Re: [PATCH v2 02/12] device-core: Add dev->lock_class to enable device_lock() lockdep validation

[Dan Williams]

On Wed, Apr 13, 2022 at 1:43 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> On Tue, Apr 12, 2022 at 11:01:38PM -0700, Dan Williams wrote:
> > The device_lock() is hidden from lockdep by default because, for
> > example, a device subsystem may do something like:
> >
> > ---
> > device_add(dev1);
> > ...in driver core...
> > device_lock(dev1);
> > bus->probe(dev1); /* where bus->probe() calls driver1_probe() */
> >
> > driver1_probe(struct device *dev)
> > {
> >       ...do some enumeration...
> >       dev2->parent = dev;
> >       /* this triggers probe under device_lock(dev2); */
> >       device_add(dev2);
> > }
> > ---
> >
> > To lockdep, that device_lock(dev2) looks like a deadlock because lockdep
>
> Recursion, you're meaning to say it looks like same lock recursion.

Yes, wrong terminology on my part.

>
> > only sees lock classes, not individual lock instances. All device_lock()
> > instances across the entire kernel are the same class. However, this is
> > not a deadlock in practice because the locking is strictly hierarchical.
> > I.e. device_lock(dev1) is held over device_lock(dev2), but never the
> > reverse.
>
> I have some very vague memories from a conversation with Alan Stern,
> some maybe 10 years ago, where I think he was explaining to me this was
> not in fact a simple hierarchy.
>
> > In order for lockdep to be satisfied and see that it is
> > hierarchical in practice the mutex_lock() call in device_lock() needs to
> > be moved to mutex_lock_nested() where the @subclass argument to
> > mutex_lock_nested() represents the nesting level, i.e.:
>
> That's not an obvious conclusion; lockdep has lots of funny annotations,
> subclasses is just one.
>
> I think the big new development in lockdep since that time with Alan
> Stern is that lockdep now has support for dynamic keys; that is lock
> keys in heap memory (as opposed to static storage).

Ah, I was not aware of that, that should allow for deep cleanups of
this proposal.

>
> > s/device_lock(dev1)/mutex_lock_nested(&dev1->mutex, 1)/
> >
> > s/device_lock(dev2)/mutex_lock_nested(&dev2->mutex, 2)/
> >
> > Now, what if the internals of the device_lock() could be annotated with
> > the right @subclass argument to call mutex_lock_nested()?
> >
> > With device_set_lock_class() a subsystem can optionally add that
> > metadata. The device_lock() still takes dev->mutex, but when
> > dev->lock_class is >= 0 it additionally takes dev->lockdep_mutex with
> > the proper nesting. Unlike dev->mutex, dev->lockdep_mutex is not marked
> > lockdep_set_novalidate_class() and lockdep will become useful... at
> > least for one subsystem at a time.
> >
> > It is still the case that only one subsystem can be using lockdep with
> > lockdep_mutex at a time because different subsystems will collide class
> > numbers. You might say "well, how about subsystem1 gets class ids 0 to 9
> > and subsystem2 gets class ids 10 to 20?". MAX_LOCKDEP_SUBCLASSES is 8,
> > and 8 is just enough class ids for one subsystem of moderate complexity.
>
> Again, that doesn't seem like an obvious suggestion at all. Why not give
> each subsystem a different lock key?
>

Yes, that would also save a source of merge conflicts if every
subsystem needed to add conditional extensions to 'struct device' for
an array of lock metadata.

