On Thu, Jul 27, 2017 at 11:48:53AM -0400, Waiman Long wrote:
> On 07/26/2017 04:17 PM, Prateek Sood wrote:
> > If a spinner is present, there is a chance that the load of
> > rwsem_has_spinner() in rwsem_wake() can be reordered with
> > respect to decrement of rwsem count in __up_write() leading
> > to wakeup being missed.
> > spinning writer up_write caller
> > --------------- -----------------------
> > [S] osq_unlock() [L] osq
> > spin_lock(wait_lock)
> > sem->count=0xFFFFFFFF00000001
> > +0xFFFFFFFF00000000
> > count=sem->count
> > MB
> > sem->count=0xFFFFFFFE00000001
> > -0xFFFFFFFF00000001
> > spin_trylock(wait_lock)
> > return
> > rwsem_try_write_lock(count)
> > spin_unlock(wait_lock)
> > schedule()
> > Reordering of atomic_long_sub_return_release() in __up_write()
> > and rwsem_has_spinner() in rwsem_wake() can cause missing of
> > wakeup in up_write() context. In spinning writer, sem->count
> > and local variable count is 0XFFFFFFFE00000001. It would result
> > in rwsem_try_write_lock() failing to acquire rwsem and spinning
> > writer going to sleep in rwsem_down_write_failed().
> > The smp_rmb() will make sure that the spinner state is
> > consulted after sem->count is updated in up_write context.
> > Signed-off-by: Prateek Sood <prs...@codeaurora.org>
> Did you actually observe that the reordering happens?
> I am not sure if some architectures can actually speculatively execute
> instruction ahead of a branch and then ahead into a function call. I
> know it can happen if the function call is inlined, but rwsem_wake()
> will not be inlined into __up_read() or __up_write().
Branches/control dependencies targeting a read do not necessarily preserve
program order; this is for example the case for PowerPC and ARM.
I'd not expect more than a compiler barrier from a function call (in fact,
not even that if the function happens to be inlined).
> Even if that is the case, I am not sure if smp_rmb() alone is enough to
> guarantee the ordering as I think it will depend on how the
> atomic_long_sub_return_release() is implmented.
AFAICT, the pattern under discussion is MP with:
- a store-release to osq->tail(unlock) followed by a store to sem->count,
separated by a MB (from atomic_long_add_return()) on CPU0;
- a load of sem->count (for atomic_long_sub_return_release()) followed by
a load of osq->tail (rwsem_has_spinner()) on CPU1.
Thus a RMW between the two loads suffices to forbid the weak behaviour.