[03/18] dma-fence: basic lockdep annotations
diff mbox series

Message ID 20200604081224.863494-4-daniel.vetter@ffwll.ch
State New
Headers show
Series
  • dma-fence lockdep annotations, round 2
Related show

Commit Message

Daniel Vetter June 4, 2020, 8:12 a.m. UTC
Design is similar to the lockdep annotations for workers, but with
some twists:

- We use a read-lock for the execution/worker/completion side, so that
  this explicit annotation can be more liberally sprinkled around.
  With read locks lockdep isn't going to complain if the read-side
  isn't nested the same way under all circumstances, so ABBA deadlocks
  are ok. Which they are, since this is an annotation only.

- We're using non-recursive lockdep read lock mode, since in recursive
  read lock mode lockdep does not catch read side hazards. And we
  _very_ much want read side hazards to be caught. For full details of
  this limitation see

  commit e91498589746065e3ae95d9a00b068e525eec34f
  Author: Peter Zijlstra <peterz@infradead.org>
  Date:   Wed Aug 23 13:13:11 2017 +0200

      locking/lockdep/selftests: Add mixed read-write ABBA tests

- To allow nesting of the read-side explicit annotations we explicitly
  keep track of the nesting. lock_is_held() allows us to do that.

- The wait-side annotation is a write lock, and entirely done within
  dma_fence_wait() for everyone by default.

- To be able to freely annotate helper functions I want to make it ok
  to call dma_fence_begin/end_signalling from soft/hardirq context.
  First attempt was using the hardirq locking context for the write
  side in lockdep, but this forces all normal spinlocks nested within
  dma_fence_begin/end_signalling to be spinlocks. That bollocks.

  The approach now is to simple check in_atomic(), and for these cases
  entirely rely on the might_sleep() check in dma_fence_wait(). That
  will catch any wrong nesting against spinlocks from soft/hardirq
  contexts.

The idea here is that every code path that's critical for eventually
signalling a dma_fence should be annotated with
dma_fence_begin/end_signalling. The annotation ideally starts right
after a dma_fence is published (added to a dma_resv, exposed as a
sync_file fd, attached to a drm_syncobj fd, or anything else that
makes the dma_fence visible to other kernel threads), up to and
including the dma_fence_wait(). Examples are irq handlers, the
scheduler rt threads, the tail of execbuf (after the corresponding
fences are visible), any workers that end up signalling dma_fences and
really anything else. Not annotated should be code paths that only
complete fences opportunistically as the gpu progresses, like e.g.
shrinker/eviction code.

The main class of deadlocks this is supposed to catch are:

Thread A:

	mutex_lock(A);
	mutex_unlock(A);

	dma_fence_signal();

Thread B:

	mutex_lock(A);
	dma_fence_wait();
	mutex_unlock(A);

Thread B is blocked on A signalling the fence, but A never gets around
to that because it cannot acquire the lock A.

Note that dma_fence_wait() is allowed to be nested within
dma_fence_begin/end_signalling sections. To allow this to happen the
read lock needs to be upgraded to a write lock, which means that any
other lock is acquired between the dma_fence_begin_signalling() call and
the call to dma_fence_wait(), and still held, this will result in an
immediate lockdep complaint. The only other option would be to not
annotate such calls, defeating the point. Therefore these annotations
cannot be sprinkled over the code entirely mindless to avoid false
positives.

v2: handle soft/hardirq ctx better against write side and dont forget
EXPORT_SYMBOL, drivers can't use this otherwise.

v3: Kerneldoc.

v4: Some spelling fixes from Mika

Cc: Mika Kuoppala <mika.kuoppala@intel.com>
Cc: Thomas Hellstrom <thomas.hellstrom@intel.com>
Cc: linux-media@vger.kernel.org
Cc: linaro-mm-sig@lists.linaro.org
Cc: linux-rdma@vger.kernel.org
Cc: amd-gfx@lists.freedesktop.org
Cc: intel-gfx@lists.freedesktop.org
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Cc: Christian König <christian.koenig@amd.com>
Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>
---
 Documentation/driver-api/dma-buf.rst |  12 +-
 drivers/dma-buf/dma-fence.c          | 161 +++++++++++++++++++++++++++
 include/linux/dma-fence.h            |  12 ++
 3 files changed, 182 insertions(+), 3 deletions(-)

Comments

Thomas Hellström (Intel) June 4, 2020, 8:57 a.m. UTC | #1
On 6/4/20 10:12 AM, Daniel Vetter wrote:
...
> Thread A:
>
> 	mutex_lock(A);
> 	mutex_unlock(A);
>
> 	dma_fence_signal();
>
> Thread B:
>
> 	mutex_lock(A);
> 	dma_fence_wait();
> 	mutex_unlock(A);
>
> Thread B is blocked on A signalling the fence, but A never gets around
> to that because it cannot acquire the lock A.
>
> Note that dma_fence_wait() is allowed to be nested within
> dma_fence_begin/end_signalling sections. To allow this to happen the
> read lock needs to be upgraded to a write lock, which means that any
> other lock is acquired between the dma_fence_begin_signalling() call and
> the call to dma_fence_wait(), and still held, this will result in an
> immediate lockdep complaint. The only other option would be to not
> annotate such calls, defeating the point. Therefore these annotations
> cannot be sprinkled over the code entirely mindless to avoid false
> positives.

Just realized, isn't that example actually a true positive, or at least 
a great candidate for a true positive, since if another thread reenters 
that signaling path, it will block on that mutex, and the fence would 
never be signaled unless there is another signaling path?

Although I agree the conclusion is sound: These annotations cannot be 
sprinkled mindlessly over the code.

/Thomas






>
> v2: handle soft/hardirq ctx better against write side and dont forget
> EXPORT_SYMBOL, drivers can't use this otherwise.
>
> v3: Kerneldoc.
>
> v4: Some spelling fixes from Mika
>
> Cc: Mika Kuoppala <mika.kuoppala@intel.com>
> Cc: Thomas Hellstrom <thomas.hellstrom@intel.com>
> Cc: linux-media@vger.kernel.org
> Cc: linaro-mm-sig@lists.linaro.org
> Cc: linux-rdma@vger.kernel.org
> Cc: amd-gfx@lists.freedesktop.org
> Cc: intel-gfx@lists.freedesktop.org
> Cc: Chris Wilson <chris@chris-wilson.co.uk>
> Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
> Cc: Christian König <christian.koenig@amd.com>
> Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>
> ---
>   Documentation/driver-api/dma-buf.rst |  12 +-
>   drivers/dma-buf/dma-fence.c          | 161 +++++++++++++++++++++++++++
>   include/linux/dma-fence.h            |  12 ++
>   3 files changed, 182 insertions(+), 3 deletions(-)
>
> diff --git a/Documentation/driver-api/dma-buf.rst b/Documentation/driver-api/dma-buf.rst
> index 63dec76d1d8d..05d856131140 100644
> --- a/Documentation/driver-api/dma-buf.rst
> +++ b/Documentation/driver-api/dma-buf.rst
> @@ -100,11 +100,11 @@ CPU Access to DMA Buffer Objects
>   .. kernel-doc:: drivers/dma-buf/dma-buf.c
>      :doc: cpu access
>   
> -Fence Poll Support
> -~~~~~~~~~~~~~~~~~~
> +Implicit Fence Poll Support
> +~~~~~~~~~~~~~~~~~~~~~~~~~~~
>   
>   .. kernel-doc:: drivers/dma-buf/dma-buf.c
> -   :doc: fence polling
> +   :doc: implicit fence polling
>   
>   Kernel Functions and Structures Reference
>   ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> @@ -133,6 +133,12 @@ DMA Fences
>   .. kernel-doc:: drivers/dma-buf/dma-fence.c
>      :doc: DMA fences overview
>   
> +DMA Fence Signalling Annotations
> +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> +
> +.. kernel-doc:: drivers/dma-buf/dma-fence.c
> +   :doc: fence signalling annotation
> +
>   DMA Fences Functions Reference
>   ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>   
> diff --git a/drivers/dma-buf/dma-fence.c b/drivers/dma-buf/dma-fence.c
> index 656e9ac2d028..0005bc002529 100644
> --- a/drivers/dma-buf/dma-fence.c
> +++ b/drivers/dma-buf/dma-fence.c
> @@ -110,6 +110,160 @@ u64 dma_fence_context_alloc(unsigned num)
>   }
>   EXPORT_SYMBOL(dma_fence_context_alloc);
>   
> +/**
> + * DOC: fence signalling annotation
> + *
> + * Proving correctness of all the kernel code around &dma_fence through code
> + * review and testing is tricky for a few reasons:
> + *
> + * * It is a cross-driver contract, and therefore all drivers must follow the
> + *   same rules for lock nesting order, calling contexts for various functions
> + *   and anything else significant for in-kernel interfaces. But it is also
> + *   impossible to test all drivers in a single machine, hence brute-force N vs.
> + *   N testing of all combinations is impossible. Even just limiting to the
> + *   possible combinations is infeasible.
> + *
> + * * There is an enormous amount of driver code involved. For render drivers
> + *   there's the tail of command submission, after fences are published,
> + *   scheduler code, interrupt and workers to process job completion,
> + *   and timeout, gpu reset and gpu hang recovery code. Plus for integration
> + *   with core mm with have &mmu_notifier, respectively &mmu_interval_notifier,
> + *   and &shrinker. For modesetting drivers there's the commit tail functions
> + *   between when fences for an atomic modeset are published, and when the
> + *   corresponding vblank completes, including any interrupt processing and
> + *   related workers. Auditing all that code, across all drivers, is not
> + *   feasible.
> + *
> + * * Due to how many other subsystems are involved and the locking hierarchies
> + *   this pulls in there is extremely thin wiggle-room for driver-specific
> + *   differences. &dma_fence interacts with almost all of the core memory
> + *   handling through page fault handlers via &dma_resv, dma_resv_lock() and
> + *   dma_resv_unlock(). On the other side it also interacts through all
> + *   allocation sites through &mmu_notifier and &shrinker.
> + *
> + * Furthermore lockdep does not handle cross-release dependencies, which means
> + * any deadlocks between dma_fence_wait() and dma_fence_signal() can't be caught
> + * at runtime with some quick testing. The simplest example is one thread
> + * waiting on a &dma_fence while holding a lock::
> + *
> + *     lock(A);
> + *     dma_fence_wait(B);
> + *     unlock(A);
> + *
> + * while the other thread is stuck trying to acquire the same lock, which
> + * prevents it from signalling the fence the previous thread is stuck waiting
> + * on::
> + *
> + *     lock(A);
> + *     unlock(A);
> + *     dma_fence_signal(B);
> + *
> + * By manually annotating all code relevant to signalling a &dma_fence we can
> + * teach lockdep about these dependencies, which also helps with the validation
> + * headache since now lockdep can check all the rules for us::
> + *
> + *    cookie = dma_fence_begin_signalling();
> + *    lock(A);
> + *    unlock(A);
> + *    dma_fence_signal(B);
> + *    dma_fence_end_signalling(cookie);
> + *
> + * For using dma_fence_begin_signalling() and dma_fence_end_signalling() to
> + * annotate critical sections the following rules need to be observed:
> + *
> + * * All code necessary to complete a &dma_fence must be annotated, from the
> + *   point where a fence is accessible to other threads, to the point where
> + *   dma_fence_signal() is called. Un-annotated code can contain deadlock issues,
> + *   and due to the very strict rules and many corner cases it is infeasible to
> + *   catch these just with review or normal stress testing.
> + *
> + * * &struct dma_resv deserves a special note, since the readers are only
> + *   protected by rcu. This means the signalling critical section starts as soon
> + *   as the new fences are installed, even before dma_resv_unlock() is called.
> + *
> + * * The only exception are fast paths and opportunistic signalling code, which
> + *   calls dma_fence_signal() purely as an optimization, but is not required to
> + *   guarantee completion of a &dma_fence. The usual example is a wait IOCTL
> + *   which calls dma_fence_signal(), while the mandatory completion path goes
> + *   through a hardware interrupt and possible job completion worker.
> + *
> + * * To aid composability of code, the annotations can be freely nested, as long
> + *   as the overall locking hierarchy is consistent. The annotations also work
> + *   both in interrupt and process context. Due to implementation details this
> + *   requires that callers pass an opaque cookie from
> + *   dma_fence_begin_signalling() to dma_fence_end_signalling().
> + *
> + * * Validation against the cross driver contract is implemented by priming
> + *   lockdep with the relevant hierarchy at boot-up. This means even just
> + *   testing with a single device is enough to validate a driver, at least as
> + *   far as deadlocks with dma_fence_wait() against dma_fence_signal() are
> + *   concerned.
> + */
> +#ifdef CONFIG_LOCKDEP
> +struct lockdep_map	dma_fence_lockdep_map = {
> +	.name = "dma_fence_map"
> +};
> +
> +/**
> + * dma_fence_begin_signalling - begin a critical DMA fence signalling section
> + *
> + * Drivers should use this to annotate the beginning of any code section
> + * required to eventually complete &dma_fence by calling dma_fence_signal().
> + *
> + * The end of these critical sections are annotated with
> + * dma_fence_end_signalling().
> + *
> + * Returns:
> + *
> + * Opaque cookie needed by the implementation, which needs to be passed to
> + * dma_fence_end_signalling().
> + */
> +bool dma_fence_begin_signalling(void)
> +{
> +	/* explicitly nesting ... */
> +	if (lock_is_held_type(&dma_fence_lockdep_map, 1))
> +		return true;
> +
> +	/* rely on might_sleep check for soft/hardirq locks */
> +	if (in_atomic())
> +		return true;
> +
> +	/* ... and non-recursive readlock */
> +	lock_acquire(&dma_fence_lockdep_map, 0, 0, 1, 1, NULL, _RET_IP_);
> +
> +	return false;
> +}
> +EXPORT_SYMBOL(dma_fence_begin_signalling);
> +
> +/**
> + * dma_fence_end_signalling - end a critical DMA fence signalling section
> + *
> + * Closes a critical section annotation opened by dma_fence_begin_signalling().
> + */
> +void dma_fence_end_signalling(bool cookie)
> +{
> +	if (cookie)
> +		return;
> +
> +	lock_release(&dma_fence_lockdep_map, _RET_IP_);
> +}
> +EXPORT_SYMBOL(dma_fence_end_signalling);
> +
> +void __dma_fence_might_wait(void)
> +{
> +	bool tmp;
> +
> +	tmp = lock_is_held_type(&dma_fence_lockdep_map, 1);
> +	if (tmp)
> +		lock_release(&dma_fence_lockdep_map, _THIS_IP_);
> +	lock_map_acquire(&dma_fence_lockdep_map);
> +	lock_map_release(&dma_fence_lockdep_map);
> +	if (tmp)
> +		lock_acquire(&dma_fence_lockdep_map, 0, 0, 1, 1, NULL, _THIS_IP_);
> +}
> +#endif
> +
> +
>   /**
>    * dma_fence_signal_locked - signal completion of a fence
>    * @fence: the fence to signal
> @@ -170,14 +324,19 @@ int dma_fence_signal(struct dma_fence *fence)
>   {
>   	unsigned long flags;
>   	int ret;
> +	bool tmp;
>   
>   	if (!fence)
>   		return -EINVAL;
>   
> +	tmp = dma_fence_begin_signalling();
> +
>   	spin_lock_irqsave(fence->lock, flags);
>   	ret = dma_fence_signal_locked(fence);
>   	spin_unlock_irqrestore(fence->lock, flags);
>   
> +	dma_fence_end_signalling(tmp);
> +
>   	return ret;
>   }
>   EXPORT_SYMBOL(dma_fence_signal);
> @@ -210,6 +369,8 @@ dma_fence_wait_timeout(struct dma_fence *fence, bool intr, signed long timeout)
>   
>   	might_sleep();
>   
> +	__dma_fence_might_wait();
> +
>   	trace_dma_fence_wait_start(fence);
>   	if (fence->ops->wait)
>   		ret = fence->ops->wait(fence, intr, timeout);
> diff --git a/include/linux/dma-fence.h b/include/linux/dma-fence.h
> index 3347c54f3a87..3f288f7db2ef 100644
> --- a/include/linux/dma-fence.h
> +++ b/include/linux/dma-fence.h
> @@ -357,6 +357,18 @@ dma_fence_get_rcu_safe(struct dma_fence __rcu **fencep)
>   	} while (1);
>   }
>   
> +#ifdef CONFIG_LOCKDEP
> +bool dma_fence_begin_signalling(void);
> +void dma_fence_end_signalling(bool cookie);
> +#else
> +static inline bool dma_fence_begin_signalling(void)
> +{
> +	return true;
> +}
> +static inline void dma_fence_end_signalling(bool cookie) {}
> +static inline void __dma_fence_might_wait(void) {}
> +#endif
> +
>   int dma_fence_signal(struct dma_fence *fence);
>   int dma_fence_signal_locked(struct dma_fence *fence);
>   signed long dma_fence_default_wait(struct dma_fence *fence,
Daniel Vetter June 4, 2020, 9:21 a.m. UTC | #2
On Thu, Jun 4, 2020 at 10:57 AM Thomas Hellström (Intel)
<thomas_os@shipmail.org> wrote:
>
>
> On 6/4/20 10:12 AM, Daniel Vetter wrote:
> ...
> > Thread A:
> >
> >       mutex_lock(A);
> >       mutex_unlock(A);
> >
> >       dma_fence_signal();
> >
> > Thread B:
> >
> >       mutex_lock(A);
> >       dma_fence_wait();
> >       mutex_unlock(A);
> >
> > Thread B is blocked on A signalling the fence, but A never gets around
> > to that because it cannot acquire the lock A.
> >
> > Note that dma_fence_wait() is allowed to be nested within
> > dma_fence_begin/end_signalling sections. To allow this to happen the
> > read lock needs to be upgraded to a write lock, which means that any
> > other lock is acquired between the dma_fence_begin_signalling() call and
> > the call to dma_fence_wait(), and still held, this will result in an
> > immediate lockdep complaint. The only other option would be to not
> > annotate such calls, defeating the point. Therefore these annotations
> > cannot be sprinkled over the code entirely mindless to avoid false
> > positives.
>
> Just realized, isn't that example actually a true positive, or at least
> a great candidate for a true positive, since if another thread reenters
> that signaling path, it will block on that mutex, and the fence would
> never be signaled unless there is another signaling path?

Not sure I understand fully, but I think the answer is "it's complicated".

dma_fence are meant to be a DAG (directed acyclic graph). Now it would
be nice to enforce that, and i915 has some attempts to that effect,
but these annotations here don't try to pull off that miracle. I'm
assuming that all the dependencies between dma_fence don't create a
loop, and instead I'm only focusing on deadlocks between dma_fences
and other locks. Usually an async work looks like this:

1. wait for a bunch of dma_fence that we have as dependencies
2. do work (e.g. atomic commit)
3. signal the dma_fence that represents our work

This can happen on the cpu in a kthread or worker, or on the gpu. Now
for reasons you might want to have a per-work mutex or something and
hold that while going through all this, and this is the false positive
I'm thinking off. Of course, if your fences aren't a DAG, or if you're
holding a mutex that's shared with some other work which is part of
your dependency chain, then this goes boom. But it doesn't have to.

I think in general it's best to purely rely on ordering, and remove as
much locking as possible. This is the design behind the atomic modeset
commit code, which is does not take any mutexes in the commit path, at
least not in the helpers. Drivers can still do stuff of course. Then
the only locks you're left with are spinlocks (maybe irq safe ones) to
coordinate with interrupt handlers, workers, handle the wait/wake
queues, manage work/scheduler run queues and all that stuff, and no
spinlocks.

Now for the case where you have something like the below:

thread 1:

dma_fence_begin_signalling()
mutex_lock(a);
dma_fence_wait(b1);
mutex_unlock(a);

dma_fence_signal(b2);
dma_fence_end_signalling();

That's indeed a bit problematic, assuming you're annotating stuff
correctly, and the locking is actually required. I've seen a few of
these, and annotating the properly needs care:

- often the mutex_lock/unlock is not needed, and just gets in the way.
This was the case for the original atomic modeset commit work patches,
which again locked all the modeset locks. But strict ordering of
commit work was all that was needed to make this work, plus making
sure data structure lifetimes are handled correctly too. I think the
tendency to abuse locking to handle lifetime and ordering problems is
fairly common, but it can lead to lots of trouble. Ime all async work
items with the above problematic pattern can be fixed like this.

- other often case is that the dma_fence_begin_signalling() can&should
be pushed down past the mutex_lock, and maybe even past the
dma_fence_wait, depending upon when/how the dma_fence is published.
The fence signalling critical section can still extend past the
mutex_unlock, lockdep and semantics are fine with that (I think at
least). This is more the case for execbuf tails, where you take locks,
set up some async work, publish the fences and then begin to process
these fences (which could just be pushing the work to the job
scheduler, but could also involve running it directly in the userspace
process thread context, but with locks already dropped).

So I wouldn't go out and say these are true positives, just maybe
unecessary locking and over-eager annotations, without any real bugs
in the code.

Or am I completely off the track and you're thinking of something else?

> Although I agree the conclusion is sound: These annotations cannot be
> sprinkled mindlessly over the code.

