Re: [PATCH 4/9] firewire: don't use PREPARE_DELAYED_WORK

[James Bottomley <James.Bottomley@HansenPartnership.com>]

On Fri, 2014-02-21 at 18:01 -0500, Peter Hurley wrote:
> On 02/21/2014 11:57 AM, Tejun Heo wrote:
> > Yo,
> >
> > On Fri, Feb 21, 2014 at 11:53:46AM -0500, Peter Hurley wrote:
> >> Ok, I can do that. But AFAIK it'll have to be an smp_rmb(); there is
> >> no mb__after unlock.
> >
> > We do have smp_mb__after_unlock_lock().
> >
> >> [ After thinking about it some, I don't think preventing speculative
> >>    writes before clearing PENDING if useful or necessary, so that's
> >>    why I'm suggesting only the rmb. ]
> >
> > But smp_mb__after_unlock_lock() would be cheaper on most popular
> > archs, I think.
> 
> smp_mb__after_unlock_lock() is only for ordering memory operations
> between two spin-locked sections on either the same lock or by
> the same task/cpu. Like:
> 
>     i = 1
>     spin_unlock(lock1)
>     spin_lock(lock2)
>     smp_mb__after_unlock_lock()
>     j = 1
> 
> This guarantees that the store to j happens after the store to i.
> Without it, a cpu can
> 
>     spin_lock(lock2)
>     j = 1
>     i = 1
>     spin_unlock(lock1)

No the CPU cannot.  If the CPU were allowed to reorder locking
sequences, we'd get speculation induced ABBA deadlocks.  The rules are
quite simple: loads and stores cannot speculate out of critical
sections.  All architectures have barriers in place to prevent this ...
I know from personal experience because the barriers on PARISC were
originally too weak and we did get some speculation out of the critical
sections, which was very nasty to debug.

Stuff may speculate into critical sections from non-critical but never
out of them and critical section boundaries may not reorder to cause an
overlap.

James