Possible scenario:

0. Assuming the context wasn't initialized yet
1. Threads A and B loaded HAVE_PREALLOCATED_CONTEXT
2. Thread A is suspended by the OS
3. Thread B initializes the context and stores DONE
4. Thread A is scheduled and swaps DONE for WORKING
5. Thread C loads WORKING
6. Thread A sees the mistake and stores DONE again
7. Thread C can now continue

In this scenario C needlessly waited for A even though the context was initialized. I suggest to use compare exchange to avoid erroneous swapping in the first place.

There's also break missing so B always makes one more iteration of the loop and may even ironically become thread C.