Yup, that much is for sure.
-Daniel


>
> /Thomas
>
>
>
>
>
>
> >
> > v2: handle soft/hardirq ctx better against write side and dont forget
> > EXPORT_SYMBOL, drivers can't use this otherwise.
> >
> > v3: Kerneldoc.
> >
> > v4: Some spelling fixes from Mika
> >
> > Cc: Mika Kuoppala <mika.kuoppala@intel.com>
> > Cc: Thomas Hellstrom <thomas.hellstrom@intel.com>
> > Cc: linux-media@vger.kernel.org
> > Cc: linaro-mm-sig@lists.linaro.org
> > Cc: linux-rdma@vger.kernel.org
> > Cc: amd-gfx@lists.freedesktop.org
> > Cc: intel-gfx@lists.freedesktop.org
> > Cc: Chris Wilson <chris@chris-wilson.co.uk>
> > Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
> > Cc: Christian König <christian.koenig@amd.com>
> > Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>
> > ---
> >   Documentation/driver-api/dma-buf.rst |  12 +-
> >   drivers/dma-buf/dma-fence.c          | 161 +++++++++++++++++++++++++++
> >   include/linux/dma-fence.h            |  12 ++
> >   3 files changed, 182 insertions(+), 3 deletions(-)
> >
> > diff --git a/Documentation/driver-api/dma-buf.rst b/Documentation/driver-api/dma-buf.rst
> > index 63dec76d1d8d..05d856131140 100644
> > --- a/Documentation/driver-api/dma-buf.rst
> > +++ b/Documentation/driver-api/dma-buf.rst
> > @@ -100,11 +100,11 @@ CPU Access to DMA Buffer Objects
> >   .. kernel-doc:: drivers/dma-buf/dma-buf.c
> >      :doc: cpu access
> >
> > -Fence Poll Support
> > -~~~~~~~~~~~~~~~~~~
> > +Implicit Fence Poll Support
> > +~~~~~~~~~~~~~~~~~~~~~~~~~~~
> >
> >   .. kernel-doc:: drivers/dma-buf/dma-buf.c
> > -   :doc: fence polling
> > +   :doc: implicit fence polling
> >
> >   Kernel Functions and Structures Reference
> >   ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> > @@ -133,6 +133,12 @@ DMA Fences
> >   .. kernel-doc:: drivers/dma-buf/dma-fence.c
> >      :doc: DMA fences overview
> >
> > +DMA Fence Signalling Annotations
> > +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> > +
> > +.. kernel-doc:: drivers/dma-buf/dma-fence.c
> > +   :doc: fence signalling annotation
> > +
> >   DMA Fences Functions Reference
> >   ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> >
> > diff --git a/drivers/dma-buf/dma-fence.c b/drivers/dma-buf/dma-fence.c
> > index 656e9ac2d028..0005bc002529 100644
> > --- a/drivers/dma-buf/dma-fence.c
> > +++ b/drivers/dma-buf/dma-fence.c
> > @@ -110,6 +110,160 @@ u64 dma_fence_context_alloc(unsigned num)
> >   }
> >   EXPORT_SYMBOL(dma_fence_context_alloc);
> >
> > +/**
> > + * DOC: fence signalling annotation
> > + *
> > + * Proving correctness of all the kernel code around &dma_fence through code
> > + * review and testing is tricky for a few reasons:
> > + *
> > + * * It is a cross-driver contract, and therefore all drivers must follow the
> > + *   same rules for lock nesting order, calling contexts for various functions
> > + *   and anything else significant for in-kernel interfaces. But it is also
> > + *   impossible to test all drivers in a single machine, hence brute-force N vs.
> > + *   N testing of all combinations is impossible. Even just limiting to the
> > + *   possible combinations is infeasible.
> > + *
> > + * * There is an enormous amount of driver code involved. For render drivers
> > + *   there's the tail of command submission, after fences are published,
> > + *   scheduler code, interrupt and workers to process job completion,
> > + *   and timeout, gpu reset and gpu hang recovery code. Plus for integration
> > + *   with core mm with have &mmu_notifier, respectively &mmu_interval_notifier,
> > + *   and &shrinker. For modesetting drivers there's the commit tail functions
> > + *   between when fences for an atomic modeset are published, and when the
> > + *   corresponding vblank completes, including any interrupt processing and
> > + *   related workers. Auditing all that code, across all drivers, is not
> > + *   feasible.
> > + *
> > + * * Due to how many other subsystems are involved and the locking hierarchies
> > + *   this pulls in there is extremely thin wiggle-room for driver-specific
> > + *   differences. &dma_fence interacts with almost all of the core memory
> > + *   handling through page fault handlers via &dma_resv, dma_resv_lock() and
> > + *   dma_resv_unlock(). On the other side it also interacts through all
> > + *   allocation sites through &mmu_notifier and &shrinker.
> > + *
> > + * Furthermore lockdep does not handle cross-release dependencies, which means
> > + * any deadlocks between dma_fence_wait() and dma_fence_signal() can't be caught
> > + * at runtime with some quick testing. The simplest example is one thread
> > + * waiting on a &dma_fence while holding a lock::
> > + *
> > + *     lock(A);
> > + *     dma_fence_wait(B);
> > + *     unlock(A);
> > + *
> > + * while the other thread is stuck trying to acquire the same lock, which
> > + * prevents it from signalling the fence the previous thread is stuck waiting
> > + * on::
> > + *
> > + *     lock(A);
> > + *     unlock(A);
> > + *     dma_fence_signal(B);
> > + *
> > + * By manually annotating all code relevant to signalling a &dma_fence we can
> > + * teach lockdep about these dependencies, which also helps with the validation
> > + * headache since now lockdep can check all the rules for us::
> > + *
> > + *    cookie = dma_fence_begin_signalling();
> > + *    lock(A);
> > + *    unlock(A);
> > + *    dma_fence_signal(B);
> > + *    dma_fence_end_signalling(cookie);
> > + *
> > + * For using dma_fence_begin_signalling() and dma_fence_end_signalling() to
> > + * annotate critical sections the following rules need to be observed:
> > + *
> > + * * All code necessary to complete a &dma_fence must be annotated, from the
> > + *   point where a fence is accessible to other threads, to the point where
> > + *   dma_fence_signal() is called. Un-annotated code can contain deadlock issues,
> > + *   and due to the very strict rules and many corner cases it is infeasible to
> > + *   catch these just with review or normal stress testing.
> > + *
> > + * * &struct dma_resv deserves a special note, since the readers are only
> > + *   protected by rcu. This means the signalling critical section starts as soon
> > + *   as the new fences are installed, even before dma_resv_unlock() is called.
> > + *
> > + * * The only exception are fast paths and opportunistic signalling code, which
> > + *   calls dma_fence_signal() purely as an optimization, but is not required to
> > + *   guarantee completion of a &dma_fence. The usual example is a wait IOCTL
> > + *   which calls dma_fence_signal(), while the mandatory completion path goes
> > + *   through a hardware interrupt and possible job completion worker.
> > + *
> > + * * To aid composability of code, the annotations can be freely nested, as long
> > + *   as the overall locking hierarchy is consistent. The annotations also work
> > + *   both in interrupt and process context. Due to implementation details this
> > + *   requires that callers pass an opaque cookie from
> > + *   dma_fence_begin_signalling() to dma_fence_end_signalling().
> > + *
> > + * * Validation against the cross driver contract is implemented by priming
> > + *   lockdep with the relevant hierarchy at boot-up. This means even just
> > + *   testing with a single device is enough to validate a driver, at least as
> > + *   far as deadlocks with dma_fence_wait() against dma_fence_signal() are
> > + *   concerned.
> > + */
> > +#ifdef CONFIG_LOCKDEP
> > +struct lockdep_map   dma_fence_lockdep_map = {
> > +     .name = "dma_fence_map"
> > +};
> > +
> > +/**
> > + * dma_fence_begin_signalling - begin a critical DMA fence signalling section
> > + *
> > + * Drivers should use this to annotate the beginning of any code section
> > + * required to eventually complete &dma_fence by calling dma_fence_signal().
> > + *
> > + * The end of these critical sections are annotated with
> > + * dma_fence_end_signalling().
> > + *
> > + * Returns:
> > + *
> > + * Opaque cookie needed by the implementation, which needs to be passed to
> > + * dma_fence_end_signalling().
> > + */
> > +bool dma_fence_begin_signalling(void)
> > +{
> > +     /* explicitly nesting ... */
> > +     if (lock_is_held_type(&dma_fence_lockdep_map, 1))
> > +             return true;
> > +
> > +     /* rely on might_sleep check for soft/hardirq locks */
> > +     if (in_atomic())
> > +             return true;
> > +
> > +     /* ... and non-recursive readlock */
> > +     lock_acquire(&dma_fence_lockdep_map, 0, 0, 1, 1, NULL, _RET_IP_);
> > +
> > +     return false;
> > +}
> > +EXPORT_SYMBOL(dma_fence_begin_signalling);
> > +
> > +/**
> > + * dma_fence_end_signalling - end a critical DMA fence signalling section
> > + *
> > + * Closes a critical section annotation opened by dma_fence_begin_signalling().
> > + */
> > +void dma_fence_end_signalling(bool cookie)
> > +{
> > +     if (cookie)
> > +             return;
> > +
> > +     lock_release(&dma_fence_lockdep_map, _RET_IP_);
> > +}
> > +EXPORT_SYMBOL(dma_fence_end_signalling);
> > +
> > +void __dma_fence_might_wait(void)
> > +{
> > +     bool tmp;
> > +
> > +     tmp = lock_is_held_type(&dma_fence_lockdep_map, 1);
> > +     if (tmp)
> > +             lock_release(&dma_fence_lockdep_map, _THIS_IP_);
> > +     lock_map_acquire(&dma_fence_lockdep_map);
> > +     lock_map_release(&dma_fence_lockdep_map);
> > +     if (tmp)
> > +             lock_acquire(&dma_fence_lockdep_map, 0, 0, 1, 1, NULL, _THIS_IP_);
> > +}
> > +#endif
> > +
> > +
> >   /**
> >    * dma_fence_signal_locked - signal completion of a fence
> >    * @fence: the fence to signal
> > @@ -170,14 +324,19 @@ int dma_fence_signal(struct dma_fence *fence)
> >   {
> >       unsigned long flags;
> >       int ret;
> > +     bool tmp;
> >
> >       if (!fence)
> >               return -EINVAL;
> >
> > +     tmp = dma_fence_begin_signalling();
> > +
> >       spin_lock_irqsave(fence->lock, flags);
> >       ret = dma_fence_signal_locked(fence);
> >       spin_unlock_irqrestore(fence->lock, flags);
> >
> > +     dma_fence_end_signalling(tmp);
> > +
> >       return ret;
> >   }
> >   EXPORT_SYMBOL(dma_fence_signal);
> > @@ -210,6 +369,8 @@ dma_fence_wait_timeout(struct dma_fence *fence, bool intr, signed long timeout)
> >
> >       might_sleep();
> >
> > +     __dma_fence_might_wait();
> > +
> >       trace_dma_fence_wait_start(fence);
> >       if (fence->ops->wait)
> >               ret = fence->ops->wait(fence, intr, timeout);
> > diff --git a/include/linux/dma-fence.h b/include/linux/dma-fence.h
> > index 3347c54f3a87..3f288f7db2ef 100644
> > --- a/include/linux/dma-fence.h
> > +++ b/include/linux/dma-fence.h
> > @@ -357,6 +357,18 @@ dma_fence_get_rcu_safe(struct dma_fence __rcu **fencep)
> >       } while (1);
> >   }
> >
> > +#ifdef CONFIG_LOCKDEP
> > +bool dma_fence_begin_signalling(void);
> > +void dma_fence_end_signalling(bool cookie);
> > +#else
> > +static inline bool dma_fence_begin_signalling(void)
> > +{
> > +     return true;
> > +}
> > +static inline void dma_fence_end_signalling(bool cookie) {}
> > +static inline void __dma_fence_might_wait(void) {}
> > +#endif
> > +
> >   int dma_fence_signal(struct dma_fence *fence);
> >   int dma_fence_signal_locked(struct dma_fence *fence);
> >   signed long dma_fence_default_wait(struct dma_fence *fence,
Chris Wilson June 4, 2020, 9:26 a.m. UTC | #3
Quoting Daniel Vetter (2020-06-04 10:21:46)
> On Thu, Jun 4, 2020 at 10:57 AM Thomas Hellström (Intel)
> <thomas_os@shipmail.org> wrote:
> >
> >
> > On 6/4/20 10:12 AM, Daniel Vetter wrote:
> > ...
> > > Thread A:
> > >
> > >       mutex_lock(A);
> > >       mutex_unlock(A);
> > >
> > >       dma_fence_signal();
> > >
> > > Thread B:
> > >
> > >       mutex_lock(A);
> > >       dma_fence_wait();
> > >       mutex_unlock(A);
> > >
> > > Thread B is blocked on A signalling the fence, but A never gets around
> > > to that because it cannot acquire the lock A.
> > >
> > > Note that dma_fence_wait() is allowed to be nested within
> > > dma_fence_begin/end_signalling sections. To allow this to happen the
> > > read lock needs to be upgraded to a write lock, which means that any
> > > other lock is acquired between the dma_fence_begin_signalling() call and
> > > the call to dma_fence_wait(), and still held, this will result in an
> > > immediate lockdep complaint. The only other option would be to not
> > > annotate such calls, defeating the point. Therefore these annotations
> > > cannot be sprinkled over the code entirely mindless to avoid false
> > > positives.
> >
> > Just realized, isn't that example actually a true positive, or at least
> > a great candidate for a true positive, since if another thread reenters
> > that signaling path, it will block on that mutex, and the fence would
> > never be signaled unless there is another signaling path?
> 
> Not sure I understand fully, but I think the answer is "it's complicated".

See cd8084f91c02 ("locking/lockdep: Apply crossrelease to completions")

dma_fence usage here is nothing but another name for a completion.
-Chris
Daniel Vetter June 4, 2020, 9:36 a.m. UTC | #4
On Thu, Jun 4, 2020 at 11:27 AM Chris Wilson <chris@chris-wilson.co.uk> wrote:
>
> Quoting Daniel Vetter (2020-06-04 10:21:46)
> > On Thu, Jun 4, 2020 at 10:57 AM Thomas Hellström (Intel)
> > <thomas_os@shipmail.org> wrote:
> > >
> > >
> > > On 6/4/20 10:12 AM, Daniel Vetter wrote:
> > > ...
> > > > Thread A:
> > > >
> > > >       mutex_lock(A);
> > > >       mutex_unlock(A);
> > > >
> > > >       dma_fence_signal();
> > > >
> > > > Thread B:
> > > >
> > > >       mutex_lock(A);
> > > >       dma_fence_wait();
> > > >       mutex_unlock(A);
> > > >
> > > > Thread B is blocked on A signalling the fence, but A never gets around
> > > > to that because it cannot acquire the lock A.
> > > >
> > > > Note that dma_fence_wait() is allowed to be nested within
> > > > dma_fence_begin/end_signalling sections. To allow this to happen the
> > > > read lock needs to be upgraded to a write lock, which means that any
> > > > other lock is acquired between the dma_fence_begin_signalling() call and
> > > > the call to dma_fence_wait(), and still held, this will result in an
> > > > immediate lockdep complaint. The only other option would be to not
> > > > annotate such calls, defeating the point. Therefore these annotations
> > > > cannot be sprinkled over the code entirely mindless to avoid false
> > > > positives.
> > >
> > > Just realized, isn't that example actually a true positive, or at least
> > > a great candidate for a true positive, since if another thread reenters
> > > that signaling path, it will block on that mutex, and the fence would
> > > never be signaled unless there is another signaling path?
> >
> > Not sure I understand fully, but I think the answer is "it's complicated".
>
> See cd8084f91c02 ("locking/lockdep: Apply crossrelease to completions")
>
> dma_fence usage here is nothing but another name for a completion.

Quoting from my previous cover letter:

"I've dragged my feet for years on this, hoping that cross-release lockdep
would do this for us, but well that never really happened unfortunately.
So here we are."

I discussed this with Peter, cross-release not getting in is pretty
final it seems. The trouble is false positives without explicit
begin/end annotations reviewed by humans - ime from just these few
examples you just can't guess this stuff by computeres, you need real
brains thinking about all the edge cases, and where exactly the
critical section starts and ends. Without that you're just going to
drown in a sea of false positives and yuck.

So yeah I had hopes for cross-release too, unfortunately that was
entirely in vain and a distraction.

Now I guess it would be nice if there's a per-class
completion_begin/end annotation for the more generic problem. But then
also most people don't have a cross-driver completion api contract
like dma_fence is, with some of the most ridiculous over the top
constraints of what's possible and what's not possible on each side of
the cross-release. We do have a bit an outsized benefit (in pain
reduction) vs cost ratio here.
-Daniel
Tvrtko Ursulin June 10, 2020, 2:21 p.m. UTC | #5
On 04/06/2020 09:12, Daniel Vetter wrote:
> Design is similar to the lockdep annotations for workers, but with
> some twists:
> 
> - We use a read-lock for the execution/worker/completion side, so that
>    this explicit annotation can be more liberally sprinkled around.
>    With read locks lockdep isn't going to complain if the read-side
>    isn't nested the same way under all circumstances, so ABBA deadlocks
>    are ok. Which they are, since this is an annotation only.
> 
> - We're using non-recursive lockdep read lock mode, since in recursive
>    read lock mode lockdep does not catch read side hazards. And we
>    _very_ much want read side hazards to be caught. For full details of
>    this limitation see
> 
>    commit e91498589746065e3ae95d9a00b068e525eec34f
>    Author: Peter Zijlstra <peterz@infradead.org>
>    Date:   Wed Aug 23 13:13:11 2017 +0200
> 
>        locking/lockdep/selftests: Add mixed read-write ABBA tests
> 
> - To allow nesting of the read-side explicit annotations we explicitly
>    keep track of the nesting. lock_is_held() allows us to do that.
> 
> - The wait-side annotation is a write lock, and entirely done within
>    dma_fence_wait() for everyone by default.
> 
> - To be able to freely annotate helper functions I want to make it ok
>    to call dma_fence_begin/end_signalling from soft/hardirq context.
>    First attempt was using the hardirq locking context for the write
>    side in lockdep, but this forces all normal spinlocks nested within
>    dma_fence_begin/end_signalling to be spinlocks. That bollocks.
> 
>    The approach now is to simple check in_atomic(), and for these cases
>    entirely rely on the might_sleep() check in dma_fence_wait(). That
>    will catch any wrong nesting against spinlocks from soft/hardirq
>    contexts.
> 
> The idea here is that every code path that's critical for eventually
> signalling a dma_fence should be annotated with
> dma_fence_begin/end_signalling. The annotation ideally starts right
> after a dma_fence is published (added to a dma_resv, exposed as a
> sync_file fd, attached to a drm_syncobj fd, or anything else that
> makes the dma_fence visible to other kernel threads), up to and
> including the dma_fence_wait(). Examples are irq handlers, the
> scheduler rt threads, the tail of execbuf (after the corresponding
> fences are visible), any workers that end up signalling dma_fences and
> really anything else. Not annotated should be code paths that only
> complete fences opportunistically as the gpu progresses, like e.g.
> shrinker/eviction code.
> 
> The main class of deadlocks this is supposed to catch are:
> 
> Thread A:
> 
> 	mutex_lock(A);
> 	mutex_unlock(A);
> 
> 	dma_fence_signal();
> 
> Thread B:
> 
> 	mutex_lock(A);
> 	dma_fence_wait();
> 	mutex_unlock(A);
> 
> Thread B is blocked on A signalling the fence, but A never gets around
> to that because it cannot acquire the lock A.
> 
> Note that dma_fence_wait() is allowed to be nested within
> dma_fence_begin/end_signalling sections. To allow this to happen the
> read lock needs to be upgraded to a write lock, which means that any
> other lock is acquired between the dma_fence_begin_signalling() call and
> the call to dma_fence_wait(), and still held, this will result in an
> immediate lockdep complaint. The only other option would be to not
> annotate such calls, defeating the point. Therefore these annotations
> cannot be sprinkled over the code entirely mindless to avoid false
> positives.
> 
> v2: handle soft/hardirq ctx better against write side and dont forget
> EXPORT_SYMBOL, drivers can't use this otherwise.
> 
> v3: Kerneldoc.
> 
> v4: Some spelling fixes from Mika
> 
> Cc: Mika Kuoppala <mika.kuoppala@intel.com>
> Cc: Thomas Hellstrom <thomas.hellstrom@intel.com>
> Cc: linux-media@vger.kernel.org
> Cc: linaro-mm-sig@lists.linaro.org
> Cc: linux-rdma@vger.kernel.org
> Cc: amd-gfx@lists.freedesktop.org
> Cc: intel-gfx@lists.freedesktop.org
> Cc: Chris Wilson <chris@chris-wilson.co.uk>
> Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
> Cc: Christian König <christian.koenig@amd.com>
> Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>
> ---
>   Documentation/driver-api/dma-buf.rst |  12 +-
>   drivers/dma-buf/dma-fence.c          | 161 +++++++++++++++++++++++++++
>   include/linux/dma-fence.h            |  12 ++
>   3 files changed, 182 insertions(+), 3 deletions(-)
> 
> diff --git a/Documentation/driver-api/dma-buf.rst b/Documentation/driver-api/dma-buf.rst
> index 63dec76d1d8d..05d856131140 100644
> --- a/Documentation/driver-api/dma-buf.rst
> +++ b/Documentation/driver-api/dma-buf.rst
> @@ -100,11 +100,11 @@ CPU Access to DMA Buffer Objects
>   .. kernel-doc:: drivers/dma-buf/dma-buf.c
>      :doc: cpu access
>   
> -Fence Poll Support
> -~~~~~~~~~~~~~~~~~~
> +Implicit Fence Poll Support
> +~~~~~~~~~~~~~~~~~~~~~~~~~~~
>   
>   .. kernel-doc:: drivers/dma-buf/dma-buf.c
> -   :doc: fence polling
> +   :doc: implicit fence polling
>   
>   Kernel Functions and Structures Reference
>   ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> @@ -133,6 +133,12 @@ DMA Fences
>   .. kernel-doc:: drivers/dma-buf/dma-fence.c
>      :doc: DMA fences overview
>   
> +DMA Fence Signalling Annotations
> +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> +
> +.. kernel-doc:: drivers/dma-buf/dma-fence.c
> +   :doc: fence signalling annotation
> +
>   DMA Fences Functions Reference
>   ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>   
> diff --git a/drivers/dma-buf/dma-fence.c b/drivers/dma-buf/dma-fence.c
> index 656e9ac2d028..0005bc002529 100644
> --- a/drivers/dma-buf/dma-fence.c
> +++ b/drivers/dma-buf/dma-fence.c
> @@ -110,6 +110,160 @@ u64 dma_fence_context_alloc(unsigned num)
>   }
>   EXPORT_SYMBOL(dma_fence_context_alloc);
>   
> +/**
> + * DOC: fence signalling annotation
> + *
> + * Proving correctness of all the kernel code around &dma_fence through code
> + * review and testing is tricky for a few reasons:
> + *
> + * * It is a cross-driver contract, and therefore all drivers must follow the
> + *   same rules for lock nesting order, calling contexts for various functions
> + *   and anything else significant for in-kernel interfaces. But it is also
> + *   impossible to test all drivers in a single machine, hence brute-force N vs.
> + *   N testing of all combinations is impossible. Even just limiting to the
> + *   possible combinations is infeasible.
> + *
> + * * There is an enormous amount of driver code involved. For render drivers
> + *   there's the tail of command submission, after fences are published,
> + *   scheduler code, interrupt and workers to process job completion,
> + *   and timeout, gpu reset and gpu hang recovery code. Plus for integration
> + *   with core mm with have &mmu_notifier, respectively &mmu_interval_notifier,
> + *   and &shrinker. For modesetting drivers there's the commit tail functions
> + *   between when fences for an atomic modeset are published, and when the
> + *   corresponding vblank completes, including any interrupt processing and
> + *   related workers. Auditing all that code, across all drivers, is not
> + *   feasible.
> + *
> + * * Due to how many other subsystems are involved and the locking hierarchies
> + *   this pulls in there is extremely thin wiggle-room for driver-specific
> + *   differences. &dma_fence interacts with almost all of the core memory
> + *   handling through page fault handlers via &dma_resv, dma_resv_lock() and
> + *   dma_resv_unlock(). On the other side it also interacts through all
> + *   allocation sites through &mmu_notifier and &shrinker.
> + *
> + * Furthermore lockdep does not handle cross-release dependencies, which means
> + * any deadlocks between dma_fence_wait() and dma_fence_signal() can't be caught
> + * at runtime with some quick testing. The simplest example is one thread
> + * waiting on a &dma_fence while holding a lock::
> + *
> + *     lock(A);
> + *     dma_fence_wait(B);
> + *     unlock(A);
> + *
> + * while the other thread is stuck trying to acquire the same lock, which
> + * prevents it from signalling the fence the previous thread is stuck waiting
> + * on::
> + *
> + *     lock(A);
> + *     unlock(A);
> + *     dma_fence_signal(B);
> + *
> + * By manually annotating all code relevant to signalling a &dma_fence we can
> + * teach lockdep about these dependencies, which also helps with the validation
> + * headache since now lockdep can check all the rules for us::
> + *
> + *    cookie = dma_fence_begin_signalling();
> + *    lock(A);
> + *    unlock(A);
> + *    dma_fence_signal(B);
> + *    dma_fence_end_signalling(cookie);
> + *
> + * For using dma_fence_begin_signalling() and dma_fence_end_signalling() to
> + * annotate critical sections the following rules need to be observed:
> + *
> + * * All code necessary to complete a &dma_fence must be annotated, from the
> + *   point where a fence is accessible to other threads, to the point where
> + *   dma_fence_signal() is called. Un-annotated code can contain deadlock issues,
> + *   and due to the very strict rules and many corner cases it is infeasible to
> + *   catch these just with review or normal stress testing.
> + *
> + * * &struct dma_resv deserves a special note, since the readers are only
> + *   protected by rcu. This means the signalling critical section starts as soon
> + *   as the new fences are installed, even before dma_resv_unlock() is called.
> + *
> + * * The only exception are fast paths and opportunistic signalling code, which
> + *   calls dma_fence_signal() purely as an optimization, but is not required to
> + *   guarantee completion of a &dma_fence. The usual example is a wait IOCTL
> + *   which calls dma_fence_signal(), while the mandatory completion path goes
> + *   through a hardware interrupt and possible job completion worker.
> + *
> + * * To aid composability of code, the annotations can be freely nested, as long
> + *   as the overall locking hierarchy is consistent. The annotations also work
> + *   both in interrupt and process context. Due to implementation details this
> + *   requires that callers pass an opaque cookie from
> + *   dma_fence_begin_signalling() to dma_fence_end_signalling().
> + *
> + * * Validation against the cross driver contract is implemented by priming
> + *   lockdep with the relevant hierarchy at boot-up. This means even just
> + *   testing with a single device is enough to validate a driver, at least as
> + *   far as deadlocks with dma_fence_wait() against dma_fence_signal() are
> + *   concerned.
> + */
> +#ifdef CONFIG_LOCKDEP
> +struct lockdep_map	dma_fence_lockdep_map = {
> +	.name = "dma_fence_map"
> +};

Maybe a stupid question because this is definitely complicated, but.. If 
you have a single/static/global lockdep map, doesn't this mean _all_ 
locks, from _all_ drivers happening to use dma-fences will get recorded 
in it. Will this work and not cause false positives?

Sounds like it could create a common link between two completely 
unconnected usages. Because below you do add annotations to generic 
dma_fence_signal and dma_fence_wait.

> +
> +/**
> + * dma_fence_begin_signalling - begin a critical DMA fence signalling section
> + *
> + * Drivers should use this to annotate the beginning of any code section
> + * required to eventually complete &dma_fence by calling dma_fence_signal().
> + *
> + * The end of these critical sections are annotated with
> + * dma_fence_end_signalling().
> + *
> + * Returns:
> + *
> + * Opaque cookie needed by the implementation, which needs to be passed to
> + * dma_fence_end_signalling().
> + */
> +bool dma_fence_begin_signalling(void)
> +{
> +	/* explicitly nesting ... */
> +	if (lock_is_held_type(&dma_fence_lockdep_map, 1))
> +		return true;
> +
> +	/* rely on might_sleep check for soft/hardirq locks */
> +	if (in_atomic())
> +		return true;
> +
> +	/* ... and non-recursive readlock */
> +	lock_acquire(&dma_fence_lockdep_map, 0, 0, 1, 1, NULL, _RET_IP_);

Would it work if signalling path would mark itself as a write lock? I am 
thinking it would be nice to see in lockdep splats what are signals and 
what are waits.

