gpu: host1x: Free the IOMMU domain when there is no device to attach

Submitted by Paul Kocialkowski on July 10, 2017, 7:33 p.m.

Details

Message ID 20170710193305.5987-1-contact@paulk.fr
State New
Headers show

Commit Message

Paul Kocialkowski July 10, 2017, 7:33 p.m.
When there is no device to attach to the IOMMU domain, as may be the
case when the device-tree does not contain the proper iommu node, it is
best to keep going without IOMMU support rather than failing.
This allows the driver to probe and function instead of taking down
all of the tegra drm driver, leading to missing display support.

Signed-off-by: Paul Kocialkowski <contact@paulk.fr>
---
 drivers/gpu/host1x/dev.c | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

Comments

Mikko Perttunen July 11, 2017, 6:43 a.m.
Thanks for the patch, didn't consider this case. I really need to get 
together some system to automatically test on multiple platforms.. :)

Reviewed-by: Mikko Perttunen <mperttunen@nvidia.com>

On 10.07.2017 22:33, Paul Kocialkowski wrote:
> When there is no device to attach to the IOMMU domain, as may be the
> case when the device-tree does not contain the proper iommu node, it is
> best to keep going without IOMMU support rather than failing.
> This allows the driver to probe and function instead of taking down
> all of the tegra drm driver, leading to missing display support.
>
> Signed-off-by: Paul Kocialkowski <contact@paulk.fr>
> ---
>  drivers/gpu/host1x/dev.c | 8 +++++++-
>  1 file changed, 7 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/gpu/host1x/dev.c b/drivers/gpu/host1x/dev.c
> index ac65f52850a6..f296738d0de8 100644
> --- a/drivers/gpu/host1x/dev.c
> +++ b/drivers/gpu/host1x/dev.c
> @@ -186,8 +186,13 @@ static int host1x_probe(struct platform_device *pdev)
>  			return -ENOMEM;
>
>  		err = iommu_attach_device(host->domain, &pdev->dev);
> -		if (err)
> +		if (err == -ENODEV) {
> +			iommu_domain_free(host->domain);
> +			host->domain = NULL;
> +			goto iommu_skip;
> +		} else if (err) {
>  			goto fail_free_domain;
> +		}
>
>  		geometry = &host->domain->geometry;
>
> @@ -198,6 +203,7 @@ static int host1x_probe(struct platform_device *pdev)
>  		host->iova_end = geometry->aperture_end;
>  	}
>
> +iommu_skip:
>  	err = host1x_channel_list_init(host);
>  	if (err) {
>  		dev_err(&pdev->dev, "failed to initialize channel list\n");
>
--
To unsubscribe from this list: send the line "unsubscribe linux-tegra" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Paul Kocialkowski July 11, 2017, 8:49 a.m.
On Mon, 2017-07-10 at 21:33 +0200, Paul Kocialkowski wrote:
> When there is no device to attach to the IOMMU domain, as may be the
> case when the device-tree does not contain the proper iommu node, it
> is
> best to keep going without IOMMU support rather than failing.
> This allows the driver to probe and function instead of taking down
> all of the tegra drm driver, leading to missing display support.

Fixes: 404bfb78daf3 ("gpu: host1x: Add IOMMU support")

> Signed-off-by: Paul Kocialkowski <contact@paulk.fr>
> ---
>  drivers/gpu/host1x/dev.c | 8 +++++++-
>  1 file changed, 7 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/gpu/host1x/dev.c b/drivers/gpu/host1x/dev.c
> index ac65f52850a6..f296738d0de8 100644
> --- a/drivers/gpu/host1x/dev.c
> +++ b/drivers/gpu/host1x/dev.c
> @@ -186,8 +186,13 @@ static int host1x_probe(struct platform_device
> *pdev)
>  			return -ENOMEM;
>  
>  		err = iommu_attach_device(host->domain, &pdev->dev);
> -		if (err)
> +		if (err == -ENODEV) {
> +			iommu_domain_free(host->domain);
> +			host->domain = NULL;
> +			goto iommu_skip;
> +		} else if (err) {
>  			goto fail_free_domain;
> +		}
>  
>  		geometry = &host->domain->geometry;
>  
> @@ -198,6 +203,7 @@ static int host1x_probe(struct platform_device
> *pdev)
>  		host->iova_end = geometry->aperture_end;
>  	}
>  
> +iommu_skip:
>  	err = host1x_channel_list_init(host);
>  	if (err) {
>  		dev_err(&pdev->dev, "failed to initialize channel
> list\n");
Marcel Ziswiler July 11, 2017, 2:37 p.m.
On Tue, 2017-07-11 at 11:49 +0300, Paul Kocialkowski wrote:
> On Mon, 2017-07-10 at 21:33 +0200, Paul Kocialkowski wrote:

