Re: [question] Linux read/write spinlocks and readers starving awriter

[Manfred Spraul]

Ben LaHaise wrote:
> 
> On Mon, 26 Jun 2000, Tigran Aivazian wrote:
> 
> > I know that Linux read-write semaphores implementation is quite clever
> > with respect to putting the new readers to queue behind the waiting
> > writer. Now, is something similar true for read-write spinlocks also?
> 
> Not at present.  Right now what happens for the contended lock case is
> that the writer will release its bias of the lock and wait for the lock to
> be released by all readers before attempting to reaquire the lock.  Is
> this the right thing to do?  I'm not certain.
> 
But I'm certain :-(

The "no writer in interrupt" rule for rw spinlocks forces us to release
the bias, and without that rule we would need to add lots of _irqsave()
calls.

CPU1:			CPU2
read_lock(&tasklist_lock);
			write_lock_irq(&tasklist_lock);
			* spins
interrupt
within interrupt
read_lock(&tasklist_lock);

And IIRC your initial rw lock rewrite also held the lock bias, and you
had to change it. Perhaps you should add a comment?

--
	Manfred