Re: Avoiding the dentry d_lock on final dput(), part deux: transactional memory

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Tue, Oct 01, 2013 at 05:16:54AM -0700, Paul E. McKenney wrote:
> On Tue, Oct 01, 2013 at 02:52:28PM +1000, Michael Neuling wrote:
> > >> Well we don't have to, I think Mikey wasn't totally clear about that
> > >> "making all registers volatile" business :-) This is just something we
> > >> need to handle in assembly if we are going to reclaim the suspended
> > >> transaction.
> > 
> > Yeah, sorry.  The slow path with all registers as volatile is only
> > needed if we get pre-empted during the transaction.
> > 
> > >>
> > >> So basically, what we need is something along the lines of
> > >> enable_kernel_tm() which checks if there's a suspended user transaction
> > >> and if yes, kills/reclaims it.
> > >>
> > >> Then we also need to handle in our interrupt handlers that we have an
> > >> active/suspended transaction from a kernel state, which we don't deal
> > >> with at this point, and do whatever has to be done to kill it... we
> > >> might get away with something simple if we can state that we only allow
> > >> kernel transactions at task level and not from interrupt/softirq
> > >> contexts, at least initially.
> > >
> > > Call me a coward, but this is starting to sound a bit scary.  ;-)
> > 
> > We are just wanting to prototype it for now to see if we could make it
> > go faster.  If it's worth it, then we'd consider the additional
> > complexity this would bring.
> > 
> > I don't think it'll be that bad, but I'd certainly want to make sure
> > it's worth it before trying :-)
> 
> OK, fair point.  ;-)

That is, a fair point -assuming- that we also try the memory-barrier-free
cmpxchg that Linus suggested...

							Thanx, Paul