Re: [PATCH v7 2/6] MCS Lock: optimizations and extra comments

From: Peter Zijlstra
Date: Mon Jan 20 2014 - 08:59:07 EST


On Thu, Jan 16, 2014 at 04:08:20PM -0800, Tim Chen wrote:
> Remove unnecessary operation and make the cmpxchg(lock, node, NULL) == node
> check in mcs_spin_unlock() likely() as it is likely that a race did not occur
> most of the time.

It might be good to describe why the node->locked=1 is thought
unnecessary. I concur it is, but upon reading this changelog I was left
wondering and had to go read the code and run through the logic to
convince myself.

Having done so, I'm now wondering if we think so for the same reason --
although I'm fairly sure we are.

The argument goes like: everybody only looks at his own ->locked value,
therefore the only one possibly interested in node->locked is the lock
owner. However the lock owner doesn't care what's in it, it simply
assumes its 1 but really doesn't care one way or another.

That said, a possible DEBUG mode might want to actually set it, validate
that all other linked nodes are 0 and upon unlock verify the same before
flipping next->locked to 1.


--
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/