BOOST_THREAD_HAS_EINTR_BUG behavior changed

[austin-beer]

In 5b209c2, the code that uses BOOST_THREAD_HAS_EINTR_BUG was moved from pthread/mutex.hpp to pthread/pthread_mutex_scoped_lock.hpp. However, pthread/mutex.hpp includes pthread/pthread_mutex_scoped_lock.hpp before it defines BOOST_THREAD_HAS_EINTR_BUG. Therefore, the code in pthread/pthread_mutex_scoped_lock.hpp no longer sees that BOOST_THREAD_HAS_EINTR_BUG has been defined.

Old Behavior (Before 5b209c2): The fix for the EINTR bug is used unless the user defines BOOST_THREAD_HAS_NO_EINTR_BUG.

New Behavior (After 5b209c2): The fix for the EINTR bug is not used unless the user defines BOOST_THREAD_HAS_EINTR_BUG.

Which behavior do we want moving forward? If the old behavior is desired, then we need to move the following three lines from pthread/mutex.hpp to pthread/pthread_mutex_scoped_lock.hpp:

#ifndef BOOST_THREAD_HAS_NO_EINTR_BUG
#define BOOST_THREAD_HAS_EINTR_BUG
#endif

If the new behavior is desired, then we need to delete the above three lines from pthread/mutex.hpp since they no longer do anything and so are confusing.

I think the new behavior is preferred because very few POSIX implementations suffer from the EINTR bug, but there may be other factors that I'm not aware of.

[viboes]

Ideally this flag should be deduced by a test.

In the absence of an automatic test, I believe we should set it to BOOST_THREAD_HAS_NO_EINTR_BUG or BOOST_THREAD_HAS_EINTR_BUG depending on whether we know it for a specific platform.

In the other cases where we don't know, I believe we should suppose we have the bug until the user states the opposite.

[austin-beer]

I don't think we can easily determine this via a test.

And I don't think we know which specific platforms have this bug. The original ticket (https://svn.boost.org/trac10/ticket/6200) doesn't say.

So as you said we should assume that we have the bug until the user states the opposite. So we need to move

#ifndef BOOST_THREAD_HAS_NO_EINTR_BUG
#define BOOST_THREAD_HAS_EINTR_BUG
#endif

from pthread/mutex.hpp to either pthread/pthread_mutex_scoped_lock.hpp (where it is used) or detail/config.hpp.

[austin-beer]

I can submit a PR if you would like.

[viboes]

Yes, I said "Ideally".

If we don't know "which specific platforms have this bug" we are on the "In the other cases where we don't know" part.

And yes, you can sumbit a PR ;-) please.

[austin-beer]

PR submitted: #276