Received-SPF: pass (google.com: domain of linux-kernel+bounces-20031-linux.lists.archive=gmail.com@vger.kernel.org designates 139.178.88.99 as permitted sender) client-ip=139.178.88.99;
Date: Mon, 8 Jan 2024 20:19:25 +0100
From: Daniel Vetter <daniel@ffwll.ch>
To: Paul Cercueil <paul@crapouillou.net>
Cc: Daniel Vetter <daniel@ffwll.ch>,
	Greg Kroah-Hartman <gregkh@linuxfoundation.org>,
	Sumit Semwal <sumit.semwal@linaro.org>,
	Christian =?iso-8859-1?Q?K=F6nig?= <christian.koenig@amd.com>,
	Jonathan Corbet <corbet@lwn.net>,
	Michael Hennerich <Michael.Hennerich@analog.com>,
	linux-doc@vger.kernel.org, linux-usb@vger.kernel.org,
	linux-kernel@vger.kernel.org, dri-devel@lists.freedesktop.org,
	Andrzej Pietrasiewicz <andrzej.p@collabora.com>,
	linaro-mm-sig@lists.linaro.org,
	Nuno =?iso-8859-1?Q?S=E1?= <noname.nuno@gmail.com>,
	Jonathan Cameron <jic23@kernel.org>, linux-media@vger.kernel.org
Subject: Re: [PATCH v3 3/4] usb: gadget: functionfs: Add DMABUF import
 interface
Message-ID: <ZZxKvR9gjH8D5qxj@phenom.ffwll.local>
Mail-Followup-To: Paul Cercueil <paul@crapouillou.net>,
	Greg Kroah-Hartman <gregkh@linuxfoundation.org>,
	Sumit Semwal <sumit.semwal@linaro.org>,
	Christian =?iso-8859-1?Q?K=F6nig?= <christian.koenig@amd.com>,
	Jonathan Corbet <corbet@lwn.net>,
	Michael Hennerich <Michael.Hennerich@analog.com>,
	linux-doc@vger.kernel.org, linux-usb@vger.kernel.org,
	linux-kernel@vger.kernel.org, dri-devel@lists.freedesktop.org,
	Andrzej Pietrasiewicz <andrzej.p@collabora.com>,
	linaro-mm-sig@lists.linaro.org,
	Nuno =?iso-8859-1?Q?S=E1?= <noname.nuno@gmail.com>,
	Jonathan Cameron <jic23@kernel.org>, linux-media@vger.kernel.org
References: <20240108120056.22165-1-paul@crapouillou.net>
 <20240108120056.22165-4-paul@crapouillou.net>
 <ZZvtEXL8DLPPdtPs@phenom.ffwll.local>
 <a44aca93adc60ce56a64c50797a029631900172e.camel@crapouillou.net>
 <ZZwU827NMHbx7bsO@phenom.ffwll.local>
 <2c0d4ef1b657c56ea2290fe16d757ce563a3e71b.camel@crapouillou.net>
Precedence: bulk
MIME-Version: 1.0
Content-Type: text/plain; charset=iso-8859-1
Content-Disposition: inline
Content-Transfer-Encoding: 8bit
In-Reply-To: <2c0d4ef1b657c56ea2290fe16d757ce563a3e71b.camel@crapouillou.net>

On Mon, Jan 08, 2024 at 05:27:33PM +0100, Paul Cercueil wrote:
> Le lundi 08 janvier 2024 ? 16:29 +0100, Daniel Vetter a ?crit?:
> > On Mon, Jan 08, 2024 at 03:21:21PM +0100, Paul Cercueil wrote:
> > > Hi Daniel (Sima?),
> > > 
> > > Le lundi 08 janvier 2024 ? 13:39 +0100, Daniel Vetter a ?crit?:
> > > > On Mon, Jan 08, 2024 at 01:00:55PM +0100, Paul Cercueil wrote:
> > > > > +static void ffs_dmabuf_signal_done(struct ffs_dma_fence
> > > > > *dma_fence, int ret)
> > > > > +{
> > > > > +	struct ffs_dmabuf_priv *priv = dma_fence->priv;
> > > > > +	struct dma_fence *fence = &dma_fence->base;
> > > > > +
> > > > > +	dma_fence_get(fence);
> > > > > +	fence->error = ret;
> > > > > +	dma_fence_signal(fence);
> > > > > +
> > > > > +	dma_buf_unmap_attachment(priv->attach, dma_fence->sgt,
> > > > > dma_fence->dir);
> > > > > +	dma_fence_put(fence);
> > > > > +	ffs_dmabuf_put(priv->attach);
> > > > 
> > > > So this can in theory take the dma_resv lock, and if the usb
> > > > completion
> > > > isn't an unlimited worker this could hold up completion of future
> > > > dma_fence, resulting in a deadlock.
> > > > 
> > > > Needs to be checked how usb works, and if stalling indefinitely
> > > > in
> > > > the
> > > > io_complete callback can hold up the usb stack you need to:
> > > > 
> > > > - drop a dma_fence_begin/end_signalling annotations in here
> > > > - pull out the unref stuff into a separate preallocated worker
> > > > (or at
> > > > ? least the final unrefs for ffs_dma_buf).
> > > 
> > > Only ffs_dmabuf_put() can attempt to take the dma_resv and would
> > > have
> > > to be in a worker, right? Everything else would be inside the
> > > dma_fence_begin/end_signalling() annotations?
> > 
> > Yup. Also I noticed that unlike the iio patches you don't have the
> > dma_buf_unmap here in the completion path (or I'm blind?), which
> > helps a
> > lot with avoiding trouble.
> 
> They both call dma_buf_unmap_attachment() in the "signal done"
> callback, the only difference I see is that it is called after the
> dma_fence_put() in the iio patches, while it's called before
> dma_fence_put() here.

I was indeed blind ...

So the trouble is this wont work because:
- dma_buf_unmap_attachment() requires dma_resv_lock. This is a somewhat
  recent-ish change from 47e982d5195d ("dma-buf: Move
  dma_buf_map_attachment() to dynamic locking specification"), so maybe
  old kernel or you don't have full lockdep enabled to get the right
  splat.

- dma_fence critical section forbids dma_resv_lock

Which means you need to move this out, but then there's the potential
cache management issue. Which current gpu drivers just kinda ignore
because it doesn't matter for current use-case, they all cache the mapping
for about as long as the attachment exists. You might want to do the same,
unless that somehow breaks a use-case you have, I have no idea about that.
If something breaks with unmap_attachment moved out of the fence handling
then I guess it's high time to add separate cache-management only to
dma_buf (and that's probably going to be quite some wiring up, not sure
even how easy that would be to do nor what exactly the interface should
look like).

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