Re: [PATCH 1/3] habanalabs: implement dma-fence mechanism

[Daniel Vetter <daniel.vetter@ffwll.ch>]

On Mon, Jul 13, 2020 at 9:03 PM Jason Gunthorpe <jgg@ziepe.ca> wrote:
>
> On Mon, Jul 13, 2020 at 08:34:12PM +0200, Daniel Vetter wrote:
> > On Mon, Jul 13, 2020 at 5:57 PM Greg Kroah-Hartman
> > <gregkh@linuxfoundation.org> wrote:
> > >
> > > On Mon, Jul 13, 2020 at 06:54:22PM +0300, Oded Gabbay wrote:
> > > > From: Ofir Bitton <obitton@habana.ai>
> > > >
> > > > Instead of using standard dma-fence mechanism designed for GPU's, we
> > > > introduce our own implementation based on the former one. This
> > > > implementation is much more sparse than the original, contains only
> > > > mandatory functionality required by the driver.
> > >
> > > Sad you can't use the in-kernel code for this, I really don't understand
> > > what's wrong with using it as-is.
> > >
> > > Daniel, why do we need/want duplicate code floating around in the tree
> > > like this?
> >
> > The rules around dma-fence are ridiculously strict, and it only makes
> > sense to inflict that upon you if you actually want to participate in
> > the cross driver uapi built up around dma-buf and dma-fence.
> >
> > I've recently started some lockdep annotations to better enforce these
> > rules (and document them), and it's finding tons of subtle bugs even
> > in drivers/gpu (and I only just started with annotating drivers:
> >
> > https://lore.kernel.org/dri-devel/20200707201229.472834-1-daniel.vetter@ffwll.ch/
> >
> > You really don't want to deal with this if you don't have to. If
> > drivers/gpu folks (who created this) aren't good enough to understand
> > it, maybe it's not a good idea to sprinkle this all over the tree. And
> > fundamentally all this is is a slightly fancier struct completion. Use
> > that one instead, or a wait_queue.
> >
> > I discussed this a bit with Oded, and he thinks it's easier to
> > copypaste and simplify, but given that all other drivers seem to get
> > by perfectly well with completion or wait_queue, I'm not sure that's a
> > solid case.
> >
> > Also adding Jason Gunthorpe, who very much suggested this should be
> > limited to dma-buf/gpu related usage only.
>
> Without all the cross-driver stuff dma_fence is just a
> completion. Using dma_fence to get a completion is big abuse of what
> it is intended for.
>
> I think the only problem with this patch is that it keeps too much of
> the dma_fence stuff around. From what I could tell it really just
> wants to add a kref and completion to struct hl_cs_compl and delete
> everything to do with dma_fence.

Yeah, that's what I recommended doing too. error flag might be needed
too I think, but that's it.
-Daniel

> Not even sure the kref is needed ;)
>
> Jason



-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch