On Thu, Feb 22, 2018 at 06:46:54PM +0100, Peter Zijlstra wrote:
> On Thu, Feb 22, 2018 at 03:08:55PM +0800, Boqun Feng wrote:
> > There are four cases for recursive read lock realted deadlocks:
> > 
> > (--(X..Y)--> means a strong dependency path starts with a --(X*)-->
> > dependency and ends with a --(*Y)-- dependency.)
> > 
> > 1.	An irq-safe lock L1 has a dependency --(*..*)--> to an
> > 	irq-unsafe lock L2.
> > 
> > 2.	An irq-read-safe lock L1 has a dependency --(N..*)--> to an
> > 	irq-unsafe lock L2.
> > 
> > 3.	An irq-safe lock L1 has a dependency --(*..N)--> to an
> > 	irq-read-unsafe lock L2.
> > 
> > 4.	An irq-read-safe lock L1 has a dependency --(N..N)--> to an
> > 	irq-read-unsafe lock L2.
> > 
> > The current check_usage() only checks 1) and 2), so this patch adds
> > checks for 3) and 4) and makes sure when find_usage_{back,for}wards find
> > an irq-read-{,un}safe lock, the traverse path should ends at a
> > dependency --(*N)-->. Note when we search backwards, --(*N)--> indicates
> > a real dependency --(N*)-->.
> 
> This adds 4 __bfs() searches for every new link.
> 
> Can't we make the existing traversals smarter?

Haven't really thought this one through, I will try. But as you said, we
only need to do more searchs for _new_ links, so I think it's the slow
path, would the performance matter that much?

Regards,
Boqun