Re: [PATCH] dma-fence: Correct return of dma_fence_driver_name()

[Tvrtko Ursulin <tvrtko.ursulin@igalia.com>]

On 24/10/2025 11:59, Philipp Stanner wrote:
> On Fri, 2025-10-24 at 09:31 +0100, Tvrtko Ursulin wrote:
>>
>> On 24/10/2025 08:50, Philipp Stanner wrote:
>>> To decouple the dma_fence_ops lifetime from dma_fences lifetime RCU
>>> support was added to said function, coupled with using the signaled bit
>>> to detect whether the fence_ops might be gone already.
>>>
>>> When implementing that a wrong string was set as a default return
>>> parameter, indicating that every driver whose fence is already signalled
>>> must be detached, which is frankly wrong.
>>
>> Depends on how you look at it. After being signaled fence has to be
>> detached from the driver. Ie. nothing belonging to this driver must be
>> accessed via the fence.
> 
> Is that even documented btw? Many of the mysterious "dma fence rules"
> are often only obtainable by asking Christian & Co

I tried to document it in the very patch which added it. Both in the 
commit message and in the large sticky-outy comments added to these helpers:

"""
  * dma_fence_driver_name - Access the driver name
  * @fence: the fence to query
  *
  * Returns a driver name backing the dma-fence implementation.
  *
  * IMPORTANT CONSIDERATION:
  * Dma-fence contract stipulates that access to driver provided data 
(data not
  * directly embedded into the object itself), such as the 
&dma_fence.lock and
  * memory potentially accessed by the &dma_fence.ops functions, is 
forbidden
  * after the fence has been signalled. Drivers are allowed to free that 
data,
  * and some do.
  *
  * To allow safe access drivers are mandated to guarantee a RCU grace 
period
  * between signalling the fence and freeing said data.
  *
  * As such access to the driver name is only valid inside a RCU locked 
section.
  * The pointer MUST be both queried and USED ONLY WITHIN a SINGLE block 
guarded
  * by the &rcu_read_lock and &rcu_read_unlock pair.
"""

> 
>>
>> I started with names and Christian has recently continued with ops.
>>
>>> Reported-by: Danilo Krummrich <dakr@kernel.org>
>>> Fixes: 506aa8b02a8d ("dma-fence: Add safe access helpers and document the rules")
>>> Signed-off-by: Philipp Stanner <phasta@kernel.org>
>>> ---
>>> When this was merged, it sadly slipped by me. I think this entire RCU
>>> mechanism was / is an overengineered idea.
>>>
>>> If we look at who actually uses dma_fence_driver_name() and
>>> dma_fence_timeline_name() – functions from which the largest share of
>>> the fence_ops vs. fence lifetime issue stems from – we discover that
>>> there is a single user:
>>>
>>> i915.
>>
> 
> […]
> 
>>
>>
>> That would be nice, I also do not see much value in exporting names to
>> userspace. But first more conversation around breaking the sync file ABI
>> needs to happen. I think we had a little bit of it when changing the
>> names of signalled fences and thinking was existing tools which look at
>> the names will mostly survive it. Not sure if they would if unsignalled
>> names would change.
> 
> I mean, what you and Christian are addressing in recent weeks are real
> problems, and I was / am about to write similar solutions for our Rust
> dma_fence.
> 
> In the case of those names, however, I'll likely just not support that
> in Rust, saving me from adding those RCU guards and delivering output
> of questionable use to users.
> (could ofc be added later by someone who really needs it…)

Sounds like a good plan to start without the problematic parts whenever 
possible. More than that I cannot comment since I have no idea how rust 
stuff will work and interact with the existing uapi entry points such as 
the sync file.

>>> P.
>>> ---
>>>    drivers/dma-buf/dma-fence.c | 2 +-
>>>    1 file changed, 1 insertion(+), 1 deletion(-)
>>>
>>> diff --git a/drivers/dma-buf/dma-fence.c b/drivers/dma-buf/dma-fence.c
>>> index 3f78c56b58dc..1875a0abebd3 100644
>>> --- a/drivers/dma-buf/dma-fence.c
>>> +++ b/drivers/dma-buf/dma-fence.c
>>> @@ -1111,7 +1111,7 @@ const char __rcu *dma_fence_driver_name(struct dma_fence *fence)
>>>    	if (!test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
>>>    		return fence->ops->get_driver_name(fence);
>>>    	else
>>> -		return "detached-driver";
>>> +		return "driver-whose-fence-is-already-signalled";
>>
>> IMHO unnecessarily verbose and whether or not changing it to anything
>> different warrants a Fixes: tag is debatable.
> 
> IMO the output is just wrong and confusing. It's easy to imagine that
> some user starts wondering and searching why his driver has been
> unloaded, opening support tickets and so on.
> 
> Could be less verbose, though. Dunno. I let the maintainer decide.

Driver and timeline usually come together so the signalled info is 
already there ie. "detached-driver signaled-timeline". For example in 
debugfs via dma_fence_describe().

So changing that to "driver-whose-fence-is-already-signalled 
signaled-timeline" still looks too much.

Also, the short name can be reduced from a verbose starting point 
similar to yours:

  "unknown-driver-is-detached-from-the-signaled-fence"
  "driver-detached-from-the-fence"
  "driver-detached"

Or keep "detached-driver" as good enough. Mea culpa for typing it up 
transposed. :)

Regards,

Tvrtko

> 
> P.
> 
>>
>> Regards,
>>
>> Tvrtko
>>
>>>    }
>>>    EXPORT_SYMBOL(dma_fence_driver_name);
>>>    
>>
>