On Tue, Aug 22, 2017 at 03:49:22PM +0200, Peter Zijlstra wrote:
> Now, this means I also have to consider the existing
> lock_map_acquire_read() users and if they really wanted to be recursive
> or not. When I change lock_map_acquire_read() to use
> lock_acquire_shared() this annotation no longer suffices and the splat
> comes back.
>
>
> Also, the acquire_read() annotation will (obviously) no longer work to
> cure this problem when we switch to normal read (1), because then the
> generated chain:
>
> W(1) -> A(0) -> C(0) -> W(1)
>
> spells deadlock, since W isn't allowed to recurse.
>
>
> /me goes dig through commit:
>
> e159489baa71 ("workqueue: relax lockdep annotation on flush_work()")
>
> to figure out wth the existing users really want.
Yep, they really want recursive, the pattern there is one work flushing
another on the same workqueue, which ends up being:
Work W1: Work W2: Task:
AR(Q) AR(Q) M(A)
A(W1) A(W2) flush_workqueue(Q)
flush_work(W2) M(A) A(Q)
A(W2) R(W2) R(Q)
R(W2) R(Q)
AR(Q)
R(Q)
R(W1)
R(Q)
should spell deadlock (AQ-QA), and W1 takes Q recursively.
I am however slightly puzzled by the need of flush_work() to take Q,
what deadlock potential is there?
Task: Work-W1: Work-W2:
M(A) AR(Q) AR(Q)
flush_work(W1) A(W1) A(W2)
A(W1) M(A)
R(W1)
AR(Q)
R(Q)
Spells deadlock on AQ-QA, but why? Why is flush_work() linked to any lock
taken inside random other works. If we can get rid of flush_work()'s
usage of Q, we can drop the recursive nature.
It was added by Oleg in commit:
a67da70dc095 ("workqueues: lockdep annotations for flush_work()")
Which has a distinct lack of Changelog. However, that is still very much
the old workqueue code, where I think the annotation makes sense because
that was a single thread running the works consecutively. But I don't
see it making sense for the current workqueue that runs works
concurrently.
TJ, Oleg, can we agree flush_work() no longer needs the dependency on Q?
Also, TJ, what protects @pwq in start_flush_work() at the time of
lock_map_*() ?
Also^2, TJ, what's the purpose of using atomic_long_t for work->data?
All it ever seems to do is atomic_long_read() and atomic_long_set(),
neither of which provides anything stronger than
READ_ONCE()/WRITE_ONCE() respectively.
