Re: MSB in rwlock?

[Rich Felker <dalias@libc.org>]

On Tue, Jan 22, 2019 at 09:28:47PM +0100, Markus Wichmann wrote:
> Hi all,
> 
> would someone kindly explain what the highest bit in the rwlock value is
> doing? As far as I can see, it is set whenever one of the locking
> functions enters a contended case, and cleared when the unlock function
> sets the lock to 0 (so when the last reader or the writer unlocks the
> rwlock). But at any other point the bit is masked out, AFAICS. So what
> is it doing? Or is it just a vestige that can be done away with?

In most of the locking primitives, the high bit of the lock word is
used as a maybe-has-waiters flag. It looks like it's redundant with
the waiters count field, but it's not.

If we used *only* this bit, then threads obtaining the lock after
contending for it would have to set it to true (since they couldn't
know if they were the last waiter), thereby performing a
potentially-useless futex wake when they subsequently unlock it.

If on the other hand we used only the waiters count, there would be a
race condition at unlock time: between the time the waiter count was
read and the time the lock word was unlocked, a new waiter could
arrive, and go to sleep on the futex without getting woken up. (The
unlocking thread cannot check the waiter count *after* unlocking the
lock word, since the lock object may no longer exist at that point.)

Rich