Re: [PATCH RFC 1/1] dmaengine: ptdma: use SLAB_TYPESAFE_BY_RCU for the DMA descriptor slab

[Eder Zulian <ezulian@redhat.com>]

Hello Nathan,

On Thu, Apr 17, 2025 at 04:02:23PM -0500, Nathan Lynch wrote:
> Eder Zulian <ezulian@redhat.com> writes:
> > The SLAB_TYPESAFE_BY_RCU flag prevents a change of type for objects
> > allocated from the slab cache (although the memory may be reallocated to
> > a completetly different object of the same type.) Moreover, when the
> > last reference to an object is dropped the finalization code must not
> > run until all __rcu pointers referencing the object have been updated,
> > and then a grace period has passed.
> >
> > Signed-off-by: Eder Zulian <ezulian@redhat.com>
> > ---
> >  drivers/dma/amd/ptdma/ptdma-dmaengine.c | 3 ++-
> >  1 file changed, 2 insertions(+), 1 deletion(-)
> >
> > diff --git a/drivers/dma/amd/ptdma/ptdma-dmaengine.c b/drivers/dma/amd/ptdma/ptdma-dmaengine.c
> > index 715ac3ae067b..b70dd1b0b9fb 100644
> > --- a/drivers/dma/amd/ptdma/ptdma-dmaengine.c
> > +++ b/drivers/dma/amd/ptdma/ptdma-dmaengine.c
> > @@ -597,7 +597,8 @@ int pt_dmaengine_register(struct pt_device *pt)
> >  
> >  	pt->dma_desc_cache = kmem_cache_create(desc_cache_name,
> >  					       sizeof(struct pt_dma_desc), 0,
> > -					       SLAB_HWCACHE_ALIGN, NULL);
> > +					       SLAB_HWCACHE_ALIGN |
> > +					       SLAB_TYPESAFE_BY_RCU, NULL);
> 
> No, this code wasn't written to exploit SLAB_TYPESAFE_BY_RCU and this
> change can only obscure the problem. There's likely a data race in the
> driver.
> 

Ack. Let's conclude my RFC and discard the proposed patch then.
Thank you very much for your feedback. 

> I suspect pt_cmd_callback_work() has a bug:
> 
>         spin_lock_irqsave(&chan->vc.lock, flags);
>         if (desc) {
>                 if (desc->status != DMA_COMPLETE) {
>                         if (desc->status != DMA_ERROR)
>                                 desc->status = DMA_COMPLETE;
> 
>                         dma_cookie_complete(tx_desc);
>                         dma_descriptor_unmap(tx_desc);
>                 } else {
>                         tx_desc = NULL;
>                 }
>         }
>         spin_unlock_irqrestore(&chan->vc.lock, flags);
> 
>         if (tx_desc) {
>                 dmaengine_desc_get_callback_invoke(tx_desc, NULL);
>                 dma_run_dependencies(tx_desc);
> >>>>            list_del(&desc->vd.node); <<< must be done under vc.lock
>                 vchan_vdesc_fini(vd);
>         }
> 
> But that's relatively new code that may not be in the kernel you're
> running.
> 

True. pt_cmd_callback_work() wasn't in the kernel used for tests and it
seems to be used only if 'pt->ver == AE4_DMA_VERSION'. In that kernel
pt_cmd_callback() would call pt_handle_active_desc() which seemed to have
the same bug.

Eder