The recursive usage wouldn't work then right? Would write annotation on 
the wait path work?

Regards,

Tvrtko

> +
> +	return false;
> +}
> +EXPORT_SYMBOL(dma_fence_begin_signalling);
> +
> +/**
> + * dma_fence_end_signalling - end a critical DMA fence signalling section
> + *
> + * Closes a critical section annotation opened by dma_fence_begin_signalling().
> + */
> +void dma_fence_end_signalling(bool cookie)
> +{
> +	if (cookie)
> +		return;
> +
> +	lock_release(&dma_fence_lockdep_map, _RET_IP_);
> +}
> +EXPORT_SYMBOL(dma_fence_end_signalling);
> +
> +void __dma_fence_might_wait(void)
> +{
> +	bool tmp;
> +
> +	tmp = lock_is_held_type(&dma_fence_lockdep_map, 1);
> +	if (tmp)
> +		lock_release(&dma_fence_lockdep_map, _THIS_IP_);
> +	lock_map_acquire(&dma_fence_lockdep_map);
> +	lock_map_release(&dma_fence_lockdep_map);
> +	if (tmp)
> +		lock_acquire(&dma_fence_lockdep_map, 0, 0, 1, 1, NULL, _THIS_IP_);
> +}
> +#endif
> +
> +
>   /**
>    * dma_fence_signal_locked - signal completion of a fence
>    * @fence: the fence to signal
> @@ -170,14 +324,19 @@ int dma_fence_signal(struct dma_fence *fence)
>   {
>   	unsigned long flags;
>   	int ret;
> +	bool tmp;
>   
>   	if (!fence)
>   		return -EINVAL;
>   
> +	tmp = dma_fence_begin_signalling();
> +
>   	spin_lock_irqsave(fence->lock, flags);
>   	ret = dma_fence_signal_locked(fence);
>   	spin_unlock_irqrestore(fence->lock, flags);
>   
> +	dma_fence_end_signalling(tmp);
> +
>   	return ret;
>   }
>   EXPORT_SYMBOL(dma_fence_signal);
> @@ -210,6 +369,8 @@ dma_fence_wait_timeout(struct dma_fence *fence, bool intr, signed long timeout)
>   
>   	might_sleep();
>   
> +	__dma_fence_might_wait();
> +
>   	trace_dma_fence_wait_start(fence);
>   	if (fence->ops->wait)
>   		ret = fence->ops->wait(fence, intr, timeout);
> diff --git a/include/linux/dma-fence.h b/include/linux/dma-fence.h
> index 3347c54f3a87..3f288f7db2ef 100644
> --- a/include/linux/dma-fence.h
> +++ b/include/linux/dma-fence.h
> @@ -357,6 +357,18 @@ dma_fence_get_rcu_safe(struct dma_fence __rcu **fencep)
>   	} while (1);
>   }
>   
> +#ifdef CONFIG_LOCKDEP
> +bool dma_fence_begin_signalling(void);
> +void dma_fence_end_signalling(bool cookie);
> +#else
> +static inline bool dma_fence_begin_signalling(void)
> +{
> +	return true;
> +}
> +static inline void dma_fence_end_signalling(bool cookie) {}
> +static inline void __dma_fence_might_wait(void) {}
> +#endif
> +
>   int dma_fence_signal(struct dma_fence *fence);
>   int dma_fence_signal_locked(struct dma_fence *fence);
>   signed long dma_fence_default_wait(struct dma_fence *fence,
>
Daniel Vetter June 10, 2020, 3:17 p.m. UTC | #6
On Wed, Jun 10, 2020 at 4:22 PM Tvrtko Ursulin
<tvrtko.ursulin@linux.intel.com> wrote:
>
>
> On 04/06/2020 09:12, Daniel Vetter wrote:
> > Design is similar to the lockdep annotations for workers, but with
> > some twists:
> >
> > - We use a read-lock for the execution/worker/completion side, so that
> >    this explicit annotation can be more liberally sprinkled around.
> >    With read locks lockdep isn't going to complain if the read-side
> >    isn't nested the same way under all circumstances, so ABBA deadlocks
> >    are ok. Which they are, since this is an annotation only.
> >
> > - We're using non-recursive lockdep read lock mode, since in recursive
> >    read lock mode lockdep does not catch read side hazards. And we
> >    _very_ much want read side hazards to be caught. For full details of
> >    this limitation see
> >
> >    commit e91498589746065e3ae95d9a00b068e525eec34f
> >    Author: Peter Zijlstra <peterz@infradead.org>
> >    Date:   Wed Aug 23 13:13:11 2017 +0200
> >
> >        locking/lockdep/selftests: Add mixed read-write ABBA tests
> >
> > - To allow nesting of the read-side explicit annotations we explicitly
> >    keep track of the nesting. lock_is_held() allows us to do that.
> >
> > - The wait-side annotation is a write lock, and entirely done within
> >    dma_fence_wait() for everyone by default.
> >
> > - To be able to freely annotate helper functions I want to make it ok
> >    to call dma_fence_begin/end_signalling from soft/hardirq context.
> >    First attempt was using the hardirq locking context for the write
> >    side in lockdep, but this forces all normal spinlocks nested within
> >    dma_fence_begin/end_signalling to be spinlocks. That bollocks.
> >
> >    The approach now is to simple check in_atomic(), and for these cases
> >    entirely rely on the might_sleep() check in dma_fence_wait(). That
> >    will catch any wrong nesting against spinlocks from soft/hardirq
> >    contexts.
> >
> > The idea here is that every code path that's critical for eventually
> > signalling a dma_fence should be annotated with
> > dma_fence_begin/end_signalling. The annotation ideally starts right
> > after a dma_fence is published (added to a dma_resv, exposed as a
> > sync_file fd, attached to a drm_syncobj fd, or anything else that
> > makes the dma_fence visible to other kernel threads), up to and
> > including the dma_fence_wait(). Examples are irq handlers, the
> > scheduler rt threads, the tail of execbuf (after the corresponding
> > fences are visible), any workers that end up signalling dma_fences and
> > really anything else. Not annotated should be code paths that only
> > complete fences opportunistically as the gpu progresses, like e.g.
> > shrinker/eviction code.
> >
> > The main class of deadlocks this is supposed to catch are:
> >
> > Thread A:
> >
> >       mutex_lock(A);
> >       mutex_unlock(A);
> >
> >       dma_fence_signal();
> >
> > Thread B:
> >
> >       mutex_lock(A);
> >       dma_fence_wait();
> >       mutex_unlock(A);
> >
> > Thread B is blocked on A signalling the fence, but A never gets around
> > to that because it cannot acquire the lock A.
> >
> > Note that dma_fence_wait() is allowed to be nested within
> > dma_fence_begin/end_signalling sections. To allow this to happen the
> > read lock needs to be upgraded to a write lock, which means that any
> > other lock is acquired between the dma_fence_begin_signalling() call and
> > the call to dma_fence_wait(), and still held, this will result in an
> > immediate lockdep complaint. The only other option would be to not
> > annotate such calls, defeating the point. Therefore these annotations
> > cannot be sprinkled over the code entirely mindless to avoid false
> > positives.
> >
> > v2: handle soft/hardirq ctx better against write side and dont forget
> > EXPORT_SYMBOL, drivers can't use this otherwise.
> >
> > v3: Kerneldoc.
> >
> > v4: Some spelling fixes from Mika
> >
> > Cc: Mika Kuoppala <mika.kuoppala@intel.com>
> > Cc: Thomas Hellstrom <thomas.hellstrom@intel.com>
> > Cc: linux-media@vger.kernel.org
> > Cc: linaro-mm-sig@lists.linaro.org
> > Cc: linux-rdma@vger.kernel.org
> > Cc: amd-gfx@lists.freedesktop.org
> > Cc: intel-gfx@lists.freedesktop.org
> > Cc: Chris Wilson <chris@chris-wilson.co.uk>
> > Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
> > Cc: Christian König <christian.koenig@amd.com>
> > Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>
> > ---
> >   Documentation/driver-api/dma-buf.rst |  12 +-
> >   drivers/dma-buf/dma-fence.c          | 161 +++++++++++++++++++++++++++
> >   include/linux/dma-fence.h            |  12 ++
> >   3 files changed, 182 insertions(+), 3 deletions(-)
> >
> > diff --git a/Documentation/driver-api/dma-buf.rst b/Documentation/driver-api/dma-buf.rst
> > index 63dec76d1d8d..05d856131140 100644
> > --- a/Documentation/driver-api/dma-buf.rst
> > +++ b/Documentation/driver-api/dma-buf.rst
> > @@ -100,11 +100,11 @@ CPU Access to DMA Buffer Objects
> >   .. kernel-doc:: drivers/dma-buf/dma-buf.c
> >      :doc: cpu access
> >
> > -Fence Poll Support
> > -~~~~~~~~~~~~~~~~~~
> > +Implicit Fence Poll Support
> > +~~~~~~~~~~~~~~~~~~~~~~~~~~~
> >
> >   .. kernel-doc:: drivers/dma-buf/dma-buf.c
> > -   :doc: fence polling
> > +   :doc: implicit fence polling
> >
> >   Kernel Functions and Structures Reference
> >   ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> > @@ -133,6 +133,12 @@ DMA Fences
> >   .. kernel-doc:: drivers/dma-buf/dma-fence.c
> >      :doc: DMA fences overview
> >
> > +DMA Fence Signalling Annotations
> > +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> > +
> > +.. kernel-doc:: drivers/dma-buf/dma-fence.c
> > +   :doc: fence signalling annotation
> > +
> >   DMA Fences Functions Reference
> >   ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> >
> > diff --git a/drivers/dma-buf/dma-fence.c b/drivers/dma-buf/dma-fence.c
> > index 656e9ac2d028..0005bc002529 100644
> > --- a/drivers/dma-buf/dma-fence.c
> > +++ b/drivers/dma-buf/dma-fence.c
> > @@ -110,6 +110,160 @@ u64 dma_fence_context_alloc(unsigned num)
> >   }
> >   EXPORT_SYMBOL(dma_fence_context_alloc);
> >
> > +/**
> > + * DOC: fence signalling annotation
> > + *
> > + * Proving correctness of all the kernel code around &dma_fence through code
> > + * review and testing is tricky for a few reasons:
> > + *
> > + * * It is a cross-driver contract, and therefore all drivers must follow the
> > + *   same rules for lock nesting order, calling contexts for various functions
> > + *   and anything else significant for in-kernel interfaces. But it is also
> > + *   impossible to test all drivers in a single machine, hence brute-force N vs.
> > + *   N testing of all combinations is impossible. Even just limiting to the
> > + *   possible combinations is infeasible.
> > + *
> > + * * There is an enormous amount of driver code involved. For render drivers
> > + *   there's the tail of command submission, after fences are published,
> > + *   scheduler code, interrupt and workers to process job completion,
> > + *   and timeout, gpu reset and gpu hang recovery code. Plus for integration
> > + *   with core mm with have &mmu_notifier, respectively &mmu_interval_notifier,
> > + *   and &shrinker. For modesetting drivers there's the commit tail functions
> > + *   between when fences for an atomic modeset are published, and when the
> > + *   corresponding vblank completes, including any interrupt processing and
> > + *   related workers. Auditing all that code, across all drivers, is not
> > + *   feasible.
> > + *
> > + * * Due to how many other subsystems are involved and the locking hierarchies
> > + *   this pulls in there is extremely thin wiggle-room for driver-specific
> > + *   differences. &dma_fence interacts with almost all of the core memory
> > + *   handling through page fault handlers via &dma_resv, dma_resv_lock() and
> > + *   dma_resv_unlock(). On the other side it also interacts through all
> > + *   allocation sites through &mmu_notifier and &shrinker.
> > + *
> > + * Furthermore lockdep does not handle cross-release dependencies, which means
> > + * any deadlocks between dma_fence_wait() and dma_fence_signal() can't be caught
> > + * at runtime with some quick testing. The simplest example is one thread
> > + * waiting on a &dma_fence while holding a lock::
> > + *
> > + *     lock(A);
> > + *     dma_fence_wait(B);
> > + *     unlock(A);
> > + *
> > + * while the other thread is stuck trying to acquire the same lock, which
> > + * prevents it from signalling the fence the previous thread is stuck waiting
> > + * on::
> > + *
> > + *     lock(A);
> > + *     unlock(A);
> > + *     dma_fence_signal(B);
> > + *
> > + * By manually annotating all code relevant to signalling a &dma_fence we can
> > + * teach lockdep about these dependencies, which also helps with the validation
> > + * headache since now lockdep can check all the rules for us::
> > + *
> > + *    cookie = dma_fence_begin_signalling();
> > + *    lock(A);
> > + *    unlock(A);
> > + *    dma_fence_signal(B);
> > + *    dma_fence_end_signalling(cookie);
> > + *
> > + * For using dma_fence_begin_signalling() and dma_fence_end_signalling() to
> > + * annotate critical sections the following rules need to be observed:
> > + *
> > + * * All code necessary to complete a &dma_fence must be annotated, from the
> > + *   point where a fence is accessible to other threads, to the point where
> > + *   dma_fence_signal() is called. Un-annotated code can contain deadlock issues,
> > + *   and due to the very strict rules and many corner cases it is infeasible to
> > + *   catch these just with review or normal stress testing.
> > + *
> > + * * &struct dma_resv deserves a special note, since the readers are only
> > + *   protected by rcu. This means the signalling critical section starts as soon
> > + *   as the new fences are installed, even before dma_resv_unlock() is called.
> > + *
> > + * * The only exception are fast paths and opportunistic signalling code, which
> > + *   calls dma_fence_signal() purely as an optimization, but is not required to
> > + *   guarantee completion of a &dma_fence. The usual example is a wait IOCTL
> > + *   which calls dma_fence_signal(), while the mandatory completion path goes
> > + *   through a hardware interrupt and possible job completion worker.
> > + *
> > + * * To aid composability of code, the annotations can be freely nested, as long
> > + *   as the overall locking hierarchy is consistent. The annotations also work
> > + *   both in interrupt and process context. Due to implementation details this
> > + *   requires that callers pass an opaque cookie from
> > + *   dma_fence_begin_signalling() to dma_fence_end_signalling().
> > + *
> > + * * Validation against the cross driver contract is implemented by priming
> > + *   lockdep with the relevant hierarchy at boot-up. This means even just
> > + *   testing with a single device is enough to validate a driver, at least as
> > + *   far as deadlocks with dma_fence_wait() against dma_fence_signal() are
> > + *   concerned.
> > + */
> > +#ifdef CONFIG_LOCKDEP
> > +struct lockdep_map   dma_fence_lockdep_map = {
> > +     .name = "dma_fence_map"
> > +};
>
> Maybe a stupid question because this is definitely complicated, but.. If
> you have a single/static/global lockdep map, doesn't this mean _all_
> locks, from _all_ drivers happening to use dma-fences will get recorded
> in it. Will this work and not cause false positives?
>
> Sounds like it could create a common link between two completely
> unconnected usages. Because below you do add annotations to generic
> dma_fence_signal and dma_fence_wait.

This is fully intentional. dma-fence is a cross-driver interface, if
every driver invents its own rules about how this should work we have
an unmaintainable and unreviewable mess.

I've typed up the full length rant already here:

https://lore.kernel.org/dri-devel/CAKMK7uGnFhbpuurRsnZ4dvRV9gQ_3-rmSJaoqSFY=+Kvepz_CA@mail.gmail.com/

> > +
> > +/**
> > + * dma_fence_begin_signalling - begin a critical DMA fence signalling section
> > + *
> > + * Drivers should use this to annotate the beginning of any code section
> > + * required to eventually complete &dma_fence by calling dma_fence_signal().
> > + *
> > + * The end of these critical sections are annotated with
> > + * dma_fence_end_signalling().
> > + *
> > + * Returns:
> > + *
> > + * Opaque cookie needed by the implementation, which needs to be passed to
> > + * dma_fence_end_signalling().
> > + */
> > +bool dma_fence_begin_signalling(void)
> > +{
> > +     /* explicitly nesting ... */
> > +     if (lock_is_held_type(&dma_fence_lockdep_map, 1))
> > +             return true;
> > +
> > +     /* rely on might_sleep check for soft/hardirq locks */
> > +     if (in_atomic())
> > +             return true;
> > +
> > +     /* ... and non-recursive readlock */
> > +     lock_acquire(&dma_fence_lockdep_map, 0, 0, 1, 1, NULL, _RET_IP_);
>
> Would it work if signalling path would mark itself as a write lock? I am
> thinking it would be nice to see in lockdep splats what are signals and
> what are waits.

Yeah it'd be nice to have a read vs write name for the lock. But we
already have this problem for e.g. flush_work(), from which I've
stolen this idea. So it's not really new. Essentially look at the
backtraces lockdep gives you, and reconstruct the deadlock. I'm hoping
that people will notice the special functions on the backtrace, e.g.
dma_fence_begin_signalling will be listed as offending function/lock
holder, and then read the kerneldoc.

> The recursive usage wouldn't work then right? Would write annotation on
> the wait path work?

Wait path is write annotations already, but yeah annotating the
signalling side as write would cause endless amounts of alse
positives. Also it makes composability of these e.g. what I've done in
amdgpu with annotations in tdr work in drm/scheduler, annotations in
the amdgpu gpu reset code and then also annotations in atomic code,
which all nest within each other in some call chains, but not others.
Dropping the recursion would break that and make it really awkward to
annotate such cases correctly.

And the recursion only works if it's read locks, otherwise lockdep
complains if you have inconsistent annotations on the signalling side
(which again would make it more or less impossible to annotate the
above case fully).

Cheers, Daniel


>
> Regards,
>
> Tvrtko
>
> > +
> > +     return false;
> > +}
> > +EXPORT_SYMBOL(dma_fence_begin_signalling);
> > +
> > +/**
> > + * dma_fence_end_signalling - end a critical DMA fence signalling section
> > + *
> > + * Closes a critical section annotation opened by dma_fence_begin_signalling().
> > + */
> > +void dma_fence_end_signalling(bool cookie)
> > +{
> > +     if (cookie)
> > +             return;
> > +
> > +     lock_release(&dma_fence_lockdep_map, _RET_IP_);
> > +}
> > +EXPORT_SYMBOL(dma_fence_end_signalling);
> > +
> > +void __dma_fence_might_wait(void)
> > +{
> > +     bool tmp;
> > +
> > +     tmp = lock_is_held_type(&dma_fence_lockdep_map, 1);
> > +     if (tmp)
> > +             lock_release(&dma_fence_lockdep_map, _THIS_IP_);
> > +     lock_map_acquire(&dma_fence_lockdep_map);
> > +     lock_map_release(&dma_fence_lockdep_map);
> > +     if (tmp)
> > +             lock_acquire(&dma_fence_lockdep_map, 0, 0, 1, 1, NULL, _THIS_IP_);
> > +}
> > +#endif
> > +
> > +
> >   /**
> >    * dma_fence_signal_locked - signal completion of a fence
> >    * @fence: the fence to signal
> > @@ -170,14 +324,19 @@ int dma_fence_signal(struct dma_fence *fence)
> >   {
> >       unsigned long flags;
> >       int ret;
> > +     bool tmp;
> >
> >       if (!fence)
> >               return -EINVAL;
> >
> > +     tmp = dma_fence_begin_signalling();
> > +
> >       spin_lock_irqsave(fence->lock, flags);
> >       ret = dma_fence_signal_locked(fence);
> >       spin_unlock_irqrestore(fence->lock, flags);
> >
> > +     dma_fence_end_signalling(tmp);
> > +
> >       return ret;
> >   }
> >   EXPORT_SYMBOL(dma_fence_signal);
> > @@ -210,6 +369,8 @@ dma_fence_wait_timeout(struct dma_fence *fence, bool intr, signed long timeout)
> >
> >       might_sleep();
> >
> > +     __dma_fence_might_wait();
> > +
> >       trace_dma_fence_wait_start(fence);
> >       if (fence->ops->wait)
> >               ret = fence->ops->wait(fence, intr, timeout);
> > diff --git a/include/linux/dma-fence.h b/include/linux/dma-fence.h
> > index 3347c54f3a87..3f288f7db2ef 100644
> > --- a/include/linux/dma-fence.h
> > +++ b/include/linux/dma-fence.h
> > @@ -357,6 +357,18 @@ dma_fence_get_rcu_safe(struct dma_fence __rcu **fencep)
> >       } while (1);
> >   }
> >
> > +#ifdef CONFIG_LOCKDEP
> > +bool dma_fence_begin_signalling(void);
> > +void dma_fence_end_signalling(bool cookie);
> > +#else
> > +static inline bool dma_fence_begin_signalling(void)
> > +{
> > +     return true;
> > +}
> > +static inline void dma_fence_end_signalling(bool cookie) {}
> > +static inline void __dma_fence_might_wait(void) {}
> > +#endif
> > +
> >   int dma_fence_signal(struct dma_fence *fence);
> >   int dma_fence_signal_locked(struct dma_fence *fence);
> >   signed long dma_fence_default_wait(struct dma_fence *fence,
> >
Chris Wilson June 11, 2020, 8 a.m. UTC | #7
Quoting Daniel Vetter (2020-06-04 09:12:09)
> Design is similar to the lockdep annotations for workers, but with
> some twists:
> 
> - We use a read-lock for the execution/worker/completion side, so that
>   this explicit annotation can be more liberally sprinkled around.
>   With read locks lockdep isn't going to complain if the read-side
>   isn't nested the same way under all circumstances, so ABBA deadlocks
>   are ok. Which they are, since this is an annotation only.
> 
> - We're using non-recursive lockdep read lock mode, since in recursive
>   read lock mode lockdep does not catch read side hazards. And we
>   _very_ much want read side hazards to be caught. For full details of
>   this limitation see
> 
>   commit e91498589746065e3ae95d9a00b068e525eec34f
>   Author: Peter Zijlstra <peterz@infradead.org>
>   Date:   Wed Aug 23 13:13:11 2017 +0200
> 
>       locking/lockdep/selftests: Add mixed read-write ABBA tests
> 
> - To allow nesting of the read-side explicit annotations we explicitly
>   keep track of the nesting. lock_is_held() allows us to do that.
> 
> - The wait-side annotation is a write lock, and entirely done within
>   dma_fence_wait() for everyone by default.
> 
> - To be able to freely annotate helper functions I want to make it ok
>   to call dma_fence_begin/end_signalling from soft/hardirq context.
>   First attempt was using the hardirq locking context for the write
>   side in lockdep, but this forces all normal spinlocks nested within
>   dma_fence_begin/end_signalling to be spinlocks. That bollocks.
> 
>   The approach now is to simple check in_atomic(), and for these cases
>   entirely rely on the might_sleep() check in dma_fence_wait(). That
>   will catch any wrong nesting against spinlocks from soft/hardirq
>   contexts.
> 
> The idea here is that every code path that's critical for eventually
> signalling a dma_fence should be annotated with
> dma_fence_begin/end_signalling. The annotation ideally starts right
> after a dma_fence is published (added to a dma_resv, exposed as a
> sync_file fd, attached to a drm_syncobj fd, or anything else that
> makes the dma_fence visible to other kernel threads), up to and
> including the dma_fence_wait(). Examples are irq handlers, the
> scheduler rt threads, the tail of execbuf (after the corresponding
> fences are visible), any workers that end up signalling dma_fences and
> really anything else. Not annotated should be code paths that only
> complete fences opportunistically as the gpu progresses, like e.g.
> shrinker/eviction code.
> 
> The main class of deadlocks this is supposed to catch are:
> 
> Thread A:
> 
>         mutex_lock(A);
>         mutex_unlock(A);
> 
>         dma_fence_signal();
> 
> Thread B:
> 
>         mutex_lock(A);
>         dma_fence_wait();
>         mutex_unlock(A);
> 
> Thread B is blocked on A signalling the fence, but A never gets around
> to that because it cannot acquire the lock A.
> 
> Note that dma_fence_wait() is allowed to be nested within
> dma_fence_begin/end_signalling sections. To allow this to happen the
> read lock needs to be upgraded to a write lock, which means that any
> other lock is acquired between the dma_fence_begin_signalling() call and
> the call to dma_fence_wait(), and still held, this will result in an
> immediate lockdep complaint. The only other option would be to not
> annotate such calls, defeating the point. Therefore these annotations
> cannot be sprinkled over the code entirely mindless to avoid false
> positives.
> 
> v2: handle soft/hardirq ctx better against write side and dont forget
> EXPORT_SYMBOL, drivers can't use this otherwise.
> 
> v3: Kerneldoc.
> 
> v4: Some spelling fixes from Mika
> 
> Cc: Mika Kuoppala <mika.kuoppala@intel.com>
> Cc: Thomas Hellstrom <thomas.hellstrom@intel.com>
> Cc: linux-media@vger.kernel.org
> Cc: linaro-mm-sig@lists.linaro.org
> Cc: linux-rdma@vger.kernel.org
> Cc: amd-gfx@lists.freedesktop.org
> Cc: intel-gfx@lists.freedesktop.org
> Cc: Chris Wilson <chris@chris-wilson.co.uk>
> Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
> Cc: Christian König <christian.koenig@amd.com>
> Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>

