Re: [RFC] Avoid mutex starvation when optimistic spinning is disabled
From: Imre Deak
Date: Fri Jul 22 2016 - 15:53:36 EST
On Fri, 2016-07-22 at 12:26 -0700, Davidlohr Bueso wrote:
> On Fri, 22 Jul 2016, Imre Deak wrote:
> > On Fri, 2016-07-22 at 11:03 -0700, Davidlohr Bueso wrote:
> > > On Fri, 22 Jul 2016, Waiman Long wrote:
> > >
> > > > I think making mutex_trylock() fail maybe a bit too far. Do we
> > > > really
> > > > have any real workload that cause starvation problemÂÂbecause
> > > > of
> > > > that.
> > > > Code that does mutex_trylock() in a loop can certainly cause
> > > > lock
> > > > starvation, but it is not how mutex_trylock() is supposed to be
> > > > used.
> > > > We can't build in safeguard for all the possible abuses of the
> > > > mutex
> > > > APIs.
> > >
> > > True, and that's actually why I think that 'fixing' the
> > > !SPIN_ON_OWNER case
> > > is a bit too far in the first place: most of the archs that will
> > > care
> > > about
> > > this already have ARCH_SUPPORTS_ATOMIC_RMW. The extra code for
> > > dealing with
> > > this is not worth it imo.
> > SPIN_ON_OWNER is also disabled in case of DEBUG_MUTEXES, which is
> > the
> > config where I wanted to avoid starvation in the first place.
> Well yes, but know of course that that option is even less common
> archs with non atomic Rmw.
The point is that it's broken atm.