Re: [PATCH 1/8] lockdep: fix recursive read lock validation

[Peter Zijlstra]

On Tue, 2008-04-29 at 18:03 +0200, Bart Van Assche wrote:
> On Tue, Apr 29, 2008 at 5:15 PM, Peter Zijlstra <a.p.zijlstra@xxxxxxxxx> wrote:
> >
> > On Tue, 2008-04-29 at 17:03 +0200, Bart Van Assche wrote:
> >  > On Tue, Apr 29, 2008 at 4:57 PM, Peter Zijlstra <a.p.zijlstra@xxxxxxxxx> wrote:
> >  > >
> >  > > On Tue, 2008-04-29 at 15:16 +0200, Bart Van Assche wrote:
> >  > >  > On Tue, Apr 29, 2008 at 2:57 PM, Gautham R Shenoy <ego@xxxxxxxxxx> wrote:
> >  > >  > > Subject: lockdep: fix recursive read lock validation
> >  > >  > >  This means that the following sequence is now invalid, whereas previously
> >  > >  > >  it was considered valid:
> >  > >  > >
> >  > >  > >   rlock(a); rlock(b); runlock(b); runlock(a)
> >  > >  > >   rlock(b); rlock(a);
> >  > >  >
> >  > >  > Why are you marking this sequence as invalid ? Although it can be
> >  > >  > debated whether it is good programming practice to be inconsistent
> >  > >  > about the order of read-locking, the above sequence can't be involved
> >  > >  > in a deadlock.
> >  > >
> >  > >  Not for pure read locks, but when you add write locks to it, it does get
> >  > >  deadlocky. Lockdep does not keep separate chains for read and write
> >  > >  locks.
> >  >
> >  > Nesting writer locks inside reader locks is always a bad idea. So
> >  > please come up with an example of how varying the reader lock nesting
> >  > order can trigger a deadlock (when no writer locks are nested inside
> >  > reader locks and nested writer locks are always nested in the same
> >  > order).
> >
> >  It can't deadlock when only readers are involved, but lockdep will not
> >  be able to distinguish between the cases where only read locks are
> >  involved and a mix of readers and writers is involved.
> >
> >  Hence disallow both.
> >
> >  But hitting this requires you do a series of rather unfortunate things:
> >
> >   1) use recursive locking
> >   2) don't have strict lock order
> >   3) make it work by using read locks only
> >
> >  Seriously, any code that triggers this might want to have its locking
> >  re-throught.
> 
> You did not get my point.
> 
> My point is that if you follow the following locking discipline, a
> deadlock will never be triggered:
> * Always obtain writer locks in a consistent order.
> * Never nest writer locks inside reader locks.
> * Nesting reader locks inside writer locks is okay, and nesting reader
> locks inside other reader locks is also OK.
> 
> Again: if you do not agree with the above, please post an example that
> proves me wrong.

Using a lock that does not allow reader nesting would be cheating,
right?

> Or: whether or not to allow a sequence like "rlock(a); rlock(b);
> runlock(b); runlock(a); rlock(b); rlock(a);" is something we can
> choose. We do not have to forbid this sequence -- we can choose
> whether or not we allow this sequence.

I'm utterly confused now; I never argued that it would get deadlocks;
and I said I choose to not allow it from a lockdep pov. What else do you
want?



--
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/