Introducing a global lockmap that cannot capture the rules correctly,
Nacked-by: Chris Wilson <chris@chris-wilson.co.uk>
-Chris
Dave Airlie June 11, 2020, 8:44 a.m. UTC | #8
On Thu, 11 Jun 2020 at 18:01, Chris Wilson <chris@chris-wilson.co.uk> wrote:
>
> Quoting Daniel Vetter (2020-06-04 09:12:09)
> > Design is similar to the lockdep annotations for workers, but with
> > some twists:
> >
> > - We use a read-lock for the execution/worker/completion side, so that
> >   this explicit annotation can be more liberally sprinkled around.
> >   With read locks lockdep isn't going to complain if the read-side
> >   isn't nested the same way under all circumstances, so ABBA deadlocks
> >   are ok. Which they are, since this is an annotation only.
> >
> > - We're using non-recursive lockdep read lock mode, since in recursive
> >   read lock mode lockdep does not catch read side hazards. And we
> >   _very_ much want read side hazards to be caught. For full details of
> >   this limitation see
> >
> >   commit e91498589746065e3ae95d9a00b068e525eec34f
> >   Author: Peter Zijlstra <peterz@infradead.org>
> >   Date:   Wed Aug 23 13:13:11 2017 +0200
> >
> >       locking/lockdep/selftests: Add mixed read-write ABBA tests
> >
> > - To allow nesting of the read-side explicit annotations we explicitly
> >   keep track of the nesting. lock_is_held() allows us to do that.
> >
> > - The wait-side annotation is a write lock, and entirely done within
> >   dma_fence_wait() for everyone by default.
> >
> > - To be able to freely annotate helper functions I want to make it ok
> >   to call dma_fence_begin/end_signalling from soft/hardirq context.
> >   First attempt was using the hardirq locking context for the write
> >   side in lockdep, but this forces all normal spinlocks nested within
> >   dma_fence_begin/end_signalling to be spinlocks. That bollocks.
> >
> >   The approach now is to simple check in_atomic(), and for these cases
> >   entirely rely on the might_sleep() check in dma_fence_wait(). That
> >   will catch any wrong nesting against spinlocks from soft/hardirq
> >   contexts.
> >
> > The idea here is that every code path that's critical for eventually
> > signalling a dma_fence should be annotated with
> > dma_fence_begin/end_signalling. The annotation ideally starts right
> > after a dma_fence is published (added to a dma_resv, exposed as a
> > sync_file fd, attached to a drm_syncobj fd, or anything else that
> > makes the dma_fence visible to other kernel threads), up to and
> > including the dma_fence_wait(). Examples are irq handlers, the
> > scheduler rt threads, the tail of execbuf (after the corresponding
> > fences are visible), any workers that end up signalling dma_fences and
> > really anything else. Not annotated should be code paths that only
> > complete fences opportunistically as the gpu progresses, like e.g.
> > shrinker/eviction code.
> >
> > The main class of deadlocks this is supposed to catch are:
> >
> > Thread A:
> >
> >         mutex_lock(A);
> >         mutex_unlock(A);
> >
> >         dma_fence_signal();
> >
> > Thread B:
> >
> >         mutex_lock(A);
> >         dma_fence_wait();
> >         mutex_unlock(A);
> >
> > Thread B is blocked on A signalling the fence, but A never gets around
> > to that because it cannot acquire the lock A.
> >
> > Note that dma_fence_wait() is allowed to be nested within
> > dma_fence_begin/end_signalling sections. To allow this to happen the
> > read lock needs to be upgraded to a write lock, which means that any
> > other lock is acquired between the dma_fence_begin_signalling() call and
> > the call to dma_fence_wait(), and still held, this will result in an
> > immediate lockdep complaint. The only other option would be to not
> > annotate such calls, defeating the point. Therefore these annotations
> > cannot be sprinkled over the code entirely mindless to avoid false
> > positives.
> >
> > v2: handle soft/hardirq ctx better against write side and dont forget
> > EXPORT_SYMBOL, drivers can't use this otherwise.
> >
> > v3: Kerneldoc.
> >
> > v4: Some spelling fixes from Mika
> >
> > Cc: Mika Kuoppala <mika.kuoppala@intel.com>
> > Cc: Thomas Hellstrom <thomas.hellstrom@intel.com>
> > Cc: linux-media@vger.kernel.org
> > Cc: linaro-mm-sig@lists.linaro.org
> > Cc: linux-rdma@vger.kernel.org
> > Cc: amd-gfx@lists.freedesktop.org
> > Cc: intel-gfx@lists.freedesktop.org
> > Cc: Chris Wilson <chris@chris-wilson.co.uk>
> > Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
> > Cc: Christian König <christian.koenig@amd.com>
> > Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>
>
> Introducing a global lockmap that cannot capture the rules correctly,

Can you document the rules all drivers should be following then,
because from here it looks to get refactored every version of i915,
and it would be nice if we could all aim for the same set of things
roughly. We've already had enough problems with amdgpu vs i915 vs
everyone else with fences, if this stops that in the future then I'd
rather we have that than just some unwritten rules per driver and
untestable.

Dave.
Daniel Stone June 11, 2020, 9:01 a.m. UTC | #9
Hi,

On Thu, 11 Jun 2020 at 09:44, Dave Airlie <airlied@gmail.com> wrote:
> On Thu, 11 Jun 2020 at 18:01, Chris Wilson <chris@chris-wilson.co.uk> wrote:
> > Introducing a global lockmap that cannot capture the rules correctly,
>
> Can you document the rules all drivers should be following then,
> because from here it looks to get refactored every version of i915,
> and it would be nice if we could all aim for the same set of things
> roughly. We've already had enough problems with amdgpu vs i915 vs
> everyone else with fences, if this stops that in the future then I'd
> rather we have that than just some unwritten rules per driver and
> untestable.

As someone who has sunk a bunch of work into explicit-fencing
awareness in my compositor so I can never be blocked, I'd be
disappointed if the infrastructure was ultimately pointless because
the documented fencing rules were \_o_/ or thereabouts. Lockdep
definitely isn't my area of expertise so I can't comment on the patch
per se, but having something to ensure we don't hit deadlocks sure
seems a lot better than nothing.

Cheers,
Daniel
Tvrtko Ursulin June 11, 2020, 10:36 a.m. UTC | #10
On 10/06/2020 16:17, Daniel Vetter wrote:
> On Wed, Jun 10, 2020 at 4:22 PM Tvrtko Ursulin
> <tvrtko.ursulin@linux.intel.com> wrote:
>>
>>
>> On 04/06/2020 09:12, Daniel Vetter wrote:
>>> Design is similar to the lockdep annotations for workers, but with
>>> some twists:
>>>
>>> - We use a read-lock for the execution/worker/completion side, so that
>>>     this explicit annotation can be more liberally sprinkled around.
>>>     With read locks lockdep isn't going to complain if the read-side
>>>     isn't nested the same way under all circumstances, so ABBA deadlocks
>>>     are ok. Which they are, since this is an annotation only.
>>>
>>> - We're using non-recursive lockdep read lock mode, since in recursive
>>>     read lock mode lockdep does not catch read side hazards. And we
>>>     _very_ much want read side hazards to be caught. For full details of
>>>     this limitation see
>>>
>>>     commit e91498589746065e3ae95d9a00b068e525eec34f
>>>     Author: Peter Zijlstra <peterz@infradead.org>
>>>     Date:   Wed Aug 23 13:13:11 2017 +0200
>>>
>>>         locking/lockdep/selftests: Add mixed read-write ABBA tests
>>>
>>> - To allow nesting of the read-side explicit annotations we explicitly
>>>     keep track of the nesting. lock_is_held() allows us to do that.
>>>
>>> - The wait-side annotation is a write lock, and entirely done within
>>>     dma_fence_wait() for everyone by default.
>>>
>>> - To be able to freely annotate helper functions I want to make it ok
>>>     to call dma_fence_begin/end_signalling from soft/hardirq context.
>>>     First attempt was using the hardirq locking context for the write
>>>     side in lockdep, but this forces all normal spinlocks nested within
>>>     dma_fence_begin/end_signalling to be spinlocks. That bollocks.
>>>
>>>     The approach now is to simple check in_atomic(), and for these cases
>>>     entirely rely on the might_sleep() check in dma_fence_wait(). That
>>>     will catch any wrong nesting against spinlocks from soft/hardirq
>>>     contexts.
>>>
>>> The idea here is that every code path that's critical for eventually
>>> signalling a dma_fence should be annotated with
>>> dma_fence_begin/end_signalling. The annotation ideally starts right
>>> after a dma_fence is published (added to a dma_resv, exposed as a
>>> sync_file fd, attached to a drm_syncobj fd, or anything else that
>>> makes the dma_fence visible to other kernel threads), up to and
>>> including the dma_fence_wait(). Examples are irq handlers, the
>>> scheduler rt threads, the tail of execbuf (after the corresponding
>>> fences are visible), any workers that end up signalling dma_fences and
>>> really anything else. Not annotated should be code paths that only
>>> complete fences opportunistically as the gpu progresses, like e.g.
>>> shrinker/eviction code.
>>>
>>> The main class of deadlocks this is supposed to catch are:
>>>
>>> Thread A:
>>>
>>>        mutex_lock(A);
>>>        mutex_unlock(A);
>>>
>>>        dma_fence_signal();
>>>
>>> Thread B:
>>>
>>>        mutex_lock(A);
>>>        dma_fence_wait();
>>>        mutex_unlock(A);
>>>
>>> Thread B is blocked on A signalling the fence, but A never gets around
>>> to that because it cannot acquire the lock A.
>>>
>>> Note that dma_fence_wait() is allowed to be nested within
>>> dma_fence_begin/end_signalling sections. To allow this to happen the
>>> read lock needs to be upgraded to a write lock, which means that any
>>> other lock is acquired between the dma_fence_begin_signalling() call and
>>> the call to dma_fence_wait(), and still held, this will result in an
>>> immediate lockdep complaint. The only other option would be to not
>>> annotate such calls, defeating the point. Therefore these annotations
>>> cannot be sprinkled over the code entirely mindless to avoid false
>>> positives.
>>>
>>> v2: handle soft/hardirq ctx better against write side and dont forget
>>> EXPORT_SYMBOL, drivers can't use this otherwise.
>>>
>>> v3: Kerneldoc.
>>>
>>> v4: Some spelling fixes from Mika
>>>
>>> Cc: Mika Kuoppala <mika.kuoppala@intel.com>
>>> Cc: Thomas Hellstrom <thomas.hellstrom@intel.com>
>>> Cc: linux-media@vger.kernel.org
>>> Cc: linaro-mm-sig@lists.linaro.org
>>> Cc: linux-rdma@vger.kernel.org
>>> Cc: amd-gfx@lists.freedesktop.org
>>> Cc: intel-gfx@lists.freedesktop.org
>>> Cc: Chris Wilson <chris@chris-wilson.co.uk>
>>> Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
>>> Cc: Christian König <christian.koenig@amd.com>
>>> Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>
>>> ---
>>>    Documentation/driver-api/dma-buf.rst |  12 +-
>>>    drivers/dma-buf/dma-fence.c          | 161 +++++++++++++++++++++++++++
>>>    include/linux/dma-fence.h            |  12 ++
>>>    3 files changed, 182 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/Documentation/driver-api/dma-buf.rst b/Documentation/driver-api/dma-buf.rst
>>> index 63dec76d1d8d..05d856131140 100644
>>> --- a/Documentation/driver-api/dma-buf.rst
>>> +++ b/Documentation/driver-api/dma-buf.rst
>>> @@ -100,11 +100,11 @@ CPU Access to DMA Buffer Objects
>>>    .. kernel-doc:: drivers/dma-buf/dma-buf.c
>>>       :doc: cpu access
>>>
>>> -Fence Poll Support
>>> -~~~~~~~~~~~~~~~~~~
>>> +Implicit Fence Poll Support
>>> +~~~~~~~~~~~~~~~~~~~~~~~~~~~
>>>
>>>    .. kernel-doc:: drivers/dma-buf/dma-buf.c
>>> -   :doc: fence polling
>>> +   :doc: implicit fence polling
>>>
>>>    Kernel Functions and Structures Reference
>>>    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>>> @@ -133,6 +133,12 @@ DMA Fences
>>>    .. kernel-doc:: drivers/dma-buf/dma-fence.c
>>>       :doc: DMA fences overview
>>>
>>> +DMA Fence Signalling Annotations
>>> +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>>> +
>>> +.. kernel-doc:: drivers/dma-buf/dma-fence.c
>>> +   :doc: fence signalling annotation
>>> +
>>>    DMA Fences Functions Reference
>>>    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>>>
>>> diff --git a/drivers/dma-buf/dma-fence.c b/drivers/dma-buf/dma-fence.c
>>> index 656e9ac2d028..0005bc002529 100644
>>> --- a/drivers/dma-buf/dma-fence.c
>>> +++ b/drivers/dma-buf/dma-fence.c
>>> @@ -110,6 +110,160 @@ u64 dma_fence_context_alloc(unsigned num)
>>>    }
>>>    EXPORT_SYMBOL(dma_fence_context_alloc);
>>>
>>> +/**
>>> + * DOC: fence signalling annotation
>>> + *
>>> + * Proving correctness of all the kernel code around &dma_fence through code
>>> + * review and testing is tricky for a few reasons:
>>> + *
>>> + * * It is a cross-driver contract, and therefore all drivers must follow the
>>> + *   same rules for lock nesting order, calling contexts for various functions
>>> + *   and anything else significant for in-kernel interfaces. But it is also
>>> + *   impossible to test all drivers in a single machine, hence brute-force N vs.
>>> + *   N testing of all combinations is impossible. Even just limiting to the
>>> + *   possible combinations is infeasible.
>>> + *
>>> + * * There is an enormous amount of driver code involved. For render drivers
>>> + *   there's the tail of command submission, after fences are published,
>>> + *   scheduler code, interrupt and workers to process job completion,
>>> + *   and timeout, gpu reset and gpu hang recovery code. Plus for integration
>>> + *   with core mm with have &mmu_notifier, respectively &mmu_interval_notifier,
>>> + *   and &shrinker. For modesetting drivers there's the commit tail functions
>>> + *   between when fences for an atomic modeset are published, and when the
>>> + *   corresponding vblank completes, including any interrupt processing and
>>> + *   related workers. Auditing all that code, across all drivers, is not
>>> + *   feasible.
>>> + *
>>> + * * Due to how many other subsystems are involved and the locking hierarchies
>>> + *   this pulls in there is extremely thin wiggle-room for driver-specific
>>> + *   differences. &dma_fence interacts with almost all of the core memory
>>> + *   handling through page fault handlers via &dma_resv, dma_resv_lock() and
>>> + *   dma_resv_unlock(). On the other side it also interacts through all
>>> + *   allocation sites through &mmu_notifier and &shrinker.
>>> + *
>>> + * Furthermore lockdep does not handle cross-release dependencies, which means
>>> + * any deadlocks between dma_fence_wait() and dma_fence_signal() can't be caught
>>> + * at runtime with some quick testing. The simplest example is one thread
>>> + * waiting on a &dma_fence while holding a lock::
>>> + *
>>> + *     lock(A);
>>> + *     dma_fence_wait(B);
>>> + *     unlock(A);
>>> + *
>>> + * while the other thread is stuck trying to acquire the same lock, which
>>> + * prevents it from signalling the fence the previous thread is stuck waiting
>>> + * on::
>>> + *
>>> + *     lock(A);
>>> + *     unlock(A);
>>> + *     dma_fence_signal(B);
>>> + *
>>> + * By manually annotating all code relevant to signalling a &dma_fence we can
>>> + * teach lockdep about these dependencies, which also helps with the validation
>>> + * headache since now lockdep can check all the rules for us::
>>> + *
>>> + *    cookie = dma_fence_begin_signalling();
>>> + *    lock(A);
>>> + *    unlock(A);
>>> + *    dma_fence_signal(B);
>>> + *    dma_fence_end_signalling(cookie);
>>> + *
>>> + * For using dma_fence_begin_signalling() and dma_fence_end_signalling() to
>>> + * annotate critical sections the following rules need to be observed:
>>> + *
>>> + * * All code necessary to complete a &dma_fence must be annotated, from the
>>> + *   point where a fence is accessible to other threads, to the point where
>>> + *   dma_fence_signal() is called. Un-annotated code can contain deadlock issues,
>>> + *   and due to the very strict rules and many corner cases it is infeasible to
>>> + *   catch these just with review or normal stress testing.
>>> + *
>>> + * * &struct dma_resv deserves a special note, since the readers are only
>>> + *   protected by rcu. This means the signalling critical section starts as soon
>>> + *   as the new fences are installed, even before dma_resv_unlock() is called.
>>> + *
>>> + * * The only exception are fast paths and opportunistic signalling code, which
>>> + *   calls dma_fence_signal() purely as an optimization, but is not required to
>>> + *   guarantee completion of a &dma_fence. The usual example is a wait IOCTL
>>> + *   which calls dma_fence_signal(), while the mandatory completion path goes
>>> + *   through a hardware interrupt and possible job completion worker.
>>> + *
>>> + * * To aid composability of code, the annotations can be freely nested, as long
>>> + *   as the overall locking hierarchy is consistent. The annotations also work
>>> + *   both in interrupt and process context. Due to implementation details this
>>> + *   requires that callers pass an opaque cookie from
>>> + *   dma_fence_begin_signalling() to dma_fence_end_signalling().
>>> + *
>>> + * * Validation against the cross driver contract is implemented by priming
>>> + *   lockdep with the relevant hierarchy at boot-up. This means even just
>>> + *   testing with a single device is enough to validate a driver, at least as
>>> + *   far as deadlocks with dma_fence_wait() against dma_fence_signal() are
>>> + *   concerned.
>>> + */
>>> +#ifdef CONFIG_LOCKDEP
>>> +struct lockdep_map   dma_fence_lockdep_map = {
>>> +     .name = "dma_fence_map"
>>> +};
>>
>> Maybe a stupid question because this is definitely complicated, but.. If
>> you have a single/static/global lockdep map, doesn't this mean _all_
>> locks, from _all_ drivers happening to use dma-fences will get recorded
>> in it. Will this work and not cause false positives?
>>
>> Sounds like it could create a common link between two completely
>> unconnected usages. Because below you do add annotations to generic
>> dma_fence_signal and dma_fence_wait.
> 
> This is fully intentional. dma-fence is a cross-driver interface, if
> every driver invents its own rules about how this should work we have
> an unmaintainable and unreviewable mess.
> 
> I've typed up the full length rant already here:
> 
> https://lore.kernel.org/dri-devel/CAKMK7uGnFhbpuurRsnZ4dvRV9gQ_3-rmSJaoqSFY=+Kvepz_CA@mail.gmail.com/

But "perfect storm" of:

 + global fence lockmap
 + mmu notifiers
 + fs reclaim
 + default annotations in dma_fence_signal / dma_fence_wait

Equals to anything ever using dma_fence will be in impossible chains with random other drivers, even if neither driver has code to export/share that fence.

Example from the CI run:

 [25.918788] Chain exists of:
  fs_reclaim --> mmu_notifier_invalidate_range_start --> dma_fence_map
 [25.918794]  Possible unsafe locking scenario:
 [25.918797]        CPU0                    CPU1
 [25.918799]        ----                    ----
 [25.918801]   lock(dma_fence_map);
 [25.918803]                                lock(mmu_notifier_invalidate_range_start);
 [25.918807]                                lock(dma_fence_map);
 [25.918809]   lock(fs_reclaim);

What about a dma_fence_export helper which would "arm" the annotations? It would be called as soon as the fence is exported. Maybe when added to dma_resv, or exported via sync_file, etc. Before that point begin/end_signaling and so would be no-ops. 

>>> +
>>> +/**
>>> + * dma_fence_begin_signalling - begin a critical DMA fence signalling section
>>> + *
>>> + * Drivers should use this to annotate the beginning of any code section
>>> + * required to eventually complete &dma_fence by calling dma_fence_signal().
>>> + *
>>> + * The end of these critical sections are annotated with
>>> + * dma_fence_end_signalling().
>>> + *
>>> + * Returns:
>>> + *
>>> + * Opaque cookie needed by the implementation, which needs to be passed to
>>> + * dma_fence_end_signalling().
>>> + */
>>> +bool dma_fence_begin_signalling(void)
>>> +{
>>> +     /* explicitly nesting ... */
>>> +     if (lock_is_held_type(&dma_fence_lockdep_map, 1))
>>> +             return true;
>>> +
>>> +     /* rely on might_sleep check for soft/hardirq locks */
>>> +     if (in_atomic())
>>> +             return true;
>>> +
>>> +     /* ... and non-recursive readlock */
>>> +     lock_acquire(&dma_fence_lockdep_map, 0, 0, 1, 1, NULL, _RET_IP_);
>>
>> Would it work if signalling path would mark itself as a write lock? I am
>> thinking it would be nice to see in lockdep splats what are signals and
>> what are waits.
> 
> Yeah it'd be nice to have a read vs write name for the lock. But we
> already have this problem for e.g. flush_work(), from which I've
> stolen this idea. So it's not really new. Essentially look at the
> backtraces lockdep gives you, and reconstruct the deadlock. I'm hoping
> that people will notice the special functions on the backtrace, e.g.
> dma_fence_begin_signalling will be listed as offending function/lock
> holder, and then read the kerneldoc.
> 
>> The recursive usage wouldn't work then right? Would write annotation on
>> the wait path work?
> 
> Wait path is write annotations already, but yeah annotating the
> signalling side as write would cause endless amounts of alse
> positives. Also it makes composability of these e.g. what I've done in
> amdgpu with annotations in tdr work in drm/scheduler, annotations in
> the amdgpu gpu reset code and then also annotations in atomic code,
> which all nest within each other in some call chains, but not others.
> Dropping the recursion would break that and make it really awkward to
> annotate such cases correctly.
> 
> And the recursion only works if it's read locks, otherwise lockdep
> complains if you have inconsistent annotations on the signalling side
> (which again would make it more or less impossible to annotate the
> above case fully).

How do I see in lockdep splats if it was a read or write user? Your patch appears to have:

dma_fence_signal:
+	lock_acquire(&dma_fence_lockdep_map, 0, 0, 1, 1, NULL, _RET_IP_);

__dma_fence_might_wait:
+	lock_acquire(&dma_fence_lockdep_map, 0, 0, 1, 1, NULL, _THIS_IP_);

Which both seem like read lock. I don't fully understand the lockdep API so I might be wrong, not sure. But neither I see a difference in splats telling me which path is which.

Regards,

Tvrtko
Daniel Vetter June 11, 2020, 11:29 a.m. UTC | #11
On Thu, Jun 11, 2020 at 12:36 PM Tvrtko Ursulin
<tvrtko.ursulin@linux.intel.com> wrote:
>
>
> On 10/06/2020 16:17, Daniel Vetter wrote:
> > On Wed, Jun 10, 2020 at 4:22 PM Tvrtko Ursulin
> > <tvrtko.ursulin@linux.intel.com> wrote:
> >>
> >>
> >> On 04/06/2020 09:12, Daniel Vetter wrote:
> >>> Design is similar to the lockdep annotations for workers, but with
> >>> some twists:
> >>>
> >>> - We use a read-lock for the execution/worker/completion side, so that
> >>>     this explicit annotation can be more liberally sprinkled around.
> >>>     With read locks lockdep isn't going to complain if the read-side
> >>>     isn't nested the same way under all circumstances, so ABBA deadlocks
> >>>     are ok. Which they are, since this is an annotation only.
> >>>
> >>> - We're using non-recursive lockdep read lock mode, since in recursive
> >>>     read lock mode lockdep does not catch read side hazards. And we
> >>>     _very_ much want read side hazards to be caught. For full details of
> >>>     this limitation see
> >>>
> >>>     commit e91498589746065e3ae95d9a00b068e525eec34f
> >>>     Author: Peter Zijlstra <peterz@infradead.org>
> >>>     Date:   Wed Aug 23 13:13:11 2017 +0200
> >>>
> >>>         locking/lockdep/selftests: Add mixed read-write ABBA tests
> >>>
> >>> - To allow nesting of the read-side explicit annotations we explicitly
> >>>     keep track of the nesting. lock_is_held() allows us to do that.
> >>>
> >>> - The wait-side annotation is a write lock, and entirely done within
> >>>     dma_fence_wait() for everyone by default.
> >>>
> >>> - To be able to freely annotate helper functions I want to make it ok
> >>>     to call dma_fence_begin/end_signalling from soft/hardirq context.
> >>>     First attempt was using the hardirq locking context for the write
> >>>     side in lockdep, but this forces all normal spinlocks nested within
> >>>     dma_fence_begin/end_signalling to be spinlocks. That bollocks.
> >>>
> >>>     The approach now is to simple check in_atomic(), and for these cases
> >>>     entirely rely on the might_sleep() check in dma_fence_wait(). That
> >>>     will catch any wrong nesting against spinlocks from soft/hardirq
> >>>     contexts.
> >>>
> >>> The idea here is that every code path that's critical for eventually
> >>> signalling a dma_fence should be annotated with
> >>> dma_fence_begin/end_signalling. The annotation ideally starts right
> >>> after a dma_fence is published (added to a dma_resv, exposed as a
> >>> sync_file fd, attached to a drm_syncobj fd, or anything else that
> >>> makes the dma_fence visible to other kernel threads), up to and
> >>> including the dma_fence_wait(). Examples are irq handlers, the
> >>> scheduler rt threads, the tail of execbuf (after the corresponding
> >>> fences are visible), any workers that end up signalling dma_fences and
> >>> really anything else. Not annotated should be code paths that only
> >>> complete fences opportunistically as the gpu progresses, like e.g.
> >>> shrinker/eviction code.
> >>>
> >>> The main class of deadlocks this is supposed to catch are:
> >>>
> >>> Thread A:
> >>>
> >>>        mutex_lock(A);
> >>>        mutex_unlock(A);
> >>>
> >>>        dma_fence_signal();
> >>>
> >>> Thread B:
> >>>
> >>>        mutex_lock(A);
> >>>        dma_fence_wait();
> >>>        mutex_unlock(A);
> >>>
> >>> Thread B is blocked on A signalling the fence, but A never gets around
> >>> to that because it cannot acquire the lock A.
> >>>
> >>> Note that dma_fence_wait() is allowed to be nested within
> >>> dma_fence_begin/end_signalling sections. To allow this to happen the
> >>> read lock needs to be upgraded to a write lock, which means that any
> >>> other lock is acquired between the dma_fence_begin_signalling() call and
> >>> the call to dma_fence_wait(), and still held, this will result in an
> >>> immediate lockdep complaint. The only other option would be to not
> >>> annotate such calls, defeating the point. Therefore these annotations
> >>> cannot be sprinkled over the code entirely mindless to avoid false
> >>> positives.
> >>>
> >>> v2: handle soft/hardirq ctx better against write side and dont forget
> >>> EXPORT_SYMBOL, drivers can't use this otherwise.
> >>>
> >>> v3: Kerneldoc.
> >>>
> >>> v4: Some spelling fixes from Mika
> >>>
> >>> Cc: Mika Kuoppala <mika.kuoppala@intel.com>
> >>> Cc: Thomas Hellstrom <thomas.hellstrom@intel.com>
> >>> Cc: linux-media@vger.kernel.org
> >>> Cc: linaro-mm-sig@lists.linaro.org
> >>> Cc: linux-rdma@vger.kernel.org
> >>> Cc: amd-gfx@lists.freedesktop.org
> >>> Cc: intel-gfx@lists.freedesktop.org
> >>> Cc: Chris Wilson <chris@chris-wilson.co.uk>
> >>> Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
> >>> Cc: Christian König <christian.koenig@amd.com>
> >>> Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>
> >>> ---
> >>>    Documentation/driver-api/dma-buf.rst |  12 +-
> >>>    drivers/dma-buf/dma-fence.c          | 161 +++++++++++++++++++++++++++
> >>>    include/linux/dma-fence.h            |  12 ++
> >>>    3 files changed, 182 insertions(+), 3 deletions(-)
> >>>
> >>> diff --git a/Documentation/driver-api/dma-buf.rst b/Documentation/driver-api/dma-buf.rst
> >>> index 63dec76d1d8d..05d856131140 100644
> >>> --- a/Documentation/driver-api/dma-buf.rst
> >>> +++ b/Documentation/driver-api/dma-buf.rst
> >>> @@ -100,11 +100,11 @@ CPU Access to DMA Buffer Objects
> >>>    .. kernel-doc:: drivers/dma-buf/dma-buf.c
> >>>       :doc: cpu access
> >>>
> >>> -Fence Poll Support
> >>> -~~~~~~~~~~~~~~~~~~
> >>> +Implicit Fence Poll Support
> >>> +~~~~~~~~~~~~~~~~~~~~~~~~~~~
> >>>
> >>>    .. kernel-doc:: drivers/dma-buf/dma-buf.c
> >>> -   :doc: fence polling
> >>> +   :doc: implicit fence polling
> >>>
> >>>    Kernel Functions and Structures Reference
> >>>    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> >>> @@ -133,6 +133,12 @@ DMA Fences
> >>>    .. kernel-doc:: drivers/dma-buf/dma-fence.c
> >>>       :doc: DMA fences overview
> >>>
> >>> +DMA Fence Signalling Annotations
> >>> +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> >>> +
> >>> +.. kernel-doc:: drivers/dma-buf/dma-fence.c
> >>> +   :doc: fence signalling annotation
> >>> +
> >>>    DMA Fences Functions Reference
> >>>    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> >>>
> >>> diff --git a/drivers/dma-buf/dma-fence.c b/drivers/dma-buf/dma-fence.c
> >>> index 656e9ac2d028..0005bc002529 100644
> >>> --- a/drivers/dma-buf/dma-fence.c
> >>> +++ b/drivers/dma-buf/dma-fence.c
> >>> @@ -110,6 +110,160 @@ u64 dma_fence_context_alloc(unsigned num)
> >>>    }
> >>>    EXPORT_SYMBOL(dma_fence_context_alloc);
> >>>
> >>> +/**
> >>> + * DOC: fence signalling annotation
> >>> + *
> >>> + * Proving correctness of all the kernel code around &dma_fence through code
> >>> + * review and testing is tricky for a few reasons:
> >>> + *
> >>> + * * It is a cross-driver contract, and therefore all drivers must follow the
> >>> + *   same rules for lock nesting order, calling contexts for various functions
> >>> + *   and anything else significant for in-kernel interfaces. But it is also
> >>> + *   impossible to test all drivers in a single machine, hence brute-force N vs.
> >>> + *   N testing of all combinations is impossible. Even just limiting to the
> >>> + *   possible combinations is infeasible.
> >>> + *
> >>> + * * There is an enormous amount of driver code involved. For render drivers
> >>> + *   there's the tail of command submission, after fences are published,
> >>> + *   scheduler code, interrupt and workers to process job completion,
> >>> + *   and timeout, gpu reset and gpu hang recovery code. Plus for integration
> >>> + *   with core mm with have &mmu_notifier, respectively &mmu_interval_notifier,
> >>> + *   and &shrinker. For modesetting drivers there's the commit tail functions
> >>> + *   between when fences for an atomic modeset are published, and when the
> >>> + *   corresponding vblank completes, including any interrupt processing and
> >>> + *   related workers. Auditing all that code, across all drivers, is not
> >>> + *   feasible.
> >>> + *
> >>> + * * Due to how many other subsystems are involved and the locking hierarchies
> >>> + *   this pulls in there is extremely thin wiggle-room for driver-specific
> >>> + *   differences. &dma_fence interacts with almost all of the core memory
> >>> + *   handling through page fault handlers via &dma_resv, dma_resv_lock() and
> >>> + *   dma_resv_unlock(). On the other side it also interacts through all
> >>> + *   allocation sites through &mmu_notifier and &shrinker.
> >>> + *
> >>> + * Furthermore lockdep does not handle cross-release dependencies, which means
> >>> + * any deadlocks between dma_fence_wait() and dma_fence_signal() can't be caught
> >>> + * at runtime with some quick testing. The simplest example is one thread
> >>> + * waiting on a &dma_fence while holding a lock::
> >>> + *
> >>> + *     lock(A);
> >>> + *     dma_fence_wait(B);
> >>> + *     unlock(A);
> >>> + *
> >>> + * while the other thread is stuck trying to acquire the same lock, which
> >>> + * prevents it from signalling the fence the previous thread is stuck waiting
> >>> + * on::
> >>> + *
> >>> + *     lock(A);
> >>> + *     unlock(A);
> >>> + *     dma_fence_signal(B);
> >>> + *
> >>> + * By manually annotating all code relevant to signalling a &dma_fence we can
> >>> + * teach lockdep about these dependencies, which also helps with the validation
> >>> + * headache since now lockdep can check all the rules for us::
> >>> + *
> >>> + *    cookie = dma_fence_begin_signalling();
> >>> + *    lock(A);
> >>> + *    unlock(A);
> >>> + *    dma_fence_signal(B);
> >>> + *    dma_fence_end_signalling(cookie);
> >>> + *
> >>> + * For using dma_fence_begin_signalling() and dma_fence_end_signalling() to
> >>> + * annotate critical sections the following rules need to be observed:
> >>> + *
> >>> + * * All code necessary to complete a &dma_fence must be annotated, from the
> >>> + *   point where a fence is accessible to other threads, to the point where
> >>> + *   dma_fence_signal() is called. Un-annotated code can contain deadlock issues,
> >>> + *   and due to the very strict rules and many corner cases it is infeasible to
> >>> + *   catch these just with review or normal stress testing.
> >>> + *
> >>> + * * &struct dma_resv deserves a special note, since the readers are only
> >>> + *   protected by rcu. This means the signalling critical section starts as soon
> >>> + *   as the new fences are installed, even before dma_resv_unlock() is called.
> >>> + *
> >>> + * * The only exception are fast paths and opportunistic signalling code, which
> >>> + *   calls dma_fence_signal() purely as an optimization, but is not required to
> >>> + *   guarantee completion of a &dma_fence. The usual example is a wait IOCTL
> >>> + *   which calls dma_fence_signal(), while the mandatory completion path goes
> >>> + *   through a hardware interrupt and possible job completion worker.
> >>> + *
> >>> + * * To aid composability of code, the annotations can be freely nested, as long
> >>> + *   as the overall locking hierarchy is consistent. The annotations also work
> >>> + *   both in interrupt and process context. Due to implementation details this
> >>> + *   requires that callers pass an opaque cookie from
> >>> + *   dma_fence_begin_signalling() to dma_fence_end_signalling().
> >>> + *
> >>> + * * Validation against the cross driver contract is implemented by priming
> >>> + *   lockdep with the relevant hierarchy at boot-up. This means even just
> >>> + *   testing with a single device is enough to validate a driver, at least as
> >>> + *   far as deadlocks with dma_fence_wait() against dma_fence_signal() are
> >>> + *   concerned.
> >>> + */
> >>> +#ifdef CONFIG_LOCKDEP
> >>> +struct lockdep_map   dma_fence_lockdep_map = {
> >>> +     .name = "dma_fence_map"
> >>> +};
> >>
> >> Maybe a stupid question because this is definitely complicated, but.. If
> >> you have a single/static/global lockdep map, doesn't this mean _all_
> >> locks, from _all_ drivers happening to use dma-fences will get recorded
> >> in it. Will this work and not cause false positives?
> >>
> >> Sounds like it could create a common link between two completely
> >> unconnected usages. Because below you do add annotations to generic
> >> dma_fence_signal and dma_fence_wait.
> >
> > This is fully intentional. dma-fence is a cross-driver interface, if
> > every driver invents its own rules about how this should work we have
> > an unmaintainable and unreviewable mess.
> >
> > I've typed up the full length rant already here:
> >
> > https://lore.kernel.org/dri-devel/CAKMK7uGnFhbpuurRsnZ4dvRV9gQ_3-rmSJaoqSFY=+Kvepz_CA@mail.gmail.com/
>
> But "perfect storm" of:
>
>  + global fence lockmap
>  + mmu notifiers
>  + fs reclaim
>  + default annotations in dma_fence_signal / dma_fence_wait
>
> Equals to anything ever using dma_fence will be in impossible chains with random other drivers, even if neither driver has code to export/share that fence.
>
> Example from the CI run:
>
>  [25.918788] Chain exists of:
>   fs_reclaim --> mmu_notifier_invalidate_range_start --> dma_fence_map
>  [25.918794]  Possible unsafe locking scenario:
>  [25.918797]        CPU0                    CPU1
>  [25.918799]        ----                    ----
>  [25.918801]   lock(dma_fence_map);
>  [25.918803]                                lock(mmu_notifier_invalidate_range_start);
>  [25.918807]                                lock(dma_fence_map);
>  [25.918809]   lock(fs_reclaim);
>
> What about a dma_fence_export helper which would "arm" the annotations? It would be called as soon as the fence is exported. Maybe when added to dma_resv, or exported via sync_file, etc. Before that point begin/end_signaling and so would be no-ops.

Run CI without the i915 annotation patch, nothing breaks.

So we can gradually fix up existing code that doesn't quite get it
right and move on.

> >>> +
> >>> +/**
> >>> + * dma_fence_begin_signalling - begin a critical DMA fence signalling section
> >>> + *
> >>> + * Drivers should use this to annotate the beginning of any code section
> >>> + * required to eventually complete &dma_fence by calling dma_fence_signal().
> >>> + *
> >>> + * The end of these critical sections are annotated with
> >>> + * dma_fence_end_signalling().
> >>> + *
> >>> + * Returns:
> >>> + *
> >>> + * Opaque cookie needed by the implementation, which needs to be passed to
> >>> + * dma_fence_end_signalling().
> >>> + */
> >>> +bool dma_fence_begin_signalling(void)
> >>> +{
> >>> +     /* explicitly nesting ... */
> >>> +     if (lock_is_held_type(&dma_fence_lockdep_map, 1))
> >>> +             return true;
> >>> +
> >>> +     /* rely on might_sleep check for soft/hardirq locks */
> >>> +     if (in_atomic())
> >>> +             return true;
> >>> +
> >>> +     /* ... and non-recursive readlock */
> >>> +     lock_acquire(&dma_fence_lockdep_map, 0, 0, 1, 1, NULL, _RET_IP_);
> >>
> >> Would it work if signalling path would mark itself as a write lock? I am
> >> thinking it would be nice to see in lockdep splats what are signals and
> >> what are waits.
> >
> > Yeah it'd be nice to have a read vs write name for the lock. But we
> > already have this problem for e.g. flush_work(), from which I've
> > stolen this idea. So it's not really new. Essentially look at the
> > backtraces lockdep gives you, and reconstruct the deadlock. I'm hoping
> > that people will notice the special functions on the backtrace, e.g.
> > dma_fence_begin_signalling will be listed as offending function/lock
> > holder, and then read the kerneldoc.
> >
> >> The recursive usage wouldn't work then right? Would write annotation on
> >> the wait path work?
> >
> > Wait path is write annotations already, but yeah annotating the
> > signalling side as write would cause endless amounts of alse
> > positives. Also it makes composability of these e.g. what I've done in
> > amdgpu with annotations in tdr work in drm/scheduler, annotations in
> > the amdgpu gpu reset code and then also annotations in atomic code,
> > which all nest within each other in some call chains, but not others.
> > Dropping the recursion would break that and make it really awkward to
> > annotate such cases correctly.
> >
> > And the recursion only works if it's read locks, otherwise lockdep
> > complains if you have inconsistent annotations on the signalling side
> > (which again would make it more or less impossible to annotate the
> > above case fully).
>
> How do I see in lockdep splats if it was a read or write user? Your patch appears to have:
>
> dma_fence_signal:
> +       lock_acquire(&dma_fence_lockdep_map, 0, 0, 1, 1, NULL, _RET_IP_);
>
> __dma_fence_might_wait:
> +       lock_acquire(&dma_fence_lockdep_map, 0, 0, 1, 1, NULL, _THIS_IP_);
>
> Which both seem like read lock. I don't fully understand the lockdep API so I might be wrong, not sure. But neither I see a difference in splats telling me which path is which.

I think you got tricked by the implementation, this isn't quite what's
going on. There's two things which make the annotations special:

- we want a recursive read lock on the signalling critical section.
The problem is that lockdep doesn't implement full validation for
recursive read locks, only non-recursive read/write locks fully
validated. There's some checks for recursive read locks, but exactly
the checks we need to catch common dma_fence_wait deadlocks aren't
done. That's why we need to implement manual lock recursion on the
reader side

- now on the write side we additionally need to implement an
read2write upgrade, and a write2read downgrade. Lockdep doesn't
implement that, so again we have to hand-roll this.

Let's go through the code line-by-line:

    bool tmp;

    tmp = lock_is_held_type(&dma_fence_lockdep_map, 1);

We check whether someone is holding the non-recursive read lock already.

    if (tmp)
        lock_release(&dma_fence_lockdep_map, _THIS_IP_);

If that's the case, we drop that read lock.

    lock_map_acquire(&dma_fence_lockdep_map);

Then we do the actual might_wait annotation, the above takes the full
write lock ...

    lock_map_release(&dma_fence_lockdep_map);

... and now we release the write lock again.


    if (tmp)
        lock_acquire(&dma_fence_lockdep_map, 0, 0, 1, 1, NULL, _THIS_IP_);

Finally we need to re-acquire the read lock, if we've held that when
entering this function. This annotation naturally has to exactly match
what begin_signalling would do, otherwise the hand-rolled nesting
would fall apart.

I hope that explains what's going on here, and assures you that
might_wait() is indeed a write lock annotation, but with a big pile of
complications.
-Daniel
Tvrtko Ursulin June 11, 2020, 2:29 p.m. UTC | #12
On 11/06/2020 12:29, Daniel Vetter wrote:
> On Thu, Jun 11, 2020 at 12:36 PM Tvrtko Ursulin
> <tvrtko.ursulin@linux.intel.com> wrote:
>> On 10/06/2020 16:17, Daniel Vetter wrote:
>>> On Wed, Jun 10, 2020 at 4:22 PM Tvrtko Ursulin
>>> <tvrtko.ursulin@linux.intel.com> wrote:
>>>>
>>>>
>>>> On 04/06/2020 09:12, Daniel Vetter wrote:
>>>>> Design is similar to the lockdep annotations for workers, but with
>>>>> some twists:
>>>>>
>>>>> - We use a read-lock for the execution/worker/completion side, so that
>>>>>      this explicit annotation can be more liberally sprinkled around.
>>>>>      With read locks lockdep isn't going to complain if the read-side
>>>>>      isn't nested the same way under all circumstances, so ABBA deadlocks
>>>>>      are ok. Which they are, since this is an annotation only.
>>>>>
>>>>> - We're using non-recursive lockdep read lock mode, since in recursive
>>>>>      read lock mode lockdep does not catch read side hazards. And we
>>>>>      _very_ much want read side hazards to be caught. For full details of
>>>>>      this limitation see
>>>>>
>>>>>      commit e91498589746065e3ae95d9a00b068e525eec34f
>>>>>      Author: Peter Zijlstra <peterz@infradead.org>
>>>>>      Date:   Wed Aug 23 13:13:11 2017 +0200
>>>>>
>>>>>          locking/lockdep/selftests: Add mixed read-write ABBA tests
>>>>>
>>>>> - To allow nesting of the read-side explicit annotations we explicitly
>>>>>      keep track of the nesting. lock_is_held() allows us to do that.
>>>>>
>>>>> - The wait-side annotation is a write lock, and entirely done within
>>>>>      dma_fence_wait() for everyone by default.
>>>>>
>>>>> - To be able to freely annotate helper functions I want to make it ok
>>>>>      to call dma_fence_begin/end_signalling from soft/hardirq context.
>>>>>      First attempt was using the hardirq locking context for the write
>>>>>      side in lockdep, but this forces all normal spinlocks nested within
>>>>>      dma_fence_begin/end_signalling to be spinlocks. That bollocks.
>>>>>
>>>>>      The approach now is to simple check in_atomic(), and for these cases
>>>>>      entirely rely on the might_sleep() check in dma_fence_wait(). That
>>>>>      will catch any wrong nesting against spinlocks from soft/hardirq
>>>>>      contexts.
>>>>>
>>>>> The idea here is that every code path that's critical for eventually
>>>>> signalling a dma_fence should be annotated with
>>>>> dma_fence_begin/end_signalling. The annotation ideally starts right
>>>>> after a dma_fence is published (added to a dma_resv, exposed as a
>>>>> sync_file fd, attached to a drm_syncobj fd, or anything else that
>>>>> makes the dma_fence visible to other kernel threads), up to and
>>>>> including the dma_fence_wait(). Examples are irq handlers, the
>>>>> scheduler rt threads, the tail of execbuf (after the corresponding
>>>>> fences are visible), any workers that end up signalling dma_fences and
>>>>> really anything else. Not annotated should be code paths that only
>>>>> complete fences opportunistically as the gpu progresses, like e.g.
>>>>> shrinker/eviction code.
>>>>>
>>>>> The main class of deadlocks this is supposed to catch are:
>>>>>
>>>>> Thread A:
>>>>>
>>>>>         mutex_lock(A);
>>>>>         mutex_unlock(A);
>>>>>
>>>>>         dma_fence_signal();
>>>>>
>>>>> Thread B:
>>>>>
>>>>>         mutex_lock(A);
>>>>>         dma_fence_wait();
>>>>>         mutex_unlock(A);
>>>>>
>>>>> Thread B is blocked on A signalling the fence, but A never gets around
>>>>> to that because it cannot acquire the lock A.
>>>>>
>>>>> Note that dma_fence_wait() is allowed to be nested within
>>>>> dma_fence_begin/end_signalling sections. To allow this to happen the
>>>>> read lock needs to be upgraded to a write lock, which means that any
>>>>> other lock is acquired between the dma_fence_begin_signalling() call and
>>>>> the call to dma_fence_wait(), and still held, this will result in an
>>>>> immediate lockdep complaint. The only other option would be to not
>>>>> annotate such calls, defeating the point. Therefore these annotations
>>>>> cannot be sprinkled over the code entirely mindless to avoid false
>>>>> positives.
>>>>>
>>>>> v2: handle soft/hardirq ctx better against write side and dont forget
>>>>> EXPORT_SYMBOL, drivers can't use this otherwise.
>>>>>
>>>>> v3: Kerneldoc.
>>>>>
>>>>> v4: Some spelling fixes from Mika
>>>>>
>>>>> Cc: Mika Kuoppala <mika.kuoppala@intel.com>
>>>>> Cc: Thomas Hellstrom <thomas.hellstrom@intel.com>
>>>>> Cc: linux-media@vger.kernel.org
>>>>> Cc: linaro-mm-sig@lists.linaro.org
>>>>> Cc: linux-rdma@vger.kernel.org
>>>>> Cc: amd-gfx@lists.freedesktop.org
>>>>> Cc: intel-gfx@lists.freedesktop.org
>>>>> Cc: Chris Wilson <chris@chris-wilson.co.uk>
>>>>> Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
>>>>> Cc: Christian König <christian.koenig@amd.com>
>>>>> Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>
>>>>> ---
>>>>>     Documentation/driver-api/dma-buf.rst |  12 +-
>>>>>     drivers/dma-buf/dma-fence.c          | 161 +++++++++++++++++++++++++++
>>>>>     include/linux/dma-fence.h            |  12 ++
>>>>>     3 files changed, 182 insertions(+), 3 deletions(-)
>>>>>
>>>>> diff --git a/Documentation/driver-api/dma-buf.rst b/Documentation/driver-api/dma-buf.rst
>>>>> index 63dec76d1d8d..05d856131140 100644
>>>>> --- a/Documentation/driver-api/dma-buf.rst
>>>>> +++ b/Documentation/driver-api/dma-buf.rst
>>>>> @@ -100,11 +100,11 @@ CPU Access to DMA Buffer Objects
>>>>>     .. kernel-doc:: drivers/dma-buf/dma-buf.c
>>>>>        :doc: cpu access
>>>>>
>>>>> -Fence Poll Support
>>>>> -~~~~~~~~~~~~~~~~~~
>>>>> +Implicit Fence Poll Support
>>>>> +~~~~~~~~~~~~~~~~~~~~~~~~~~~
>>>>>
>>>>>     .. kernel-doc:: drivers/dma-buf/dma-buf.c
>>>>> -   :doc: fence polling
>>>>> +   :doc: implicit fence polling
>>>>>
>>>>>     Kernel Functions and Structures Reference
>>>>>     ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>>>>> @@ -133,6 +133,12 @@ DMA Fences
>>>>>     .. kernel-doc:: drivers/dma-buf/dma-fence.c
>>>>>        :doc: DMA fences overview
>>>>>
>>>>> +DMA Fence Signalling Annotations
>>>>> +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>>>>> +
>>>>> +.. kernel-doc:: drivers/dma-buf/dma-fence.c
>>>>> +   :doc: fence signalling annotation
>>>>> +
>>>>>     DMA Fences Functions Reference
>>>>>     ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>>>>>
>>>>> diff --git a/drivers/dma-buf/dma-fence.c b/drivers/dma-buf/dma-fence.c
>>>>> index 656e9ac2d028..0005bc002529 100644
>>>>> --- a/drivers/dma-buf/dma-fence.c
>>>>> +++ b/drivers/dma-buf/dma-fence.c
>>>>> @@ -110,6 +110,160 @@ u64 dma_fence_context_alloc(unsigned num)
>>>>>     }
>>>>>     EXPORT_SYMBOL(dma_fence_context_alloc);
>>>>>
>>>>> +/**
>>>>> + * DOC: fence signalling annotation
>>>>> + *
>>>>> + * Proving correctness of all the kernel code around &dma_fence through code
>>>>> + * review and testing is tricky for a few reasons:
>>>>> + *
>>>>> + * * It is a cross-driver contract, and therefore all drivers must follow the
>>>>> + *   same rules for lock nesting order, calling contexts for various functions
>>>>> + *   and anything else significant for in-kernel interfaces. But it is also
>>>>> + *   impossible to test all drivers in a single machine, hence brute-force N vs.
>>>>> + *   N testing of all combinations is impossible. Even just limiting to the
>>>>> + *   possible combinations is infeasible.
>>>>> + *
>>>>> + * * There is an enormous amount of driver code involved. For render drivers
>>>>> + *   there's the tail of command submission, after fences are published,
>>>>> + *   scheduler code, interrupt and workers to process job completion,
>>>>> + *   and timeout, gpu reset and gpu hang recovery code. Plus for integration
>>>>> + *   with core mm with have &mmu_notifier, respectively &mmu_interval_notifier,
>>>>> + *   and &shrinker. For modesetting drivers there's the commit tail functions
>>>>> + *   between when fences for an atomic modeset are published, and when the
>>>>> + *   corresponding vblank completes, including any interrupt processing and
>>>>> + *   related workers. Auditing all that code, across all drivers, is not
>>>>> + *   feasible.
>>>>> + *
>>>>> + * * Due to how many other subsystems are involved and the locking hierarchies
>>>>> + *   this pulls in there is extremely thin wiggle-room for driver-specific
>>>>> + *   differences. &dma_fence interacts with almost all of the core memory
>>>>> + *   handling through page fault handlers via &dma_resv, dma_resv_lock() and
>>>>> + *   dma_resv_unlock(). On the other side it also interacts through all
>>>>> + *   allocation sites through &mmu_notifier and &shrinker.
>>>>> + *
>>>>> + * Furthermore lockdep does not handle cross-release dependencies, which means
>>>>> + * any deadlocks between dma_fence_wait() and dma_fence_signal() can't be caught
>>>>> + * at runtime with some quick testing. The simplest example is one thread
>>>>> + * waiting on a &dma_fence while holding a lock::
>>>>> + *
>>>>> + *     lock(A);
>>>>> + *     dma_fence_wait(B);
>>>>> + *     unlock(A);
>>>>> + *
>>>>> + * while the other thread is stuck trying to acquire the same lock, which
>>>>> + * prevents it from signalling the fence the previous thread is stuck waiting
>>>>> + * on::
>>>>> + *
>>>>> + *     lock(A);
>>>>> + *     unlock(A);
>>>>> + *     dma_fence_signal(B);
>>>>> + *
>>>>> + * By manually annotating all code relevant to signalling a &dma_fence we can
>>>>> + * teach lockdep about these dependencies, which also helps with the validation
>>>>> + * headache since now lockdep can check all the rules for us::
>>>>> + *
>>>>> + *    cookie = dma_fence_begin_signalling();
>>>>> + *    lock(A);
>>>>> + *    unlock(A);
>>>>> + *    dma_fence_signal(B);
>>>>> + *    dma_fence_end_signalling(cookie);
>>>>> + *
>>>>> + * For using dma_fence_begin_signalling() and dma_fence_end_signalling() to
>>>>> + * annotate critical sections the following rules need to be observed:
>>>>> + *
>>>>> + * * All code necessary to complete a &dma_fence must be annotated, from the
>>>>> + *   point where a fence is accessible to other threads, to the point where
>>>>> + *   dma_fence_signal() is called. Un-annotated code can contain deadlock issues,
>>>>> + *   and due to the very strict rules and many corner cases it is infeasible to
>>>>> + *   catch these just with review or normal stress testing.
>>>>> + *
>>>>> + * * &struct dma_resv deserves a special note, since the readers are only
>>>>> + *   protected by rcu. This means the signalling critical section starts as soon
>>>>> + *   as the new fences are installed, even before dma_resv_unlock() is called.
>>>>> + *
>>>>> + * * The only exception are fast paths and opportunistic signalling code, which
>>>>> + *   calls dma_fence_signal() purely as an optimization, but is not required to
>>>>> + *   guarantee completion of a &dma_fence. The usual example is a wait IOCTL
>>>>> + *   which calls dma_fence_signal(), while the mandatory completion path goes
>>>>> + *   through a hardware interrupt and possible job completion worker.
>>>>> + *
>>>>> + * * To aid composability of code, the annotations can be freely nested, as long
>>>>> + *   as the overall locking hierarchy is consistent. The annotations also work
>>>>> + *   both in interrupt and process context. Due to implementation details this
>>>>> + *   requires that callers pass an opaque cookie from
>>>>> + *   dma_fence_begin_signalling() to dma_fence_end_signalling().
>>>>> + *
>>>>> + * * Validation against the cross driver contract is implemented by priming
>>>>> + *   lockdep with the relevant hierarchy at boot-up. This means even just
>>>>> + *   testing with a single device is enough to validate a driver, at least as
>>>>> + *   far as deadlocks with dma_fence_wait() against dma_fence_signal() are
>>>>> + *   concerned.
>>>>> + */
>>>>> +#ifdef CONFIG_LOCKDEP
>>>>> +struct lockdep_map   dma_fence_lockdep_map = {
>>>>> +     .name = "dma_fence_map"
>>>>> +};
>>>>
>>>> Maybe a stupid question because this is definitely complicated, but.. If
>>>> you have a single/static/global lockdep map, doesn't this mean _all_
>>>> locks, from _all_ drivers happening to use dma-fences will get recorded
>>>> in it. Will this work and not cause false positives?
>>>>
>>>> Sounds like it could create a common link between two completely
>>>> unconnected usages. Because below you do add annotations to generic
>>>> dma_fence_signal and dma_fence_wait.
>>>
>>> This is fully intentional. dma-fence is a cross-driver interface, if
>>> every driver invents its own rules about how this should work we have
>>> an unmaintainable and unreviewable mess.
>>>
>>> I've typed up the full length rant already here:
>>>
>>> https://lore.kernel.org/dri-devel/CAKMK7uGnFhbpuurRsnZ4dvRV9gQ_3-rmSJaoqSFY=+Kvepz_CA@mail.gmail.com/
>>
>> But "perfect storm" of:
>>
>>   + global fence lockmap
>>   + mmu notifiers
>>   + fs reclaim
>>   + default annotations in dma_fence_signal / dma_fence_wait
>>
>> Equals to anything ever using dma_fence will be in impossible chains with random other drivers, even if neither driver has code to export/share that fence.
>>
>> Example from the CI run:
>>
>>   [25.918788] Chain exists of:
>>    fs_reclaim --> mmu_notifier_invalidate_range_start --> dma_fence_map
>>   [25.918794]  Possible unsafe locking scenario:
>>   [25.918797]        CPU0                    CPU1
>>   [25.918799]        ----                    ----
>>   [25.918801]   lock(dma_fence_map);
>>   [25.918803]                                lock(mmu_notifier_invalidate_range_start);
>>   [25.918807]                                lock(dma_fence_map);
>>   [25.918809]   lock(fs_reclaim);
>>
>> What about a dma_fence_export helper which would "arm" the annotations? It would be called as soon as the fence is exported. Maybe when added to dma_resv, or exported via sync_file, etc. Before that point begin/end_signaling and so would be no-ops.
> 
> Run CI without the i915 annotation patch, nothing breaks.

