On Mon, Jun 08, 2026 at 06:54:28PM -0700, Sami Tolvanen wrote:
> On Mon, Jun 8, 2026 at 5:33 PM Sami Tolvanen <[email protected]> wrote:
> >
> > GpuVm implements Send and Sync unconditionally, but it shares and drops
> > T::VmBoData and T::Object across threads: obtain() can return ARefs to
> > the same bo on several threads, and deferred_cleanup() drops them on
> > whichever thread calls it. This is unsound unless both are Send + Sync,
> > so bound both impls accordingly.
> 
> Sashiko keeps finding more issues here. For the benefit of those not
> subscribed to dri-devel:
> 
> > This is a pre-existing issue, but does GpuVm also need Send + Sync bounds
> > for T::VaData here?
> >
> > Since GpuVm<T> explicitly bypasses the lack of Send on VaData,
> > UniqueRefGpuVm<T> automatically derives Send. This allows safe code to map
> > a VA containing !Send data on one thread, send the UniqueRefGpuVm<T> to
> > another thread, and unmap the VA there.
> >
> > Unmapping via sm_unmap() returns a GpuVaRemoved<T> which drops the !Send
> > VaData on the wrong thread, potentially causing data races.
> 
> Looking at the code, UniqueRefGpuVm seems like a better place for this
> than GpuVm. However, before I go add another patch to the series, I
> wonder if it would make more sense to just add these bounds to the
> DriverGpuVm trait instead? I'd be happy to hear some non-AI thoughts
> about this too.

Adding bounds on the trait does sound like a good idea for simplicity.
It's not like we're ever going to use GPUVM with non-thread-safe types.

If it's help, note that you can use `where GpuVm<T>: Send` bounds or
similar to "reuse" the implementations elsewhere. For example,
GpuVmBo<T> being Sync requires GpuVm<T> being Sync because of the
GpuVmBo::gpuvm() method, so you can use `where GpuVm<T>: Send` as a
bound to directly express that if you want.

But if we just add bounds to the trait, it may not be necessary.

Alice

Reply via email to