternatively, the new C++0x atomic operations library appears suitable for this use as well.)",Bugs,closed,Boost 1.54.0,thread,Boost 1.47.0,Problem,fixed,once,viboes fmhess@…