I think some parts of i915 would still break with my idea to only apply annotations on exported fences. What do you dislike about that idea? I thought the point is to enforce rules for _exported_ fences.

How you have annotated dma_fence_work you can't say, maybe it is exported maybe it isn't. I think it is btw, so splats would still be there, but I am not sure it is conceptually correct.

At least my understanding is GFP_KERNEL allocations are only disallowed by the virtue of the global dma-fence contract. If you want to enforce they are never used for anything but exporting, then that would be a bit harsh, no?

Another example from the CI run:

 [26.585357]        CPU0                    CPU1
 [26.585359]        ----                    ----
 [26.585360]   lock(dma_fence_map);
 [26.585362]                                lock(mmu_notifier_invalidate_range_start);
 [26.585365]                                lock(dma_fence_map);
 [26.585367]   lock(i915_gem_object_internal/1);
 [26.585369] 
 *** DEADLOCK ***

Lets say someone submitted an execbuf using userptr as a batch and then unmapped it immediately. That would explain CPU1 getting into the mmu notifier and waiting on this batch to unbind the object.

Meanwhile CPU0 is the async command parser for this request trying to lock the shadow batch buffer. Because it uses the dma_fence_work this is between the begin/end signalling markers.

It can be the same dma-fence I think, since we install the async parser fence on the real batch dma-resv, but dma_fence_map is not a real lock, so what is actually preventing progress in this case?

CPU1 is waiting on a fence, but CPU0 can obtain the lock(i915_gem_object_internal/1), proceed to parse the batch, and exit the signalling section. At which point CPU1 is still blocked, waiting until the execbuf finishes and then mmu notifier can finish and invalidate the pages.

Maybe I am missing something but I don't see how this one is real.

> So we can gradually fix up existing code that doesn't quite get it
> right and move on.
>
>>>>> +
>>>>> +/**
>>>>> + * dma_fence_begin_signalling - begin a critical DMA fence signalling section
>>>>> + *
>>>>> + * Drivers should use this to annotate the beginning of any code section
>>>>> + * required to eventually complete &dma_fence by calling dma_fence_signal().
>>>>> + *
>>>>> + * The end of these critical sections are annotated with
>>>>> + * dma_fence_end_signalling().
>>>>> + *
>>>>> + * Returns:
>>>>> + *
>>>>> + * Opaque cookie needed by the implementation, which needs to be passed to
>>>>> + * dma_fence_end_signalling().
>>>>> + */
>>>>> +bool dma_fence_begin_signalling(void)
>>>>> +{
>>>>> +     /* explicitly nesting ... */
>>>>> +     if (lock_is_held_type(&dma_fence_lockdep_map, 1))
>>>>> +             return true;
>>>>> +
>>>>> +     /* rely on might_sleep check for soft/hardirq locks */
>>>>> +     if (in_atomic())
>>>>> +             return true;
>>>>> +
>>>>> +     /* ... and non-recursive readlock */
>>>>> +     lock_acquire(&dma_fence_lockdep_map, 0, 0, 1, 1, NULL, _RET_IP_);
>>>>
>>>> Would it work if signalling path would mark itself as a write lock? I am
>>>> thinking it would be nice to see in lockdep splats what are signals and
>>>> what are waits.
>>>
>>> Yeah it'd be nice to have a read vs write name for the lock. But we
>>> already have this problem for e.g. flush_work(), from which I've
>>> stolen this idea. So it's not really new. Essentially look at the
>>> backtraces lockdep gives you, and reconstruct the deadlock. I'm hoping
>>> that people will notice the special functions on the backtrace, e.g.
>>> dma_fence_begin_signalling will be listed as offending function/lock
>>> holder, and then read the kerneldoc.
>>>
>>>> The recursive usage wouldn't work then right? Would write annotation on
>>>> the wait path work?
>>>
>>> Wait path is write annotations already, but yeah annotating the
>>> signalling side as write would cause endless amounts of alse
>>> positives. Also it makes composability of these e.g. what I've done in
>>> amdgpu with annotations in tdr work in drm/scheduler, annotations in
>>> the amdgpu gpu reset code and then also annotations in atomic code,
>>> which all nest within each other in some call chains, but not others.
>>> Dropping the recursion would break that and make it really awkward to
>>> annotate such cases correctly.
>>>
>>> And the recursion only works if it's read locks, otherwise lockdep
>>> complains if you have inconsistent annotations on the signalling side
>>> (which again would make it more or less impossible to annotate the
>>> above case fully).
>>
>> How do I see in lockdep splats if it was a read or write user? Your patch appears to have:
>>
>> dma_fence_signal:
>> +       lock_acquire(&dma_fence_lockdep_map, 0, 0, 1, 1, NULL, _RET_IP_);
>>
>> __dma_fence_might_wait:
>> +       lock_acquire(&dma_fence_lockdep_map, 0, 0, 1, 1, NULL, _THIS_IP_);
>>
>> Which both seem like read lock. I don't fully understand the lockdep API so I might be wrong, not sure. But neither I see a difference in splats telling me which path is which.
> 
> I think you got tricked by the implementation, this isn't quite what's
> going on. There's two things which make the annotations special:
> 
> - we want a recursive read lock on the signalling critical section.
> The problem is that lockdep doesn't implement full validation for
> recursive read locks, only non-recursive read/write locks fully
> validated. There's some checks for recursive read locks, but exactly
> the checks we need to catch common dma_fence_wait deadlocks aren't
> done. That's why we need to implement manual lock recursion on the
> reader side
> 
> - now on the write side we additionally need to implement an
> read2write upgrade, and a write2read downgrade. Lockdep doesn't
> implement that, so again we have to hand-roll this.
> 
> Let's go through the code line-by-line:
> 
>      bool tmp;
> 
>      tmp = lock_is_held_type(&dma_fence_lockdep_map, 1);
> 
> We check whether someone is holding the non-recursive read lock already.
> 
>      if (tmp)
>          lock_release(&dma_fence_lockdep_map, _THIS_IP_);
> 
> If that's the case, we drop that read lock.
> 
>      lock_map_acquire(&dma_fence_lockdep_map);
> 
> Then we do the actual might_wait annotation, the above takes the full
> write lock ...
> 
>      lock_map_release(&dma_fence_lockdep_map);
> 
> ... and now we release the write lock again.
> 
> 
>      if (tmp)
>          lock_acquire(&dma_fence_lockdep_map, 0, 0, 1, 1, NULL, _THIS_IP_);
> 
> Finally we need to re-acquire the read lock, if we've held that when
> entering this function. This annotation naturally has to exactly match
> what begin_signalling would do, otherwise the hand-rolled nesting
> would fall apart.
> 
> I hope that explains what's going on here, and assures you that
> might_wait() is indeed a write lock annotation, but with a big pile of
> complications.

I am certainly confused by the difference between lock_map_acquire/release and lock_acquire/release. What is the difference between the two?

Regards,

Tvrtko
Daniel Vetter June 11, 2020, 3:03 p.m. UTC | #13
On Thu, Jun 11, 2020 at 4:29 PM Tvrtko Ursulin
<tvrtko.ursulin@linux.intel.com> wrote:
>
>
> On 11/06/2020 12:29, Daniel Vetter wrote:
> > On Thu, Jun 11, 2020 at 12:36 PM Tvrtko Ursulin
> > <tvrtko.ursulin@linux.intel.com> wrote:
> >> On 10/06/2020 16:17, Daniel Vetter wrote:
> >>> On Wed, Jun 10, 2020 at 4:22 PM Tvrtko Ursulin
> >>> <tvrtko.ursulin@linux.intel.com> wrote:
> >>>>
> >>>>
> >>>> On 04/06/2020 09:12, Daniel Vetter wrote:
> >>>>> Design is similar to the lockdep annotations for workers, but with
> >>>>> some twists:
> >>>>>
> >>>>> - We use a read-lock for the execution/worker/completion side, so that
> >>>>>      this explicit annotation can be more liberally sprinkled around.
> >>>>>      With read locks lockdep isn't going to complain if the read-side
> >>>>>      isn't nested the same way under all circumstances, so ABBA deadlocks
> >>>>>      are ok. Which they are, since this is an annotation only.
> >>>>>
> >>>>> - We're using non-recursive lockdep read lock mode, since in recursive
> >>>>>      read lock mode lockdep does not catch read side hazards. And we
> >>>>>      _very_ much want read side hazards to be caught. For full details of
> >>>>>      this limitation see
> >>>>>
> >>>>>      commit e91498589746065e3ae95d9a00b068e525eec34f
> >>>>>      Author: Peter Zijlstra <peterz@infradead.org>
> >>>>>      Date:   Wed Aug 23 13:13:11 2017 +0200
> >>>>>
> >>>>>          locking/lockdep/selftests: Add mixed read-write ABBA tests
> >>>>>
> >>>>> - To allow nesting of the read-side explicit annotations we explicitly
> >>>>>      keep track of the nesting. lock_is_held() allows us to do that.
> >>>>>
> >>>>> - The wait-side annotation is a write lock, and entirely done within
> >>>>>      dma_fence_wait() for everyone by default.
> >>>>>
> >>>>> - To be able to freely annotate helper functions I want to make it ok
> >>>>>      to call dma_fence_begin/end_signalling from soft/hardirq context.
> >>>>>      First attempt was using the hardirq locking context for the write
> >>>>>      side in lockdep, but this forces all normal spinlocks nested within
> >>>>>      dma_fence_begin/end_signalling to be spinlocks. That bollocks.
> >>>>>
> >>>>>      The approach now is to simple check in_atomic(), and for these cases
> >>>>>      entirely rely on the might_sleep() check in dma_fence_wait(). That
> >>>>>      will catch any wrong nesting against spinlocks from soft/hardirq
> >>>>>      contexts.
> >>>>>
> >>>>> The idea here is that every code path that's critical for eventually
> >>>>> signalling a dma_fence should be annotated with
> >>>>> dma_fence_begin/end_signalling. The annotation ideally starts right
> >>>>> after a dma_fence is published (added to a dma_resv, exposed as a
> >>>>> sync_file fd, attached to a drm_syncobj fd, or anything else that
> >>>>> makes the dma_fence visible to other kernel threads), up to and
> >>>>> including the dma_fence_wait(). Examples are irq handlers, the
> >>>>> scheduler rt threads, the tail of execbuf (after the corresponding
> >>>>> fences are visible), any workers that end up signalling dma_fences and
> >>>>> really anything else. Not annotated should be code paths that only
> >>>>> complete fences opportunistically as the gpu progresses, like e.g.
> >>>>> shrinker/eviction code.
> >>>>>
> >>>>> The main class of deadlocks this is supposed to catch are:
> >>>>>
> >>>>> Thread A:
> >>>>>
> >>>>>         mutex_lock(A);
> >>>>>         mutex_unlock(A);
> >>>>>
> >>>>>         dma_fence_signal();
> >>>>>
> >>>>> Thread B:
> >>>>>
> >>>>>         mutex_lock(A);
> >>>>>         dma_fence_wait();
> >>>>>         mutex_unlock(A);
> >>>>>
> >>>>> Thread B is blocked on A signalling the fence, but A never gets around
> >>>>> to that because it cannot acquire the lock A.
> >>>>>
> >>>>> Note that dma_fence_wait() is allowed to be nested within
> >>>>> dma_fence_begin/end_signalling sections. To allow this to happen the
> >>>>> read lock needs to be upgraded to a write lock, which means that any
> >>>>> other lock is acquired between the dma_fence_begin_signalling() call and
> >>>>> the call to dma_fence_wait(), and still held, this will result in an
> >>>>> immediate lockdep complaint. The only other option would be to not
> >>>>> annotate such calls, defeating the point. Therefore these annotations
> >>>>> cannot be sprinkled over the code entirely mindless to avoid false
> >>>>> positives.
> >>>>>
> >>>>> v2: handle soft/hardirq ctx better against write side and dont forget
> >>>>> EXPORT_SYMBOL, drivers can't use this otherwise.
> >>>>>
> >>>>> v3: Kerneldoc.
> >>>>>
> >>>>> v4: Some spelling fixes from Mika
> >>>>>
> >>>>> Cc: Mika Kuoppala <mika.kuoppala@intel.com>
> >>>>> Cc: Thomas Hellstrom <thomas.hellstrom@intel.com>
> >>>>> Cc: linux-media@vger.kernel.org
> >>>>> Cc: linaro-mm-sig@lists.linaro.org
> >>>>> Cc: linux-rdma@vger.kernel.org
> >>>>> Cc: amd-gfx@lists.freedesktop.org
> >>>>> Cc: intel-gfx@lists.freedesktop.org
> >>>>> Cc: Chris Wilson <chris@chris-wilson.co.uk>
> >>>>> Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
> >>>>> Cc: Christian König <christian.koenig@amd.com>
> >>>>> Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>
> >>>>> ---
> >>>>>     Documentation/driver-api/dma-buf.rst |  12 +-
> >>>>>     drivers/dma-buf/dma-fence.c          | 161 +++++++++++++++++++++++++++
> >>>>>     include/linux/dma-fence.h            |  12 ++
> >>>>>     3 files changed, 182 insertions(+), 3 deletions(-)
> >>>>>
> >>>>> diff --git a/Documentation/driver-api/dma-buf.rst b/Documentation/driver-api/dma-buf.rst
> >>>>> index 63dec76d1d8d..05d856131140 100644
> >>>>> --- a/Documentation/driver-api/dma-buf.rst
> >>>>> +++ b/Documentation/driver-api/dma-buf.rst
> >>>>> @@ -100,11 +100,11 @@ CPU Access to DMA Buffer Objects
> >>>>>     .. kernel-doc:: drivers/dma-buf/dma-buf.c
> >>>>>        :doc: cpu access
> >>>>>
> >>>>> -Fence Poll Support
> >>>>> -~~~~~~~~~~~~~~~~~~
> >>>>> +Implicit Fence Poll Support
> >>>>> +~~~~~~~~~~~~~~~~~~~~~~~~~~~
> >>>>>
> >>>>>     .. kernel-doc:: drivers/dma-buf/dma-buf.c
> >>>>> -   :doc: fence polling
> >>>>> +   :doc: implicit fence polling
> >>>>>
> >>>>>     Kernel Functions and Structures Reference
> >>>>>     ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> >>>>> @@ -133,6 +133,12 @@ DMA Fences
> >>>>>     .. kernel-doc:: drivers/dma-buf/dma-fence.c
> >>>>>        :doc: DMA fences overview
> >>>>>
> >>>>> +DMA Fence Signalling Annotations
> >>>>> +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> >>>>> +
> >>>>> +.. kernel-doc:: drivers/dma-buf/dma-fence.c
> >>>>> +   :doc: fence signalling annotation
> >>>>> +
> >>>>>     DMA Fences Functions Reference
> >>>>>     ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> >>>>>
> >>>>> diff --git a/drivers/dma-buf/dma-fence.c b/drivers/dma-buf/dma-fence.c
> >>>>> index 656e9ac2d028..0005bc002529 100644
> >>>>> --- a/drivers/dma-buf/dma-fence.c
> >>>>> +++ b/drivers/dma-buf/dma-fence.c
> >>>>> @@ -110,6 +110,160 @@ u64 dma_fence_context_alloc(unsigned num)
> >>>>>     }
> >>>>>     EXPORT_SYMBOL(dma_fence_context_alloc);
> >>>>>
> >>>>> +/**
> >>>>> + * DOC: fence signalling annotation
> >>>>> + *
> >>>>> + * Proving correctness of all the kernel code around &dma_fence through code
> >>>>> + * review and testing is tricky for a few reasons:
> >>>>> + *
> >>>>> + * * It is a cross-driver contract, and therefore all drivers must follow the
> >>>>> + *   same rules for lock nesting order, calling contexts for various functions
> >>>>> + *   and anything else significant for in-kernel interfaces. But it is also
> >>>>> + *   impossible to test all drivers in a single machine, hence brute-force N vs.
> >>>>> + *   N testing of all combinations is impossible. Even just limiting to the
> >>>>> + *   possible combinations is infeasible.
> >>>>> + *
> >>>>> + * * There is an enormous amount of driver code involved. For render drivers
> >>>>> + *   there's the tail of command submission, after fences are published,
> >>>>> + *   scheduler code, interrupt and workers to process job completion,
> >>>>> + *   and timeout, gpu reset and gpu hang recovery code. Plus for integration
> >>>>> + *   with core mm with have &mmu_notifier, respectively &mmu_interval_notifier,
> >>>>> + *   and &shrinker. For modesetting drivers there's the commit tail functions
> >>>>> + *   between when fences for an atomic modeset are published, and when the
> >>>>> + *   corresponding vblank completes, including any interrupt processing and
> >>>>> + *   related workers. Auditing all that code, across all drivers, is not
> >>>>> + *   feasible.
> >>>>> + *
> >>>>> + * * Due to how many other subsystems are involved and the locking hierarchies
> >>>>> + *   this pulls in there is extremely thin wiggle-room for driver-specific
> >>>>> + *   differences. &dma_fence interacts with almost all of the core memory
> >>>>> + *   handling through page fault handlers via &dma_resv, dma_resv_lock() and
> >>>>> + *   dma_resv_unlock(). On the other side it also interacts through all
> >>>>> + *   allocation sites through &mmu_notifier and &shrinker.
> >>>>> + *
> >>>>> + * Furthermore lockdep does not handle cross-release dependencies, which means
> >>>>> + * any deadlocks between dma_fence_wait() and dma_fence_signal() can't be caught
> >>>>> + * at runtime with some quick testing. The simplest example is one thread
> >>>>> + * waiting on a &dma_fence while holding a lock::
> >>>>> + *
> >>>>> + *     lock(A);
> >>>>> + *     dma_fence_wait(B);
> >>>>> + *     unlock(A);
> >>>>> + *
> >>>>> + * while the other thread is stuck trying to acquire the same lock, which
> >>>>> + * prevents it from signalling the fence the previous thread is stuck waiting
> >>>>> + * on::
> >>>>> + *
> >>>>> + *     lock(A);
> >>>>> + *     unlock(A);
> >>>>> + *     dma_fence_signal(B);
> >>>>> + *
> >>>>> + * By manually annotating all code relevant to signalling a &dma_fence we can
> >>>>> + * teach lockdep about these dependencies, which also helps with the validation
> >>>>> + * headache since now lockdep can check all the rules for us::
> >>>>> + *
> >>>>> + *    cookie = dma_fence_begin_signalling();
> >>>>> + *    lock(A);
> >>>>> + *    unlock(A);
> >>>>> + *    dma_fence_signal(B);
> >>>>> + *    dma_fence_end_signalling(cookie);
> >>>>> + *
> >>>>> + * For using dma_fence_begin_signalling() and dma_fence_end_signalling() to
> >>>>> + * annotate critical sections the following rules need to be observed:
> >>>>> + *
> >>>>> + * * All code necessary to complete a &dma_fence must be annotated, from the
> >>>>> + *   point where a fence is accessible to other threads, to the point where
> >>>>> + *   dma_fence_signal() is called. Un-annotated code can contain deadlock issues,
> >>>>> + *   and due to the very strict rules and many corner cases it is infeasible to
> >>>>> + *   catch these just with review or normal stress testing.
> >>>>> + *
> >>>>> + * * &struct dma_resv deserves a special note, since the readers are only
> >>>>> + *   protected by rcu. This means the signalling critical section starts as soon
> >>>>> + *   as the new fences are installed, even before dma_resv_unlock() is called.
> >>>>> + *
> >>>>> + * * The only exception are fast paths and opportunistic signalling code, which
> >>>>> + *   calls dma_fence_signal() purely as an optimization, but is not required to
> >>>>> + *   guarantee completion of a &dma_fence. The usual example is a wait IOCTL
> >>>>> + *   which calls dma_fence_signal(), while the mandatory completion path goes
> >>>>> + *   through a hardware interrupt and possible job completion worker.
> >>>>> + *
> >>>>> + * * To aid composability of code, the annotations can be freely nested, as long
> >>>>> + *   as the overall locking hierarchy is consistent. The annotations also work
> >>>>> + *   both in interrupt and process context. Due to implementation details this
> >>>>> + *   requires that callers pass an opaque cookie from
> >>>>> + *   dma_fence_begin_signalling() to dma_fence_end_signalling().
> >>>>> + *
> >>>>> + * * Validation against the cross driver contract is implemented by priming
> >>>>> + *   lockdep with the relevant hierarchy at boot-up. This means even just
> >>>>> + *   testing with a single device is enough to validate a driver, at least as
> >>>>> + *   far as deadlocks with dma_fence_wait() against dma_fence_signal() are
> >>>>> + *   concerned.
> >>>>> + */
> >>>>> +#ifdef CONFIG_LOCKDEP
> >>>>> +struct lockdep_map   dma_fence_lockdep_map = {
> >>>>> +     .name = "dma_fence_map"
> >>>>> +};
> >>>>
> >>>> Maybe a stupid question because this is definitely complicated, but.. If
> >>>> you have a single/static/global lockdep map, doesn't this mean _all_
> >>>> locks, from _all_ drivers happening to use dma-fences will get recorded
> >>>> in it. Will this work and not cause false positives?
> >>>>
> >>>> Sounds like it could create a common link between two completely
> >>>> unconnected usages. Because below you do add annotations to generic
> >>>> dma_fence_signal and dma_fence_wait.
> >>>
> >>> This is fully intentional. dma-fence is a cross-driver interface, if
> >>> every driver invents its own rules about how this should work we have
> >>> an unmaintainable and unreviewable mess.
> >>>
> >>> I've typed up the full length rant already here:
> >>>
> >>> https://lore.kernel.org/dri-devel/CAKMK7uGnFhbpuurRsnZ4dvRV9gQ_3-rmSJaoqSFY=+Kvepz_CA@mail.gmail.com/
> >>
> >> But "perfect storm" of:
> >>
> >>   + global fence lockmap
> >>   + mmu notifiers
> >>   + fs reclaim
> >>   + default annotations in dma_fence_signal / dma_fence_wait
> >>
> >> Equals to anything ever using dma_fence will be in impossible chains with random other drivers, even if neither driver has code to export/share that fence.
> >>
> >> Example from the CI run:
> >>
> >>   [25.918788] Chain exists of:
> >>    fs_reclaim --> mmu_notifier_invalidate_range_start --> dma_fence_map
> >>   [25.918794]  Possible unsafe locking scenario:
> >>   [25.918797]        CPU0                    CPU1
> >>   [25.918799]        ----                    ----
> >>   [25.918801]   lock(dma_fence_map);
> >>   [25.918803]                                lock(mmu_notifier_invalidate_range_start);
> >>   [25.918807]                                lock(dma_fence_map);
> >>   [25.918809]   lock(fs_reclaim);
> >>
> >> What about a dma_fence_export helper which would "arm" the annotations? It would be called as soon as the fence is exported. Maybe when added to dma_resv, or exported via sync_file, etc. Before that point begin/end_signaling and so would be no-ops.
> >
> > Run CI without the i915 annotation patch, nothing breaks.
>
> I think some parts of i915 would still break with my idea to only apply annotations on exported fences. What do you dislike about that idea? I thought the point is to enforce rules for _exported_ fences.

