Re: [RFC][PATCH 4/7] lockdep: Seperate lock ids for read/write acquires

[Peter Zijlstra <a.p.zijlstra@chello.nl>]

On Sun, 2011-04-17 at 11:45 +0200, Peter Zijlstra wrote:
> In order to support recursive read locks we need to support the
> previously mentioned lock state conflict matrix:
> 
>  Conflicting_states(WRITE):             RECURSIVE_READ | READ | WRITE
>  Conflicting_states(READ):                               READ | WRITE
>  Conflicting_states(RECURSIVE_READ):                            WRITE
> 
> Since this introduces asymmetry between recursive read and write, we
> need to split the lock dependency chains such that we can traverse
> WRITE chains without observing RECURSIVE_READ|READ chains. 

So while we split off the WRITE chain from the RECURSIVE_READ|READ
chains, shouldn't we split it in three, because the READ conflict state
only has READ|WRITE, not RECURSIVE_READ.

Therefore a RECURSIVE_READ dependency in the READ chain could throw the
regular READ cycle detector, no?


 A(r) -> B(r) -> C(rr)

 C(w) -> B(r)

would close the cycle and report a problem, but doesn't match the
conflict states.