Re: [PATCH] lockdep: Allow tuning tracing keys constant.

[Ryan Huang]

Sorry for the late reply.

Our use case is we will create workqueue for each ptr in one device abc
(Rename to abc here without leaking any information)

  ptr->wq = alloc_workqueue("abc_x_%d_y_%d_ptr_%d", WQ_HIGHPRI, 1, x, y, ptr);

where x is 0~7, y is 0~127, ptr is 0~4, and we have a maximum of 8 devices.
So there is a maximum 8x128x5x8=40960 workqueue.
I found the lockdep key is filled with

  register_lock_class: nr_lock_classes=2498, name=_rs.lock
  register_lock_class: nr_lock_classes=2499, name=semaphore->lock
  register_lock_class: nr_lock_classes=2500,
name=(wq_completion)abc_x_0_y_0_ptr_1
  register_lock_class: nr_lock_classes=2501,
name=(work_completion)(&wp->write_work)
  register_lock_class: nr_lock_classes=2502,
name=(wq_completion)abc_x_0_y_0_ptr_2
  register_lock_class: nr_lock_classes=2503,
name=(wq_completion)abc_x_0_y_0_ptr_3
  ...
  register_lock_class: nr_lock_classes=8189,
name=(wq_completion)abc_x_0_y_110_ptr_2
  register_lock_class: nr_lock_classes=8190,
name=(wq_completion)abc_x_0_y_110_ptr_3
  register_lock_class: nr_lock_classes=8191,
name=(wq_completion)abc_x_0_y_110_ptr_4
  register_lock_class: nr_lock_classes=8192,
name=(wq_completion)abc_x_0_y_110_ptr_5

It seems when we queue_work on one workqueue, it will occupy one lock
class. That's why 8k is not enough for our use case.

Regarding my patch, I just move the hard code value into config.
Increasing the value just increase one byte for the held_lock structure,
and 48 bytes for task_struct structure (around 6 u64 variables) in DEBUG
mode.

Let me know if you have a better way!

Thanks
Ryan


On Fri, Sep 30, 2022 at 10:34 AM Waiman Long <longman@xxxxxxxxxx> wrote:
>
>
> On 9/30/22 13:00, Ryan Huang wrote:
> > Tetsuo Handa made a change for tuning lockdep tracing capacity constants
> > [1]. He created following tracing config constants:
> >    - LOCKDEP_BITS
> >    - LOCKDEP_CHAINS_BITS
> >    - LOCKDEP_STACK_TRACE_BITS
> > However there is a missing one for LOCKDEP_KEYS_BITS. We can see this is
> > also hitting the upper limits in [2].
> >
> > [1] https://github.com/torvalds/linux/commit/5dc33592e95534dc8455ce3e9baaaf3dae0fff82
> > [2] https://syzkaller.appspot.com/bug?id=df466e1151a7e45dd880d8c7033f1ad48acebfb4
> >
> > Fixes: 5dc33592e955 ("lockdep: Allow tuning tracing capacity constants.")
> > Signed-off-by: Ryan Huang <tzukui@xxxxxxxxxx>
> > ---
> >   include/linux/lockdep.h | 2 +-
> >   lib/Kconfig.debug       | 8 ++++++++
> >   2 files changed, 9 insertions(+), 1 deletion(-)
> >
> > diff --git a/include/linux/lockdep.h b/include/linux/lockdep.h
> > index 1f1099dac3f05..3bb1740296559 100644
> > --- a/include/linux/lockdep.h
> > +++ b/include/linux/lockdep.h
> > @@ -82,7 +82,7 @@ struct lock_chain {
> >       u64                             chain_key;
> >   };
> >
> > -#define MAX_LOCKDEP_KEYS_BITS                13
> > +#define MAX_LOCKDEP_KEYS_BITS                CONFIG_LOCKDEP_KEYS_BITS
> >   #define MAX_LOCKDEP_KEYS            (1UL << MAX_LOCKDEP_KEYS_BITS)
> >   #define INITIAL_CHAIN_KEY           -1
> >
> > diff --git a/lib/Kconfig.debug b/lib/Kconfig.debug
> > index d3e5f36bb01e0..d15024bd14f1d 100644
> > --- a/lib/Kconfig.debug
> > +++ b/lib/Kconfig.debug
> > @@ -1398,6 +1398,14 @@ config LOCKDEP_CHAINS_BITS
> >       help
> >         Try increasing this value if you hit "BUG: MAX_LOCKDEP_CHAINS too low!" message.
> >
> > +config LOCKDEP_KEYS_BITS
> > +     int "Bitsize for MAX_LOCKDEP_KEYS"
> > +     depends on LOCKDEP && !LOCKDEP_SMALL
> > +     range 10 30
> > +     default 13
> > +     help
> > +       Try increasing this value if you hit "BUG: MAX_LOCKDEP_KEYS too low!" message.
> > +
> >   config LOCKDEP_STACK_TRACE_BITS
> >       int "Bitsize for MAX_STACK_TRACE_ENTRIES"
> >       depends on LOCKDEP && !LOCKDEP_SMALL
>
> The lockdep key is embedded in a bit field within the held_lock
> structure to utilize all the 32 bits of an integer. Changing its size
> will require adjusting other bit fields or expand the bit field size
> from 32 bits to 64 bits. 13 bits allows up to 8k unique lock classes.
> Unless there is good evidence that we are going to run out of that, we
> shouldn't change it.
>
> Thanks,
> Longman
>