Re: [PATCH v3] softlockup: remove hung_task_check_count

[Frederic Weisbecker <fweisbec@gmail.com>]

On Fri, Jan 23, 2009 at 05:55:14PM -0800, Mandeep Singh Baines wrote:
> Frédéric Weisbecker (fweisbec@gmail.com) wrote:
> > 2009/1/23 Ingo Molnar <mingo@elte.hu>:
> > >
> > > not sure i like the whole idea of removing the max iterations check. In
> > > theory if there's a _ton_ of tasks, we could spend a lot of time looping
> > > there. So it always looked prudent to limit it somewhat.
> > >
> > 
> > Which means we can loose several of them. Would it hurt to iterate as
> > much as possible along the task list,
> > keeping some care about writers starvation and latency?
> > BTW I thought about the slow work framework, but I can't retrieve
> > it....  But this thread has already a slow priority.
> > 
> > Would it be interesting to provide a way for rwlocks to know if there
> > is writer waiting for the lock?
> 
> Would be cool if that API existed. You could release the CPU and/or lock as
> soon as either was contended for. You'd have the benefits of fine-grained
> locking without the overhead of locking and unlocking multiple time.
> 
> Currently, there is no bit that can tell you there is a writer waiting. You'd
> probably need to change the write_lock() implementation at a minimum. Maybe
> if the first writer left the RW_LOCK_BIAS bit clear and then waited for the
> readers to leave instead of re-trying? That would actually make write_lock()
> more efficient for the 1-writer case since you'd only need to spin doing
> a read in the failure case instead of an atomic_dec and atomic_inc.
> 


This is already what is done in the slow path (in x86):

/* rdi:	pointer to rwlock_t */
ENTRY(__write_lock_failed)
	CFI_STARTPROC
	LOCK_PREFIX
	addl $RW_LOCK_BIAS,(%rdi)
1:	rep
	nop
	cmpl $RW_LOCK_BIAS,(%rdi)
	jne 1b
	LOCK_PREFIX
	subl $RW_LOCK_BIAS,(%rdi)
	jnz  __write_lock_failed
	ret
	CFI_ENDPROC
END(__write_lock_failed)

It spins lurking at the lock value and only if there are no writers nor readers that
own the lock, it restarts its atomic_sub (and then atomic_add in fail case).

And if an implementation of writers_waiting_for_lock() is needed, I guess this
is the perfect place. One atomic_add on a "waiters_count" on entry and an atomic_sub
on it on exit.

Since this is the slow_path, I guess that wouldn't really impact the performances....