dma_fence is a shared concept, this is upstream, drivers are expected
to a) use shared concepts and b) use them in a consistent way. If
drivers do whatever they feel like then they're no maintainable in the
upstream sense of "maintainable even if the vendor walks away". This
was the reason why amd had to spend 2 refactoring from DAL (which used
all the helpers they shared with their firmware/windows driver) to DC
(which uses all the upstream kms helpers and datastructures directly).

> How you have annotated dma_fence_work you can't say, maybe it is exported maybe it isn't. I think it is btw, so splats would still be there, but I am not sure it is conceptually correct.
>
> At least my understanding is GFP_KERNEL allocations are only disallowed by the virtue of the global dma-fence contract. If you want to enforce they are never used for anything but exporting, then that would be a bit harsh, no?
>
> Another example from the CI run:
>
>  [26.585357]        CPU0                    CPU1
>  [26.585359]        ----                    ----
>  [26.585360]   lock(dma_fence_map);
>  [26.585362]                                lock(mmu_notifier_invalidate_range_start);
>  [26.585365]                                lock(dma_fence_map);
>  [26.585367]   lock(i915_gem_object_internal/1);
>  [26.585369]
>  *** DEADLOCK ***

So ime the above deadlock summaries tend to be wrong as soon as you
have more than 2 locks involved. Which we have here - they only ever
show at most 2 threads, with each thread only taking 2 locks in total,
which isn't going to deadlock if you have  more than 2 locks involved.
Which is the case above.

Personally I just ignore the above deadlock scenario and just always
look at all the locks and backtraces lockdep gives me, and then
reconstruct the dependency graph by hand myself, including deadlock
scenario.

> Lets say someone submitted an execbuf using userptr as a batch and then unmapped it immediately. That would explain CPU1 getting into the mmu notifier and waiting on this batch to unbind the object.
>
> Meanwhile CPU0 is the async command parser for this request trying to lock the shadow batch buffer. Because it uses the dma_fence_work this is between the begin/end signalling markers.
>
> It can be the same dma-fence I think, since we install the async parser fence on the real batch dma-resv, but dma_fence_map is not a real lock, so what is actually preventing progress in this case?
>
> CPU1 is waiting on a fence, but CPU0 can obtain the lock(i915_gem_object_internal/1), proceed to parse the batch, and exit the signalling section. At which point CPU1 is still blocked, waiting until the execbuf finishes and then mmu notifier can finish and invalidate the pages.
>
> Maybe I am missing something but I don't see how this one is real.

The above doesn't deadlock, and it also shouldn't result in a lockdep
splat. The trouble is when the signalling thread also grabs
i915_gem_object_internal/1 somewhere. Which if you go through full CI
results you see there's more involved (and at least one of the splats
is all just lockdep priming and might_lock, so could be an annotation
bug on top), and there is indeed a path where we lock the driver
private lock in more places, and the wrong way round. That's the thing
lockdep is complaining about, it's just not making that clear in the
summary because the summary is only ever correct for 2 locks. Not if
more is involved.

> > So we can gradually fix up existing code that doesn't quite get it
> > right and move on.
> >
> >>>>> +
> >>>>> +/**
> >>>>> + * dma_fence_begin_signalling - begin a critical DMA fence signalling section
> >>>>> + *
> >>>>> + * Drivers should use this to annotate the beginning of any code section
> >>>>> + * required to eventually complete &dma_fence by calling dma_fence_signal().
> >>>>> + *
> >>>>> + * The end of these critical sections are annotated with
> >>>>> + * dma_fence_end_signalling().
> >>>>> + *
> >>>>> + * Returns:
> >>>>> + *
> >>>>> + * Opaque cookie needed by the implementation, which needs to be passed to
> >>>>> + * dma_fence_end_signalling().
> >>>>> + */
> >>>>> +bool dma_fence_begin_signalling(void)
> >>>>> +{
> >>>>> +     /* explicitly nesting ... */
> >>>>> +     if (lock_is_held_type(&dma_fence_lockdep_map, 1))
> >>>>> +             return true;
> >>>>> +
> >>>>> +     /* rely on might_sleep check for soft/hardirq locks */
> >>>>> +     if (in_atomic())
> >>>>> +             return true;
> >>>>> +
> >>>>> +     /* ... and non-recursive readlock */
> >>>>> +     lock_acquire(&dma_fence_lockdep_map, 0, 0, 1, 1, NULL, _RET_IP_);
> >>>>
> >>>> Would it work if signalling path would mark itself as a write lock? I am
> >>>> thinking it would be nice to see in lockdep splats what are signals and
> >>>> what are waits.
> >>>
> >>> Yeah it'd be nice to have a read vs write name for the lock. But we
> >>> already have this problem for e.g. flush_work(), from which I've
> >>> stolen this idea. So it's not really new. Essentially look at the
> >>> backtraces lockdep gives you, and reconstruct the deadlock. I'm hoping
> >>> that people will notice the special functions on the backtrace, e.g.
> >>> dma_fence_begin_signalling will be listed as offending function/lock
> >>> holder, and then read the kerneldoc.
> >>>
> >>>> The recursive usage wouldn't work then right? Would write annotation on
> >>>> the wait path work?
> >>>
> >>> Wait path is write annotations already, but yeah annotating the
> >>> signalling side as write would cause endless amounts of alse
> >>> positives. Also it makes composability of these e.g. what I've done in
> >>> amdgpu with annotations in tdr work in drm/scheduler, annotations in
> >>> the amdgpu gpu reset code and then also annotations in atomic code,
> >>> which all nest within each other in some call chains, but not others.
> >>> Dropping the recursion would break that and make it really awkward to
> >>> annotate such cases correctly.
> >>>
> >>> And the recursion only works if it's read locks, otherwise lockdep
> >>> complains if you have inconsistent annotations on the signalling side
> >>> (which again would make it more or less impossible to annotate the
> >>> above case fully).
> >>
> >> How do I see in lockdep splats if it was a read or write user? Your patch appears to have:
> >>
> >> dma_fence_signal:
> >> +       lock_acquire(&dma_fence_lockdep_map, 0, 0, 1, 1, NULL, _RET_IP_);
> >>
> >> __dma_fence_might_wait:
> >> +       lock_acquire(&dma_fence_lockdep_map, 0, 0, 1, 1, NULL, _THIS_IP_);
> >>
> >> Which both seem like read lock. I don't fully understand the lockdep API so I might be wrong, not sure. But neither I see a difference in splats telling me which path is which.
> >
> > I think you got tricked by the implementation, this isn't quite what's
> > going on. There's two things which make the annotations special:
> >
> > - we want a recursive read lock on the signalling critical section.
> > The problem is that lockdep doesn't implement full validation for
> > recursive read locks, only non-recursive read/write locks fully
> > validated. There's some checks for recursive read locks, but exactly
> > the checks we need to catch common dma_fence_wait deadlocks aren't
> > done. That's why we need to implement manual lock recursion on the
> > reader side
> >
> > - now on the write side we additionally need to implement an
> > read2write upgrade, and a write2read downgrade. Lockdep doesn't
> > implement that, so again we have to hand-roll this.
> >
> > Let's go through the code line-by-line:
> >
> >      bool tmp;
> >
> >      tmp = lock_is_held_type(&dma_fence_lockdep_map, 1);
> >
> > We check whether someone is holding the non-recursive read lock already.
> >
> >      if (tmp)
> >          lock_release(&dma_fence_lockdep_map, _THIS_IP_);
> >
> > If that's the case, we drop that read lock.
> >
> >      lock_map_acquire(&dma_fence_lockdep_map);
> >
> > Then we do the actual might_wait annotation, the above takes the full
> > write lock ...
> >
> >      lock_map_release(&dma_fence_lockdep_map);
> >
> > ... and now we release the write lock again.
> >
> >
> >      if (tmp)
> >          lock_acquire(&dma_fence_lockdep_map, 0, 0, 1, 1, NULL, _THIS_IP_);
> >
> > Finally we need to re-acquire the read lock, if we've held that when
> > entering this function. This annotation naturally has to exactly match
> > what begin_signalling would do, otherwise the hand-rolled nesting
> > would fall apart.
> >
> > I hope that explains what's going on here, and assures you that
> > might_wait() is indeed a write lock annotation, but with a big pile of
> > complications.
>
> I am certainly confused by the difference between lock_map_acquire/release and lock_acquire/release. What is the difference between the two?

lock_acquire/release is a wrapper around lock_map_acquire/release.
This is all lockdep internal, it's a completely undocumented maze, so
unfortunately only option is to really careful follow all the
definitions from various locking primitives. And then compare with
lockdep self-test (which use the locking primitives, not the lockdep
internals) to see which flag controls which kind of behaviour.

That's at least what I do, and it's horrible. But yeah lockdep doesn't
have documentation for this.

If you think it's better to open code the lock_map/acquire, I guess I
can do that. But it's a mess, so I need to carefully retest everything
and make sure I've set the right flags and bits - for added fun they
also change ordering in some of the wrappers!
-Daniel
Chris Wilson June 19, 2020, 8:25 a.m. UTC | #14
Quoting Daniel Stone (2020-06-11 10:01:46)
> Hi,
> 
> On Thu, 11 Jun 2020 at 09:44, Dave Airlie <airlied@gmail.com> wrote:
> > On Thu, 11 Jun 2020 at 18:01, Chris Wilson <chris@chris-wilson.co.uk> wrote:
> > > Introducing a global lockmap that cannot capture the rules correctly,
> >
> > Can you document the rules all drivers should be following then,
> > because from here it looks to get refactored every version of i915,
> > and it would be nice if we could all aim for the same set of things
> > roughly. We've already had enough problems with amdgpu vs i915 vs
> > everyone else with fences, if this stops that in the future then I'd
> > rather we have that than just some unwritten rules per driver and
> > untestable.
> 
> As someone who has sunk a bunch of work into explicit-fencing
> awareness in my compositor so I can never be blocked, I'd be
> disappointed if the infrastructure was ultimately pointless because
> the documented fencing rules were \_o_/ or thereabouts. Lockdep
> definitely isn't my area of expertise so I can't comment on the patch
> per se, but having something to ensure we don't hit deadlocks sure
> seems a lot better than nothing.

This is doing dependency analysis on execution contexts which is a far
cry from doing the fence dependency analysis, and so has to actively
ignore the cycles that must exist on the dma side, and also the cycles
that prevent entering execution contexts on the CPU. It has to actively
ignore scheduler execution contexts, for lockdep cries, and so we do not
get analysis of the locking contexts along that path. This would be
solvable along the lines of extending lockdep ala lockdep_dma_enter().

Had i915's execution flow been marked up, it should have found the
dubious wait for external fences inside the dead GPU recovery, and
probably found a few more things to complain about with the reset locking.
[Note we already do the same annotations for wait-vs-reset, but not
reset-vs-execution.]

Determination of which waits are legal and which are not is entirely ad
hoc, for there is no status change tracking in the dependency analysis
[that is once an execution context is linked to a published fence, again
integral to lockdep.] Consider if the completion chain in atomic is
swapped out for the morally equivalent fences along intertwined timelines,
and so it does a bunch of dma_fence_wait() instead. Why are those waits
legal despite them being after we have committed to fulfilling the out
fence? [Why are the waits on and for the GPU legal, since they equally
block execution flow?]

Forcing a generic primitive to always be part of the same global map is
horrible. You forgo being able to use the primitive for unrelated tasks,
lose the ability to name particular contexts to gain more informative
dependency cycle reports from having the explicit linkage. You can add
wait_map tracking without loss of generality [in less than 10 lines],
and you can still enforce that all fences used for a common purpose
follow the same rules [the simplest way being to default to the singular
wait_map]. But it's the explicitly named execution contexts that are the
biggest boon to reading the code and reading the lockdep warns.

This is a bunch of ad hoc tracking for a very narrow purpose applied
globally, with loss of information.
-Chris
Daniel Vetter June 19, 2020, 8:51 a.m. UTC | #15
On Fri, Jun 19, 2020 at 10:25 AM Chris Wilson <chris@chris-wilson.co.uk> wrote:
>
> Quoting Daniel Stone (2020-06-11 10:01:46)
> > Hi,
> >
> > On Thu, 11 Jun 2020 at 09:44, Dave Airlie <airlied@gmail.com> wrote:
> > > On Thu, 11 Jun 2020 at 18:01, Chris Wilson <chris@chris-wilson.co.uk> wrote:
> > > > Introducing a global lockmap that cannot capture the rules correctly,
> > >
> > > Can you document the rules all drivers should be following then,
> > > because from here it looks to get refactored every version of i915,
> > > and it would be nice if we could all aim for the same set of things
> > > roughly. We've already had enough problems with amdgpu vs i915 vs
> > > everyone else with fences, if this stops that in the future then I'd
> > > rather we have that than just some unwritten rules per driver and
> > > untestable.
> >
> > As someone who has sunk a bunch of work into explicit-fencing
> > awareness in my compositor so I can never be blocked, I'd be
> > disappointed if the infrastructure was ultimately pointless because
> > the documented fencing rules were \_o_/ or thereabouts. Lockdep
> > definitely isn't my area of expertise so I can't comment on the patch
> > per se, but having something to ensure we don't hit deadlocks sure
> > seems a lot better than nothing.
>
> This is doing dependency analysis on execution contexts which is a far
> cry from doing the fence dependency analysis, and so has to actively
> ignore the cycles that must exist on the dma side, and also the cycles
> that prevent entering execution contexts on the CPU. It has to actively
> ignore scheduler execution contexts, for lockdep cries, and so we do not
> get analysis of the locking contexts along that path. This would be
> solvable along the lines of extending lockdep ala lockdep_dma_enter().

drm/scheduler is annotated, found some rather improbably to hit issues
in practice. But from the quick chat I've had with König and others I
think he agrees that it's real at least in the theoretical sense.
Probably should consider playing lottery if you hit it in practice
though :-)

> Had i915's execution flow been marked up, it should have found the
> dubious wait for external fences inside the dead GPU recovery, and
> probably found a few more things to complain about with the reset locking.
> [Note we already do the same annotations for wait-vs-reset, but not
> reset-vs-execution.]

I know it splats, that's why the tdr annotation patch comes with a
spec proposal for lifting the wait busting we do in i915 to the
dma_fence level. I included that because amdgpu has the same problem
on modern hw. Apparently their planned fix (because they've hit this
bug in testing) was to push some shared lock down into their
atomic_comit_tail function and use that in gpu reset, so don't seem
that interested in extending dma_fence.

For i915 it's just gen2/3 display, and cross-driver dma-buf/fence
usage for those is nil and won't change. Pragmatic solution imo would
be to just not annotate gpu reset on these platforms, and relying on
our wait busting plus igt tests to make sure it keeps working as-is.
The point of the explicit annotations for the signalling side is very
much that it can be rolled out gradually, and entirely left out for
old legacy paths that aren't worth fixing.

> Determination of which waits are legal and which are not is entirely ad
> hoc, for there is no status change tracking in the dependency analysis
> [that is once an execution context is linked to a published fence, again
> integral to lockdep.] Consider if the completion chain in atomic is
> swapped out for the morally equivalent fences along intertwined timelines,
> and so it does a bunch of dma_fence_wait() instead. Why are those waits
> legal despite them being after we have committed to fulfilling the out
> fence? [Why are the waits on and for the GPU legal, since they equally
> block execution flow?]

No need to consider, it's already real and resulted in some pretty
splats until I got the recursion handling right.

> Forcing a generic primitive to always be part of the same global map is
> horrible. You forgo being able to use the primitive for unrelated tasks,
> lose the ability to name particular contexts to gain more informative
> dependency cycle reports from having the explicit linkage. You can add
> wait_map tracking without loss of generality [in less than 10 lines],
> and you can still enforce that all fences used for a common purpose
> follow the same rules [the simplest way being to default to the singular
> wait_map]. But it's the explicitly named execution contexts that are the
> biggest boon to reading the code and reading the lockdep warns.

So one thing that's maybe not clear here: This doesn't track the DAG
of dependencies. Doesn't even try, I'm still faithfully assuming
drivers get that part right. Which is a gap and maybe we should fix
this, but not the goal here.

All this does is validate fences against anything else that might be
going on in the system. E.g. your recursion example for atomic is
handled by just assuming that any dma_fence_wait within a signalling
section is legit and correct. We can add this later on, but not with
lockdep, since lockdep works with classes. And proofing that
dma_fences are acyclic requires you track them all as individuals.
Entirely different things.

That still leaves the below:

> Forcing a generic primitive to always be part of the same global map is
> horrible.

And  no concrete example or reason for why that's not possible.
Because frankly it's not horrible, this is what upstream is all about:
Shared concepts, shared contracts, shared code.

The proposed patches might very well encode the wrong contract, that's
all up for discussion. But fundamentally questioning that we need one
is missing what upstream is all about.

> This is a bunch of ad hoc tracking for a very narrow purpose applied
> globally, with loss of information.

It doesn't solve every problem indeed. I'm happy to review patches to
check acyclic-ness of dma-fence at the global level from you, I
haven't figured out yet how to make that happen. I know i915-gem has
that, but this is about the cross-driver contract here.
-Daniel
Chris Wilson June 19, 2020, 9:13 a.m. UTC | #16
Quoting Daniel Vetter (2020-06-19 09:51:59)
> On Fri, Jun 19, 2020 at 10:25 AM Chris Wilson <chris@chris-wilson.co.uk> wrote:
> > Forcing a generic primitive to always be part of the same global map is
> > horrible.
> 
> And  no concrete example or reason for why that's not possible.
> Because frankly it's not horrible, this is what upstream is all about:
> Shared concepts, shared contracts, shared code.
> 
> The proposed patches might very well encode the wrong contract, that's
> all up for discussion. But fundamentally questioning that we need one
> is missing what upstream is all about.

Then I have not clearly communicated, as my opinion is not that
validation is worthless, but that the implementation is enshrining a
global property on a low level primitive that prevents it from being
used elsewhere. And I want to replace completion [chains] with fences, and
bio with fences, and closures with fences, and what other equivalencies
there are in the kernel. The fence is as central a locking construct as
struct completion and deserves to be a foundational primitive provided
by kernel/ used throughout all drivers for discrete problem domains.

This is narrowing dma_fence whereby adding
	struct lockdep_map *dma_fence::wait_map
and annotating linkage, allows you to continue to specify that all
dma_fence used for a particular purpose must follow common rules,
without restricting the primitive for uses outside of this scope.
-Chris
Daniel Vetter June 19, 2020, 9:43 a.m. UTC | #17
On Fri, Jun 19, 2020 at 10:13:35AM +0100, Chris Wilson wrote:
> Quoting Daniel Vetter (2020-06-19 09:51:59)
> > On Fri, Jun 19, 2020 at 10:25 AM Chris Wilson <chris@chris-wilson.co.uk> wrote:
> > > Forcing a generic primitive to always be part of the same global map is
> > > horrible.
> > 
> > And  no concrete example or reason for why that's not possible.
> > Because frankly it's not horrible, this is what upstream is all about:
> > Shared concepts, shared contracts, shared code.
> > 
> > The proposed patches might very well encode the wrong contract, that's
> > all up for discussion. But fundamentally questioning that we need one
> > is missing what upstream is all about.
> 
> Then I have not clearly communicated, as my opinion is not that
> validation is worthless, but that the implementation is enshrining a
> global property on a low level primitive that prevents it from being
> used elsewhere. And I want to replace completion [chains] with fences, and
> bio with fences, and closures with fences, and what other equivalencies
> there are in the kernel. The fence is as central a locking construct as
> struct completion and deserves to be a foundational primitive provided
> by kernel/ used throughout all drivers for discrete problem domains.
> 
> This is narrowing dma_fence whereby adding
> 	struct lockdep_map *dma_fence::wait_map
> and annotating linkage, allows you to continue to specify that all
> dma_fence used for a particular purpose must follow common rules,
> without restricting the primitive for uses outside of this scope.

