Re: "locking/lockdep: Consolidate lock usage bit initialization" is buggy

[Qian Cai <cai@lca.pw>]

On Fri, 2019-06-07 at 11:21 +0800, Yuyang Du wrote:
> Thanks for the report, but
> 
> On Fri, 7 Jun 2019 at 05:14, Qian Cai <cai@lca.pw> wrote:
> > 
> > The linux-next commit "locking/lockdep: Consolidate lock usage bit
> > initialization" [1] will always generate a warning below.
> 
> I never had such warning.
> 
> > Looking through the
> > commit that when mark_irqflags() returns 1 and check = 1, it will do one
> > less
> > mark_lock() call than it used to.
> 
> The four cases:
> 
> 1. When check == 1 and mark_irqflags() returns 1;
> 2. When check == 1 and mark_irqflags() returns 0;
> 3. When check == 0 and mark_irqflags() returns 1;
> 4. When check == 0 and mark_irqflags() returns 0;
> 
> Before and after have exactly the same code to do.

Reverted the commit on the top of linux-next fixed the issue.

With the commit (triggering the warning
DEBUG_LOCKS_WARN_ON(debug_atomic_read(nr_unused_locks) != nr_unused)),

# cat /proc/lockdep_stats
lock-classes:                         1110 [max: 8192]
stack-trace entries:                     0 [max: 524288]
combined max dependencies:               1
uncategorized locks:                     0
unused locks:                         1110
max locking depth:                      14
debug_locks:                             0

Without the commit (no warning),

# cat /proc/lockdep_stats
lock-classes:                         1110 [max: 8192]
stack-trace entries:                  9932 [max: 524288]
combined max dependencies:               1
uncategorized locks:                  1113
unused locks:                            0
max locking depth:                      14
debug_locks:                             1