Re: [PATCH 2/2] locking/rwsem: Use a return variable in rwsem_spin_on_owner()

From: Paul E. McKenney
Date: Thu Apr 09 2015 - 17:07:25 EST


On Thu, Apr 09, 2015 at 01:58:16PM -0700, Jason Low wrote:
> On Thu, Apr 9, 2015 at 12:58 PM, Paul E. McKenney
> <paulmck@xxxxxxxxxxxxxxxxxx> wrote:
> > On Thu, Apr 09, 2015 at 12:43:38PM -0700, Jason Low wrote:
> >> On Thu, 2015-04-09 at 11:16 -0700, Linus Torvalds wrote:
> >> > On Thu, Apr 9, 2015 at 11:08 AM, Linus Torvalds
> >> > <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
> >> > >
> >> > > The pointer is a known-safe kernel pointer - it's just that it was
> >> > > "known safe" a few instructions ago, and might be rcu-free'd at any
> >> > > time.
> >> >
> >> > Actually, we could even do something like this:
> >> >
> >> > static inline int sem_owner_on_cpu(struct semaphore *sem, struct
> >> > task_struct *owner)
> >> > {
> >> > int on_cpu;
> >> >
> >> > #ifdef CONFIG_DEBUG_PAGEALLOC
> >> > rcu_read_lock();
> >> > #endif
> >> > on_cpu = sem->owner == owner && owner->on_cpu;
> >> > #ifdef CONFIG_DEBUG_PAGEALLOC
> >> > rcu_read_unlock();
> >> > #endif
> >> > return on_cpu;
> >> > }
> >> >
> >> > because we really don't need to hold the RCU lock over the whole loop,
> >> > we just need to validate that the semaphore owner still matches, and
> >> > if so, check that it's on_cpu.
> >> >
> >> > And if CONFIG_DEBUG_PAGEALLOC is set, we don't care about performance
> >> > *at*all*. We will have worse performance problems than doing some RCU
> >> > read-locking inside the loop.
> >> >
> >> > And if CONFIG_DEBUG_PAGEALLOC isn't set, we don't really care about
> >> > locking, since at worst we just access stale memory for one iteration.
> >> >
> >> > Hmm. It's not pretty, but neither is the current "let's just take a
> >> > rcu lock that we don't really need over a loop that doesn't have very
> >> > strict bounding".
> >> >
> >> > Comments?
> >>
> >> So that looks more similar to how the original code was where the
> >> rcu_read_lock() and rcu_read_unlock() was done inside the owner_running
> >> helper function (though without the CONFIG_DEBUG_PAGEALLOC), before
> >> commit 307bf9803f25 ("sched: Simplify mutex_spin_on_owner()") modified
> >> it to be done outside the loop.
> >
> > Another approach would be to post a timer before entering the spinloop,
> > and have the timer handler set the resched bit. Then the loop would
> > be bounded, safe, and would run at full speed.
>
> Though posting a timer, ect... would also add more overhead right?

It would. Not on each pass through the loop, though. But yes, might
be too much overhead.

Thanx, Paul

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/