Somewhere else in this thread I had discussions with Jason Gunthorpe about
this topic. It might maybe change somewhat depending upon exact rules, but
his take is very much "I don't want dma_fence in rdma". Or pretty close to
that at least.

Similar discussions with habanalabs, they're using dma_fence internally
without any of the uapi. Discussion there has also now concluded that it's
best if they remove them, and simply switch over to a wait_queue or
completion like every other driver does.

The next round of the patches already have a paragraph to at least
somewhat limit how non-gpu drivers use dma_fence. And I guess actual
consensus might be pointing even more strongly at dma_fence being solely
something for gpus and closely related subsystem (maybe media) for syncing
dma-buf access.

So dma_fence as general replacement for completion chains I think just
wont happen.

What might make sense is if e.g. the lockdep annotations could be reused,
at least in design, for wait_queue or completion or anything else
really. I do think that has a fair chance compared to the automagic
cross-release annotations approach, which relied way too heavily on
guessing where barriers are. My experience from just a bit of playing
around with these patches here and discussing them with other driver
maintainers is that accurately deciding where critical sections start and
end is a job for humans only. And if you get it wrong, you will have a
false positive.

And you're indeed correct that if we'd do annotations for completions and
wait queues, then that would need to have a class per semantically
equivalent user, like we have lockdep classes for mutexes, not just one
overall.

But dma_fence otoh is something very specific, which comes with very
specific rules attached - it's not a generic wait_queue at all. Originally
it did start out as one even, but it is a very specialized wait_queue.

So there's imo two cases:

- Your completion is entirely orthogonal of dma_fences, and can never ever
  block a dma_fence. Don't use dma_fence for this, and no problem. It's
  just another wait_queue somewhere.

- Your completion can eventually, maybe through lots of convolutions and
  depdencies, block a dma_fence. In that case full dma_fence rules apply,
  and the only thing you can do with a custom annotation is make the rules
  even stricter. E.g. if a sub-timeline in the scheduler isn't allowed to
  take certain scheduler locks. But the userspace visible/published fence
  do take them, maybe as part of command submission or retirement.
  Entirely hypotethical, no idea any driver actually needs this.

Cheers, Daniel
Chris Wilson June 19, 2020, 1:12 p.m. UTC | #18
Quoting Daniel Vetter (2020-06-19 10:43:09)
> On Fri, Jun 19, 2020 at 10:13:35AM +0100, Chris Wilson wrote:
> > Quoting Daniel Vetter (2020-06-19 09:51:59)
> > > On Fri, Jun 19, 2020 at 10:25 AM Chris Wilson <chris@chris-wilson.co.uk> wrote:
> > > > Forcing a generic primitive to always be part of the same global map is
> > > > horrible.
> > > 
> > > And  no concrete example or reason for why that's not possible.
> > > Because frankly it's not horrible, this is what upstream is all about:
> > > Shared concepts, shared contracts, shared code.
> > > 
> > > The proposed patches might very well encode the wrong contract, that's
> > > all up for discussion. But fundamentally questioning that we need one
> > > is missing what upstream is all about.
> > 
> > Then I have not clearly communicated, as my opinion is not that
> > validation is worthless, but that the implementation is enshrining a
> > global property on a low level primitive that prevents it from being
> > used elsewhere. And I want to replace completion [chains] with fences, and
> > bio with fences, and closures with fences, and what other equivalencies
> > there are in the kernel. The fence is as central a locking construct as
> > struct completion and deserves to be a foundational primitive provided
> > by kernel/ used throughout all drivers for discrete problem domains.
> > 
> > This is narrowing dma_fence whereby adding
> >       struct lockdep_map *dma_fence::wait_map
> > and annotating linkage, allows you to continue to specify that all
> > dma_fence used for a particular purpose must follow common rules,
> > without restricting the primitive for uses outside of this scope.
> 
> Somewhere else in this thread I had discussions with Jason Gunthorpe about
> this topic. It might maybe change somewhat depending upon exact rules, but
> his take is very much "I don't want dma_fence in rdma". Or pretty close to
> that at least.
> 
> Similar discussions with habanalabs, they're using dma_fence internally
> without any of the uapi. Discussion there has also now concluded that it's
> best if they remove them, and simply switch over to a wait_queue or
> completion like every other driver does.
> 
> The next round of the patches already have a paragraph to at least
> somewhat limit how non-gpu drivers use dma_fence. And I guess actual
> consensus might be pointing even more strongly at dma_fence being solely
> something for gpus and closely related subsystem (maybe media) for syncing
> dma-buf access.
> 
> So dma_fence as general replacement for completion chains I think just
> wont happen.

That is sad. I cannot comprehend going back to pure completions after a
taste of fence scheduling. And we are not even close to fully utilising
them, as not all the async cpu [allocation!] tasks are fully tracked by
fences yet and are still stuck in a FIFO workqueue.

> What might make sense is if e.g. the lockdep annotations could be reused,
> at least in design, for wait_queue or completion or anything else
> really. I do think that has a fair chance compared to the automagic
> cross-release annotations approach, which relied way too heavily on
> guessing where barriers are. My experience from just a bit of playing
> around with these patches here and discussing them with other driver
> maintainers is that accurately deciding where critical sections start and
> end is a job for humans only. And if you get it wrong, you will have a
> false positive.
> 
> And you're indeed correct that if we'd do annotations for completions and
> wait queues, then that would need to have a class per semantically
> equivalent user, like we have lockdep classes for mutexes, not just one
> overall.
> 
> But dma_fence otoh is something very specific, which comes with very
> specific rules attached - it's not a generic wait_queue at all. Originally
> it did start out as one even, but it is a very specialized wait_queue.
> 
> So there's imo two cases:
> 
> - Your completion is entirely orthogonal of dma_fences, and can never ever
>   block a dma_fence. Don't use dma_fence for this, and no problem. It's
>   just another wait_queue somewhere.
> 
> - Your completion can eventually, maybe through lots of convolutions and
>   depdencies, block a dma_fence. In that case full dma_fence rules apply,
>   and the only thing you can do with a custom annotation is make the rules
>   even stricter. E.g. if a sub-timeline in the scheduler isn't allowed to
>   take certain scheduler locks. But the userspace visible/published fence
>   do take them, maybe as part of command submission or retirement.
>   Entirely hypotethical, no idea any driver actually needs this.

I think we are faced with this very real problem.

The papering we have today over userptr is so very thin, and if you
squint you can already see it is coupled into the completion signal. Just
it happens to be on the other side of the fence.

The next batch of priority inversions involve integrating the async cpu
tasks into the scheduler, and have full dependency tracking over every
internal fence. I do not see any way to avoid coupling the completion
signal from the GPU to the earliest resource allocation, as it's an
unbroken chain of work, at least from the user's perspective. [Next up
for annotations is that we need to always assume that userspace has an
implicit lock on GPU resources; having to break that lock with a GPU
reset should be a breach of our data integrity, and best avoided, for
compute does not care one iota about system integrity and insist
userspace knows best.] Such allocations have to be allowed to fail and
for that failure to propagate cancelling the queued work, such that I'm
considering what rules we need for gfp_t. That might allow enough
leverage to break any fs_reclaim loops, but userptr is likely forever
doomed [aside from its fs_reclaim loop is as preventable as the normal
shrinker paths], but we still need to suggest to pin_user_pages that
failure is better than oom and that is not clear atm. Plus the usual
failure can happen at any time after updating the user facing
bookkeeping, but that is just extra layers in the execution monitor
ready to step in and replacing failing work with the error propagation.
Or where the system grinds to a halt, requiring the monitor to patch in
a new page / resource.
-Chris
Daniel Vetter June 22, 2020, 9:16 a.m. UTC | #19
On Fri, Jun 19, 2020 at 3:12 PM Chris Wilson <chris@chris-wilson.co.uk> wrote:
>
> Quoting Daniel Vetter (2020-06-19 10:43:09)
> > On Fri, Jun 19, 2020 at 10:13:35AM +0100, Chris Wilson wrote:
> > > Quoting Daniel Vetter (2020-06-19 09:51:59)
> > > > On Fri, Jun 19, 2020 at 10:25 AM Chris Wilson <chris@chris-wilson.co.uk> wrote:
> > > > > Forcing a generic primitive to always be part of the same global map is
> > > > > horrible.
> > > >
> > > > And  no concrete example or reason for why that's not possible.
> > > > Because frankly it's not horrible, this is what upstream is all about:
> > > > Shared concepts, shared contracts, shared code.
> > > >
> > > > The proposed patches might very well encode the wrong contract, that's
> > > > all up for discussion. But fundamentally questioning that we need one
> > > > is missing what upstream is all about.
> > >
> > > Then I have not clearly communicated, as my opinion is not that
> > > validation is worthless, but that the implementation is enshrining a
> > > global property on a low level primitive that prevents it from being
> > > used elsewhere. And I want to replace completion [chains] with fences, and
> > > bio with fences, and closures with fences, and what other equivalencies
> > > there are in the kernel. The fence is as central a locking construct as
> > > struct completion and deserves to be a foundational primitive provided
> > > by kernel/ used throughout all drivers for discrete problem domains.
> > >
> > > This is narrowing dma_fence whereby adding
> > >       struct lockdep_map *dma_fence::wait_map
> > > and annotating linkage, allows you to continue to specify that all
> > > dma_fence used for a particular purpose must follow common rules,
> > > without restricting the primitive for uses outside of this scope.
> >
> > Somewhere else in this thread I had discussions with Jason Gunthorpe about
> > this topic. It might maybe change somewhat depending upon exact rules, but
> > his take is very much "I don't want dma_fence in rdma". Or pretty close to
> > that at least.
> >
> > Similar discussions with habanalabs, they're using dma_fence internally
> > without any of the uapi. Discussion there has also now concluded that it's
> > best if they remove them, and simply switch over to a wait_queue or
> > completion like every other driver does.
> >
> > The next round of the patches already have a paragraph to at least
> > somewhat limit how non-gpu drivers use dma_fence. And I guess actual
> > consensus might be pointing even more strongly at dma_fence being solely
> > something for gpus and closely related subsystem (maybe media) for syncing
> > dma-buf access.
> >
> > So dma_fence as general replacement for completion chains I think just
> > wont happen.
>
> That is sad. I cannot comprehend going back to pure completions after a
> taste of fence scheduling. And we are not even close to fully utilising
> them, as not all the async cpu [allocation!] tasks are fully tracked by
> fences yet and are still stuck in a FIFO workqueue.
>
> > What might make sense is if e.g. the lockdep annotations could be reused,
> > at least in design, for wait_queue or completion or anything else
> > really. I do think that has a fair chance compared to the automagic
> > cross-release annotations approach, which relied way too heavily on
> > guessing where barriers are. My experience from just a bit of playing
> > around with these patches here and discussing them with other driver
> > maintainers is that accurately deciding where critical sections start and
> > end is a job for humans only. And if you get it wrong, you will have a
> > false positive.
> >
> > And you're indeed correct that if we'd do annotations for completions and
> > wait queues, then that would need to have a class per semantically
> > equivalent user, like we have lockdep classes for mutexes, not just one
> > overall.
> >
> > But dma_fence otoh is something very specific, which comes with very
> > specific rules attached - it's not a generic wait_queue at all. Originally
> > it did start out as one even, but it is a very specialized wait_queue.
> >
> > So there's imo two cases:
> >
> > - Your completion is entirely orthogonal of dma_fences, and can never ever
> >   block a dma_fence. Don't use dma_fence for this, and no problem. It's
> >   just another wait_queue somewhere.
> >
> > - Your completion can eventually, maybe through lots of convolutions and
> >   depdencies, block a dma_fence. In that case full dma_fence rules apply,
> >   and the only thing you can do with a custom annotation is make the rules
> >   even stricter. E.g. if a sub-timeline in the scheduler isn't allowed to
> >   take certain scheduler locks. But the userspace visible/published fence
> >   do take them, maybe as part of command submission or retirement.
> >   Entirely hypotethical, no idea any driver actually needs this.
>
> I think we are faced with this very real problem.
>
> The papering we have today over userptr is so very thin, and if you
> squint you can already see it is coupled into the completion signal. Just
> it happens to be on the other side of the fence.
>
> The next batch of priority inversions involve integrating the async cpu
> tasks into the scheduler, and have full dependency tracking over every
> internal fence. I do not see any way to avoid coupling the completion
> signal from the GPU to the earliest resource allocation, as it's an
> unbroken chain of work, at least from the user's perspective. [Next up
> for annotations is that we need to always assume that userspace has an
> implicit lock on GPU resources; having to break that lock with a GPU
> reset should be a breach of our data integrity, and best avoided, for
> compute does not care one iota about system integrity and insist
> userspace knows best.] Such allocations have to be allowed to fail and
> for that failure to propagate cancelling the queued work, such that I'm
> considering what rules we need for gfp_t. That might allow enough
> leverage to break any fs_reclaim loops, but userptr is likely forever
> doomed [aside from its fs_reclaim loop is as preventable as the normal
> shrinker paths], but we still need to suggest to pin_user_pages that
> failure is better than oom and that is not clear atm. Plus the usual
> failure can happen at any time after updating the user facing
> bookkeeping, but that is just extra layers in the execution monitor
> ready to step in and replacing failing work with the error propagation.
> Or where the system grinds to a halt, requiring the monitor to patch in
> a new page / resource.

Zooming out a bunch, since this is a lot about the details of making
this happen, and I want to make sure I'm understanding your aim
correctly. I think we have 2 big things here interacting:

On one side the "everything async" push, for some value of everything.
Once everything is async we let either the linux scheduler (for
dma_fence_work) or the gpu scheduler (for i915_request) figure out how
to order everything, with all the dependencies. For memory allocations
there's likely quite a bit of retrying (on the allocation side) and
skipping (on the shrinker/mmu notifier side) involved to make this all
pan out. Maybe something like a GFP_NOGPU flag.

On the other side we have opinionated userspace with both very
long-running batches (they might as well be infinite, best we can do
is check that they still preempt within a reasonable amount of time,
lack of hw support for preemption in all cases notwithstanding). And
batches which synchronize across engines and whatever entirely under
userspace controls, with stuff like gpu semaphore waits entirely in
the cmd stream, without any kernel or gpu scheduler involvement. Well
maybe a slightly smarter gpu scheduler which converts the semaphore
wait from a pure busy loop into a "repoll on each scheduler
timeslice". But not actual dependency tracking awareness in the kernel
(or guc/hw fwiw) of what userspace is really trying to do.

Later is a big motivator for the former, since with arbitrary long
batches and arbitrary fences any wait for a batch to complete can take
forever, hence anything that might end up doing that needs to be done
async and without locks. That way we don't have to shoot anything if a
batch takes too long.

Finally if anything goes wrong (on the kernel side at least) we just
propagete fence error state through the entire ladder of in-flight
things (only if it goes wrong terminally ofc).

Roughly correct or did I miss a big (or small but really important) thing?

Thanks, Daniel

Patch
diff mbox series

diff --git a/Documentation/driver-api/dma-buf.rst b/Documentation/driver-api/dma-buf.rst
index 63dec76d1d8d..05d856131140 100644
--- a/Documentation/driver-api/dma-buf.rst
+++ b/Documentation/driver-api/dma-buf.rst
@@ -100,11 +100,11 @@  CPU Access to DMA Buffer Objects
 .. kernel-doc:: drivers/dma-buf/dma-buf.c
    :doc: cpu access
 
-Fence Poll Support
-~~~~~~~~~~~~~~~~~~
+Implicit Fence Poll Support
+~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 .. kernel-doc:: drivers/dma-buf/dma-buf.c
-   :doc: fence polling
+   :doc: implicit fence polling
 
 Kernel Functions and Structures Reference
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -133,6 +133,12 @@  DMA Fences
 .. kernel-doc:: drivers/dma-buf/dma-fence.c
    :doc: DMA fences overview
 
+DMA Fence Signalling Annotations
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. kernel-doc:: drivers/dma-buf/dma-fence.c
+   :doc: fence signalling annotation
+
 DMA Fences Functions Reference
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
diff --git a/drivers/dma-buf/dma-fence.c b/drivers/dma-buf/dma-fence.c
index 656e9ac2d028..0005bc002529 100644
--- a/drivers/dma-buf/dma-fence.c
+++ b/drivers/dma-buf/dma-fence.c
@@ -110,6 +110,160 @@  u64 dma_fence_context_alloc(unsigned num)
 }
 EXPORT_SYMBOL(dma_fence_context_alloc);
 
+/**
+ * DOC: fence signalling annotation
+ *
+ * Proving correctness of all the kernel code around &dma_fence through code
+ * review and testing is tricky for a few reasons:
+ *
+ * * It is a cross-driver contract, and therefore all drivers must follow the
+ *   same rules for lock nesting order, calling contexts for various functions
+ *   and anything else significant for in-kernel interfaces. But it is also
+ *   impossible to test all drivers in a single machine, hence brute-force N vs.
+ *   N testing of all combinations is impossible. Even just limiting to the
+ *   possible combinations is infeasible.
+ *
+ * * There is an enormous amount of driver code involved. For render drivers
+ *   there's the tail of command submission, after fences are published,
+ *   scheduler code, interrupt and workers to process job completion,
+ *   and timeout, gpu reset and gpu hang recovery code. Plus for integration
+ *   with core mm with have &mmu_notifier, respectively &mmu_interval_notifier,
+ *   and &shrinker. For modesetting drivers there's the commit tail functions
+ *   between when fences for an atomic modeset are published, and when the
+ *   corresponding vblank completes, including any interrupt processing and
+ *   related workers. Auditing all that code, across all drivers, is not
+ *   feasible.
+ *
+ * * Due to how many other subsystems are involved and the locking hierarchies
+ *   this pulls in there is extremely thin wiggle-room for driver-specific
+ *   differences. &dma_fence interacts with almost all of the core memory
+ *   handling through page fault handlers via &dma_resv, dma_resv_lock() and
+ *   dma_resv_unlock(). On the other side it also interacts through all
+ *   allocation sites through &mmu_notifier and &shrinker.
+ *
+ * Furthermore lockdep does not handle cross-release dependencies, which means
+ * any deadlocks between dma_fence_wait() and dma_fence_signal() can't be caught
+ * at runtime with some quick testing. The simplest example is one thread
+ * waiting on a &dma_fence while holding a lock::
+ *
+ *     lock(A);
+ *     dma_fence_wait(B);
+ *     unlock(A);
+ *
+ * while the other thread is stuck trying to acquire the same lock, which
+ * prevents it from signalling the fence the previous thread is stuck waiting
+ * on::
+ *
+ *     lock(A);
+ *     unlock(A);
+ *     dma_fence_signal(B);
+ *
+ * By manually annotating all code relevant to signalling a &dma_fence we can
+ * teach lockdep about these dependencies, which also helps with the validation
+ * headache since now lockdep can check all the rules for us::
+ *
+ *    cookie = dma_fence_begin_signalling();
+ *    lock(A);
+ *    unlock(A);
+ *    dma_fence_signal(B);
+ *    dma_fence_end_signalling(cookie);
+ *
+ * For using dma_fence_begin_signalling() and dma_fence_end_signalling() to
+ * annotate critical sections the following rules need to be observed:
+ *
+ * * All code necessary to complete a &dma_fence must be annotated, from the
+ *   point where a fence is accessible to other threads, to the point where
+ *   dma_fence_signal() is called. Un-annotated code can contain deadlock issues,
+ *   and due to the very strict rules and many corner cases it is infeasible to
+ *   catch these just with review or normal stress testing.
+ *
+ * * &struct dma_resv deserves a special note, since the readers are only
+ *   protected by rcu. This means the signalling critical section starts as soon
+ *   as the new fences are installed, even before dma_resv_unlock() is called.
+ *
+ * * The only exception are fast paths and opportunistic signalling code, which
+ *   calls dma_fence_signal() purely as an optimization, but is not required to
+ *   guarantee completion of a &dma_fence. The usual example is a wait IOCTL
+ *   which calls dma_fence_signal(), while the mandatory completion path goes
+ *   through a hardware interrupt and possible job completion worker.
+ *
+ * * To aid composability of code, the annotations can be freely nested, as long
+ *   as the overall locking hierarchy is consistent. The annotations also work
+ *   both in interrupt and process context. Due to implementation details this
+ *   requires that callers pass an opaque cookie from
+ *   dma_fence_begin_signalling() to dma_fence_end_signalling().
+ *
+ * * Validation against the cross driver contract is implemented by priming
+ *   lockdep with the relevant hierarchy at boot-up. This means even just
+ *   testing with a single device is enough to validate a driver, at least as
+ *   far as deadlocks with dma_fence_wait() against dma_fence_signal() are
+ *   concerned.
+ */
+#ifdef CONFIG_LOCKDEP
+struct lockdep_map	dma_fence_lockdep_map = {
+	.name = "dma_fence_map"
+};
+
+/**
+ * dma_fence_begin_signalling - begin a critical DMA fence signalling section
+ *
+ * Drivers should use this to annotate the beginning of any code section
+ * required to eventually complete &dma_fence by calling dma_fence_signal().
+ *
+ * The end of these critical sections are annotated with
+ * dma_fence_end_signalling().
+ *
+ * Returns:
+ *
+ * Opaque cookie needed by the implementation, which needs to be passed to
+ * dma_fence_end_signalling().
+ */
+bool dma_fence_begin_signalling(void)
+{
+	/* explicitly nesting ... */
+	if (lock_is_held_type(&dma_fence_lockdep_map, 1))
+		return true;
+
+	/* rely on might_sleep check for soft/hardirq locks */
+	if (in_atomic())
+		return true;
+
+	/* ... and non-recursive readlock */
+	lock_acquire(&dma_fence_lockdep_map, 0, 0, 1, 1, NULL, _RET_IP_);
+
+	return false;
+}
+EXPORT_SYMBOL(dma_fence_begin_signalling);
+
+/**
+ * dma_fence_end_signalling - end a critical DMA fence signalling section
+ *
+ * Closes a critical section annotation opened by dma_fence_begin_signalling().
+ */
+void dma_fence_end_signalling(bool cookie)
+{
+	if (cookie)
+		return;
+
+	lock_release(&dma_fence_lockdep_map, _RET_IP_);
+}
+EXPORT_SYMBOL(dma_fence_end_signalling);
+
+void __dma_fence_might_wait(void)
+{
+	bool tmp;
+
+	tmp = lock_is_held_type(&dma_fence_lockdep_map, 1);
+	if (tmp)
+		lock_release(&dma_fence_lockdep_map, _THIS_IP_);
+	lock_map_acquire(&dma_fence_lockdep_map);
+	lock_map_release(&dma_fence_lockdep_map);
+	if (tmp)
+		lock_acquire(&dma_fence_lockdep_map, 0, 0, 1, 1, NULL, _THIS_IP_);
+}
+#endif
+
+
 /**
  * dma_fence_signal_locked - signal completion of a fence
  * @fence: the fence to signal
@@ -170,14 +324,19 @@  int dma_fence_signal(struct dma_fence *fence)
 {
 	unsigned long flags;
 	int ret;
+	bool tmp;
 
 	if (!fence)
 		return -EINVAL;
 
+	tmp = dma_fence_begin_signalling();
+
 	spin_lock_irqsave(fence->lock, flags);
 	ret = dma_fence_signal_locked(fence);
 	spin_unlock_irqrestore(fence->lock, flags);
 
+	dma_fence_end_signalling(tmp);
+
 	return ret;
 }
 EXPORT_SYMBOL(dma_fence_signal);
@@ -210,6 +369,8 @@  dma_fence_wait_timeout(struct dma_fence *fence, bool intr, signed long timeout)
 
 	might_sleep();
 
+	__dma_fence_might_wait();
+
 	trace_dma_fence_wait_start(fence);
 	if (fence->ops->wait)
 		ret = fence->ops->wait(fence, intr, timeout);
diff --git a/include/linux/dma-fence.h b/include/linux/dma-fence.h
index 3347c54f3a87..3f288f7db2ef 100644
--- a/include/linux/dma-fence.h
+++ b/include/linux/dma-fence.h
@@ -357,6 +357,18 @@  dma_fence_get_rcu_safe(struct dma_fence __rcu **fencep)
 	} while (1);
 }
 
+#ifdef CONFIG_LOCKDEP
+bool dma_fence_begin_signalling(void);
+void dma_fence_end_signalling(bool cookie);
+#else
+static inline bool dma_fence_begin_signalling(void)
+{
+	return true;
+}
+static inline void dma_fence_end_signalling(bool cookie) {}
+static inline void __dma_fence_might_wait(void) {}
+#endif
+
 int dma_fence_signal(struct dma_fence *fence);
 int dma_fence_signal_locked(struct dma_fence *fence);
 signed long dma_fence_default_wait(struct dma_fence *fence,