Re: [PATCH 4/4] lockdep: Fix workqueue crossrelease annotation

[Byungchul Park]

On Tue, Sep 05, 2017 at 11:36:24AM +0200, Peter Zijlstra wrote:
> On Tue, Sep 05, 2017 at 05:57:27PM +0900, Byungchul Park wrote:
> > On Tue, Sep 05, 2017 at 09:19:30AM +0200, Peter Zijlstra wrote:
> > > On Tue, Sep 05, 2017 at 09:08:25AM +0200, Peter Zijlstra wrote:
> > > > So you worry about max_active==1 ? Or you worry about pool->lock or
> > > > about the thread setup? I'm still not sure.
> > > 
> > > So the thing about pool->lock is that its a leaf lock, we take nothing
> > 
> > I think the following sentence is a key, I hope...
> > 
> > Leaf locks can also create dependecies with *crosslocks*. These
> > dependencies are not built between holding locks like typical locks.
> 
> They can create dependencies, but they _cannot_ create deadlocks. So
> there's no value in those dependencies.

Let me show you a possible scenario with a leaf lock:

                        lock(A)
   lock(A)              wait_for_completion(B)
   unlock(A)            ...
   ...                  unlock(A)
   process_one_work()
      work->func()
         complete(B)

It's a deadlock by a lead lock A and completion B.

> > > And the whole setup stuff isn't properly preserved between works in any
> > > case, only the first few works would ever see that history, so why
> > > bother.
> > 
> > As I said in another reply, what about (1), (3) and (5) in my example?
> 
> So for single-threaded workqueues, I'd like to get recursive-read sorted
> and then we can make the lockdep_invariant_state() conditional.
> 
> Using recurisve-read lock for the wq lockdep_map's has the same effect
> as your might thing without having to introduce new magic.

Recursive-read and the hint I proposed(a.k.a. might) should be used for
their different specific applications. Both meaning and constraints of
them are totally different.

Using a right function semantically is more important than making it
just work, as you know. Wrong?