John Maddock schrieb:
The trouble is what do we do with 1.33.1? We could either revert to the 1.32 version (if that is actually reliable, the current regression tests don't show up these problems so we don't really know), or we could yank it out of Boost entirely. I'm leaning towards yanking it out at present, but it's a drastic step, and this presupposes that the original authors of the class can't temporarily fix things for 1.33.1 (they all look to have either gone away or be too busy at present however).
What do other folks think?
I don't know either whether someone relies on the presence of rw mutex being in the boost. However if he/she is using it, the usage itself is dangerous anyhow. In my opinion it is better to omit it until it is fixed. I am also not in the mood to dig through the sources to find out how the design was intended to work. I would also vote for basing the design on a theoretically sound ground (not claiming the current is not) that is understandable too. While trying to find out the reason for the initial reported bug I found myself in a maze of hard to understand details. Perhaps it would even be better not to base the rw mutex on level 0 primitives only, but write it to the native API of the respective platform? Also I can think of a logical extension of the interface to "restricted group mutexes", where there are multiple groups of privileges. At most n presetable members of a certain group are allowed to hold the mutex at the same time. The rw mutex is just a special case in this scenario. I didn't spent much time yet trying to find out if this concept could benefit from template partial specialization too, to provide optimizations for certain common cases. But this definitely needs much more thought! BTW.: I think we should instead put more effort to the level 0. Only recently I have seen some requests on the user list to get access to the locking primitives from various other parts of the library that are in detail currently. Regards, Roland