Re: [CHECKER] possible deadlock in 2.6.8.1 lockd code

[viro@parcelfarce.linux.theplanet.co.uk]

On Tue, Sep 07, 2004 at 07:57:28PM -0700, Dawson Engler wrote:
> Hi All,
> 
> below is a possible deadlock in the linux-2.6.8.1 lockd code found by a
> static deadlock checker I'm writing.  Let me know if it looks valid and/or
> whether the output is too cryptic.  (Note, the locking dependencies go
> across a bunch of function calls, so the paths may be infeasible.)

It's a BS - down() and lock_kernel() do not form a mutual deadlock.

Consider minimal deadlocked state.  By definition, we can exclude tasks
that didn't manage to get at least one lock (we would still have a deadlocked
set without them and we have chosen a minimal set).  Consider the task
that holds semaphore; since we have a deadlock, it would have to be spinning
in lock_kernel().  That requires another task in our set to be holding BKL _and_
having the timeslice, since BKL is dropped when task loses CPU.  But such
task would not be blocked on anything - it can't be blocked on semaphore
since it is runnable and it can't be blocked on BKL since it's already holding
it.  In other words, it could not be a part of our deadlock.  QED.