>
> > diff --git a/include/linux/device.h b/include/linux/device.h
> > index af2576ace130..6083e757e804 100644
> > --- a/include/linux/device.h
> > +++ b/include/linux/device.h
> > @@ -402,6 +402,7 @@ struct dev_msi_info {
> >   * @mutex:   Mutex to synchronize calls to its driver.
> >   * @lockdep_mutex: An optional debug lock that a subsystem can use as a
> >   *           peer lock to gain localized lockdep coverage of the device_lock.
> > + * @lock_class: per-subsystem annotated device lock class
> >   * @bus:     Type of bus device is on.
> >   * @driver:  Which driver has allocated this
> >   * @platform_data: Platform data specific to the device.
> > @@ -501,6 +502,7 @@ struct device {
> >                                          dev_set_drvdata/dev_get_drvdata */
> >  #ifdef CONFIG_PROVE_LOCKING
> >       struct mutex            lockdep_mutex;
> > +     int                     lock_class;
> >  #endif
> >       struct mutex            mutex;  /* mutex to synchronize calls to
> >                                        * its driver.
> > @@ -762,18 +764,100 @@ static inline bool dev_pm_test_driver_flags(struct device *dev, u32 flags)
> >       return !!(dev->power.driver_flags & flags);
> >  }
> >
> > +static inline void device_lock_assert(struct device *dev)
> > +{
> > +     lockdep_assert_held(&dev->mutex);
> > +}
> > +
> >  #ifdef CONFIG_PROVE_LOCKING
> >  static inline void device_lockdep_init(struct device *dev)
> >  {
> >       mutex_init(&dev->lockdep_mutex);
> > +     dev->lock_class = -1;
> >       lockdep_set_novalidate_class(&dev->mutex);
> >  }
> > -#else
> > +
> > +static inline void device_lock(struct device *dev)
> > +{
> > +     /*
> > +      * For double-lock programming errors the kernel will hang
> > +      * trying to acquire @dev->mutex before lockdep can report the
> > +      * problem acquiring @dev->lockdep_mutex, so manually assert
> > +      * before that hang.
> > +      */
> > +     lockdep_assert_not_held(&dev->lockdep_mutex);
> > +
> > +     mutex_lock(&dev->mutex);
> > +     if (dev->lock_class >= 0)
> > +             mutex_lock_nested(&dev->lockdep_mutex, dev->lock_class);
> > +}
> > +
> > +static inline int device_lock_interruptible(struct device *dev)
> > +{
> > +     int rc;
> > +
> > +     lockdep_assert_not_held(&dev->lockdep_mutex);
> > +
> > +     rc = mutex_lock_interruptible(&dev->mutex);
> > +     if (rc || dev->lock_class < 0)
> > +             return rc;
> > +
> > +     return mutex_lock_interruptible_nested(&dev->lockdep_mutex,
> > +                                            dev->lock_class);
> > +}
> > +
> > +static inline int device_trylock(struct device *dev)
> > +{
> > +     if (mutex_trylock(&dev->mutex)) {
> > +             if (dev->lock_class >= 0)
> > +                     mutex_lock_nested(&dev->lockdep_mutex, dev->lock_class);
>
> This must be the weirdest stuff I've seen in a while.
>
> > +             return 1;
> > +     }
> > +
> > +     return 0;
> > +}
> > +
> > +static inline void device_unlock(struct device *dev)
> > +{
> > +     if (dev->lock_class >= 0)
> > +             mutex_unlock(&dev->lockdep_mutex);
> > +     mutex_unlock(&dev->mutex);
> > +}
> > +
> > +/*
> > + * Note: this routine expects that the state of @dev->mutex is stable
> > + * from entry to exit. There is no support for changing lockdep
> > + * validation classes, only enabling and disabling validation.
> > + */
> > +static inline void device_set_lock_class(struct device *dev, int lock_class)
> > +{
> > +     /*
> > +      * Allow for setting or clearing the lock class while the
> > +      * device_lock() is held, in which case the paired nested lock
> > +      * might need to be acquired or released now to accommodate the
> > +      * next device_unlock().
> > +      */
> > +     if (dev->lock_class < 0 && lock_class >= 0) {
> > +             /* Enabling lockdep validation... */
> > +             if (mutex_is_locked(&dev->mutex))
> > +                     mutex_lock_nested(&dev->lockdep_mutex, lock_class);
> > +     } else if (dev->lock_class >= 0 && lock_class < 0) {
> > +             /* Disabling lockdep validation... */
> > +             if (mutex_is_locked(&dev->mutex))
> > +                     mutex_unlock(&dev->lockdep_mutex);
> > +     } else {
> > +             dev_warn(dev,
> > +                      "%s: failed to change lock_class from: %d to %d\n",
> > +                      __func__, dev->lock_class, lock_class);
> > +             return;
> > +     }
> > +     dev->lock_class = lock_class;
> > +}
> > +#else /* !CONFIG_PROVE_LOCKING */
>
> This all reads like something utterly surreal... *WHAT*!?!?

Pile of hacks to workaround cases where the lock_class needed to be
specified after the fact. For example, ACPI does not annotate its
locks, but CXL knows that an "ACPI0017" device will always be the root
of a CXL topology. So CXL subsystem can specify that as lock_class 0.

> If you want lockdep validation for one (or more) dev->mutex instances,
> why not pull them out of the no_validate class and use the normal
> locking?

Sounds perfect, just didn't know how to do that with my current
understanding of how to communicate this to lockdep.

>
> This is all quite insane.

Yes, certainly in comparison to your suggestion on the next patch.
That looks much more sane, and even better I think it allows for
optional lockdep validation without even needing to touch
include/linux/device.h.