Re: [PATCH] dmaengine: device must have at least one channel

[Viresh Kumar]

On 03-08-16, 21:34, Greg Kroah-Hartman wrote:
> On Wed, Jul 27, 2016 at 02:32:58PM -0700, Viresh Kumar wrote:
> > The DMA device can't be registered if it doesn't have any channels
> > registered at all. Moreover, it leads to memory leak and is reported by
> > kmemleak as (on 3.10 kernel, and same shall happen on mainline):
> > 
> > unreferenced object 0xffffffc09e597240 (size 64):
> >   comm "swapper/0", pid 1, jiffies 4294877736 (age 7060.280s)
> >   hex dump (first 32 bytes):
> >     00 00 00 00 c0 ff ff ff 30 00 00 ff 00 00 00 ff  ........0.......
> >     00 00 00 ff 00 00 00 ff 00 00 00 ff 00 00 00 ff  ................
> >   backtrace:
> >     [<ffffffc0003079ec>] create_object+0x148/0x2a0
> >     [<ffffffc000cc150c>] kmemleak_alloc+0x80/0xbc
> >     [<ffffffc000303a7c>] kmem_cache_alloc_trace+0x120/0x1ac
> >     [<ffffffc00054771c>] dma_async_device_register+0x160/0x46c
> >     [<ffffffc000548958>] foo_probe+0x1a0/0x264
> >     [<ffffffc0005d6658>] platform_drv_probe+0x14/0x20
> >     [<ffffffc0005d50cc>] driver_probe_device+0x160/0x374
> >     [<ffffffc0005d538c>] __driver_attach+0x60/0x90
> >     [<ffffffc0005d3e78>] bus_for_each_dev+0x7c/0xb0
> >     [<ffffffc0005d4a0c>] driver_attach+0x1c/0x28
> >     [<ffffffc0005d459c>] bus_add_driver+0x124/0x248
> >     [<ffffffc0005d59cc>] driver_register+0x90/0x110
> >     [<ffffffc0005d6bf4>] platform_driver_register+0x58/0x64
> >     [<ffffffc00142a70c>] foo_driver_init+0x10/0x1c
> >     [<ffffffc000200878>] do_one_initcall+0xac/0x148
> >     [<ffffffc00140096c>] kernel_init_freeable+0x1a0/0x258
> > 
> > Return -ENODEV from dma_async_device_register() on such a case.
> > 
> > Signed-off-by: Viresh Kumar <viresh.kumar@xxxxxxxxxx>
> > ---
> > Hi Vinod,
> > 
> > Sorry if the fundamentals behind this patch are completely incorrect,
> > i.e. We *can't* register a dma device with 0 channels. Its been long
> > that I have worked on dma stuff :)
> > 
> >  drivers/dma/dmaengine.c | 7 +++++++
> >  1 file changed, 7 insertions(+)
> > 
> > diff --git a/drivers/dma/dmaengine.c b/drivers/dma/dmaengine.c
> > index 8c9f45fd55fc..6b535262ac5d 100644
> > --- a/drivers/dma/dmaengine.c
> > +++ b/drivers/dma/dmaengine.c
> > @@ -997,6 +997,13 @@ int dma_async_device_register(struct dma_device *device)
> >  		}
> >  		chan->client_count = 0;
> >  	}
> > +
> > +	if (!chancnt) {
> > +		dev_err(device->dev, "%s: device has no channels!\n", __func__);
> > +		rc = -ENODEV;
> > +		goto err_out;
> > +	}
> > +
> >  	device->chancnt = chancnt;
> >  
> >  	mutex_lock(&dma_list_mutex);
> 
> To follow up on this, as you found out, this patch breaks working
> hardware, so please don't apply it anywhere.  It needs more debugging to
> figure out exactly what is going on with some very strange hardware that
> seems to be relying on some strange code...

Thanks for your email Greg, and I know what the issue is.

So, the dmaengine core actually allows registering a device without any channels
and the dma driver can add channels later (directly to the dma-device channel
list), for example, from the of-xlate callback once someone requests for a
channel.

Though its not the right way of doing things, but it still works. The only thing
which breaks in that case is sysfs interface to dmaengine as the channels aren't
registered as devices and so they aren't visible in userspace, as that code is
*only* executed while the dma-device is registered.

I think its dmaengine core's decision to keep supporting such drivers or not and
so I would like Vinod to comment on that..

But if we do want to support such drivers, then the core must be updated to fix
the memleak I reported here, as idr_ref is allocated and not used at all (also
dma-id allocated for it).

Or provide another API to properly and fully add channels at a later point of
time.

-- 
viresh