"Patrick Mullaney" [off-list ref] writes:

Would a better approach be?

1) clear the bit when the waiter returns(bottom of sock_wait_for_wmem)
2) add a new wait-queue to the socket to separate the writers
3) another idea?

A large reason this is tricky is because it is essentially lockless. So
alternative would be:

4) you add a spinlock which is always taken around the various condition
checks and bit manipulation. Not 100% sure if it would be bad to 
general SMP scalability (you would need to ensure that the bouncing
lock cache line isn't a performance problem), but at least it would 
be much safer and makes it much easier to verify your change is correct.

Actually the wait queue already has such a lock, so it might not be
that expensive if you switch the wait queue to a lockless version.

-Andi