Re: [PATCH V2 0/6][RFC] futex: FUTEX_LOCK with optional adaptive spinning

[Darren Hart <dvhltc@us.ibm.com>]

Thomas Gleixner wrote:
> On Tue, 6 Apr 2010, Ulrich Drepper wrote:
> 
>> On Tue, Apr 6, 2010 at 12:31, Thomas Gleixner <tglx@linutronix.de> wrote:
>>> We need to figure out a more efficient way to
>>> do the spinning in the kernel where we have all the necessary
>>> information already.
>> Really?  The owner information isn't in general available in the
>> kernel.  Futex operation doesn't require the value used to be the PID
>> (or negative of the PID).  That is a dramatic limitation of the
>> usefulness of futexes.
> 
> I know that you can do any weird stuff with the futex value, but I
> don't see the "dramatic" limitation. Care to elaborate ?
> 
>> At userlevel there is access to other fields of the data structure
>> which can contain the owner information.
>>
>> I would like to see the method using a per-thread pinned page and an
>> update of a memory location on scheduling.  For benchmarking at least.
> 
> The per thread pinned page would be unconditional, right ?
> 
> I agree that benchmarking would be interesting, but OTOH I fear that
> we open up a huge can of worms with exposing scheduler details and the
> related necessary syscalls like sys_yield_to: User space thread
> management/scheduling comes to my mind and I hope we agree that we do
> not want to revisit that.
> 
>>  I agree that a sys_yield_to() syscall would be at the very least
>> useful as well.  But it's useful for other things already.
> 
> Useful for what ?
> 
> What are the exact semantics of such a syscall ? 
> 
> How does that fit into the various scheduling constraints ?


I believe this comes back to the discussions of a directed yield. The 
idea being that a thread yields its remaining timeslice to a thread of 
it's choosing - usually because the target thread holds a resource the 
yielding thread needs access to. This makes the yield more explicit so 
the yielding thread is more likely to get some benefit out of yielding.

I believe the arguments would be either a TID or a thread group - 
however that is specified. I believe the KVM guys would like to see 
something like this as well - which might be the "other things" referred 
to above.

-- 
Darren Hart
IBM Linux Technology Center
Real-Time Linux Team