On Tue Nov 17, 2026 at 3:50 PM CET, Christian König wrote: > On 11/17/26 14:50, Danilo Krummrich wrote: >> On Tue Nov 17, 2026 at 2:40 PM CET, Christian König wrote: >>> I completely agree that we can't fix things from the dma_fence code, but it >>> is our obligation to tell the user that something very very bad happened. >> >> That's the point, we can't infer that "something very bad happened" from >> dma::DriverFence being dropped unsignaled. >> >> In Rust we use the RAII pattern, and drivers structurally embed a >> dma::DriverFence within the struct that represents the corresponding >> hardware / >> firmware queue (e.g. through Jobqueue). >> >> Consequently, if dma::DriverFence drops unsignaled it means that the >> destructor >> of the structure that embeds it (which does the required HW teardown) did run >> already. > > Oh WOW! The signaling state of the fence should *NEVER* be bound to some > object lifetime.
Let's engage with the argument and understand things before throwing overgeneralized statements like this around please. Realitiy is usually much more nuanced and what you say doesn't hold in all cases. The case described above (and clarified below) is one of them. I assume the fundamental reason why you say things like "software object running out of scope is not an indicator that the HW stops doing something" is because the lifetime model in a lot of existing C DRM drivers is backwards. Drivers often reference count representations of hardware resources and hardware states and hence keep them alive beyond what their actual lifetime should be -- even beyond driver unbind. This causes lot's of issues and I'm looking into fixing this up, see my talk in [1] and an example from Xe [2] for reference. That said, in Rust we don't have this kind of ambiguity, lifetime and ownership of resources are well defined and the core infrastructure guides drivers accordingly. Consequently, it is perfectly reasonable to assume that a driver object that owns device resources (such as DMA memory) will, in its destructor, tear down the hardware to not access any of the resources it owns anymore. For instance, think about how your logic would apply to e.g. DMA coherent allocations. Those are allocated with e.g. dma::Coherent::zeroed() and are freed on drop(). They are then owned by whatever object represents the lifetime of the HW using this buffer. Following your logic, we'd need to conclude that a dma::Coherent allocation needs an explicit free() method independent of its destructor as well, and have the destructor print a warning and then call free() anyway. Now, eventually drivers would just call free() from the destructor of the object owning the dma::Coherent allocation, so we wouldn't win anything. More in general, in Rust the object taking ownership of a device resource is responsible to deal with the device related state that makes use of this resource regardless. Insisting on an explicit method call is not improving anything in this model. To get back to dma::DriverFence: This is slightly different, as there's two different signal states, but the fundamental guarantee is the same: If it is dropped, the HW has stopped accessing the resource that it guards already *by design*. Signaling with an error on drop is arguably the correct thing to do, since if the operation would have been completed successfully before already, it would have been signaled already -- if it drops unsignaled it means the hardware was forcefully torn down (e.g. on driver unbind). [1] https://youtu.be/3A2vGbc_rxs?t=668 [2] https://lore.kernel.org/all/[email protected]/ > That sounds like an extremely bad practice to me. Just imagine anybody > grabbing an extra reference to the HW state object to inspect it for debugging > and the fence not signaling because of this. That's not how it works. Rust ownership rules will guarantee that the driver fence will be signaled with an error (if not signaled already) if the HW state object is destroyed (i.e. its destructor runs that tears down the HW). But that doesn't mean that the fence can't be signaled (and subsequently be dropped) independently through the completion of a job before that. IOW, the HW state object (for instance) owns Jobqueue, Jobqueue owns jobs, jobs own driver fences. If the HW state object drops (e.g. on driver unbind), the Jobqueue drops, which makes their jobs drop, which makes the driver fences drop. At this point the driver fence either has been signaled already because the job completed before teardown (but wasn't dropped yet), or it will be signaled with an error in its drop(), which is fine since at this point the HW state's drop() is completed already and hence the HW is stopped. Besides that, reference counting HW state objects is bad practice. HW state objects have a bounded lifetime, so there's no point in reference counting them.