> > When there is no device to attach to the IOMMU domain, as may be

> > the

> > case when the device-tree does not contain the proper iommu node,

> > it

> > is

> > best to keep going without IOMMU support rather than failing.

> > This allows the driver to probe and function instead of taking down

> > all of the tegra drm driver, leading to missing display support.

> 

> Fixes: 404bfb78daf3 ("gpu: host1x: Add IOMMU support")

> 

> > Signed-off-by: Paul Kocialkowski <contact-W9ppeneeCTY@public.gmane.

> > org>

> > ---

> >  drivers/gpu/host1x/dev.c | 8 +++++++-

> >  1 file changed, 7 insertions(+), 1 deletion(-)

> > 

> > diff --git a/drivers/gpu/host1x/dev.c b/drivers/gpu/host1x/dev.c

> > index ac65f52850a6..f296738d0de8 100644

> > --- a/drivers/gpu/host1x/dev.c

> > +++ b/drivers/gpu/host1x/dev.c

> > @@ -186,8 +186,13 @@ static int host1x_probe(struct platform_device

> > *pdev)

> >  			return -ENOMEM;

> >  

> >  		err = iommu_attach_device(host->domain, &pdev-

> > >dev);

> > -		if (err)

> > +		if (err == -ENODEV) {

> > +			iommu_domain_free(host->domain);

> > +			host->domain = NULL;

> > +			goto iommu_skip;

> > +		} else if (err) {

> >  			goto fail_free_domain;

> > +		}

> >  

> >  		geometry = &host->domain->geometry;

> >  

> > @@ -198,6 +203,7 @@ static int host1x_probe(struct platform_device

> > *pdev)

> >  		host->iova_end = geometry->aperture_end;

> >  	}

> >  

> > +iommu_skip:

> >  	err = host1x_channel_list_init(host);

> >  	if (err) {

> >  		dev_err(&pdev->dev, "failed to initialize channel

> > list\n");


Please note that this does no longer cleanly apply after Mikko's 'gpu:
host1x: Refactor channel allocation code' commit from June 15 already
applied to current -next. Other than that

Tested-by: Marcel Ziswiler <marcel.ziswiler@toradex.com>
Tested-on: Apalis TK1, Apalis T30, Beaver, Colibri T30, Jetson-TK1

Finally graphics working again, thanks guys!
Paul Kocialkowski July 11, 2017, 2:44 p.m.
On Tue, 2017-07-11 at 14:37 +0000, Marcel Ziswiler wrote:
> On Tue, 2017-07-11 at 11:49 +0300, Paul Kocialkowski wrote:
> > On Mon, 2017-07-10 at 21:33 +0200, Paul Kocialkowski wrote:
> > > When there is no device to attach to the IOMMU domain, as may be
> > > the
> > > case when the device-tree does not contain the proper iommu node,
> > > it
> > > is
> > > best to keep going without IOMMU support rather than failing.
> > > This allows the driver to probe and function instead of taking
> > > down
> > > all of the tegra drm driver, leading to missing display support.
> > 
> > Fixes: 404bfb78daf3 ("gpu: host1x: Add IOMMU support")
> > 
> > > Signed-off-by: Paul Kocialkowski <contact-W9ppeneeCTY@public.gmane
> > > .
> > > org>
> > > ---
> > >  drivers/gpu/host1x/dev.c | 8 +++++++-
> > >  1 file changed, 7 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/drivers/gpu/host1x/dev.c b/drivers/gpu/host1x/dev.c
> > > index ac65f52850a6..f296738d0de8 100644
> > > --- a/drivers/gpu/host1x/dev.c
> > > +++ b/drivers/gpu/host1x/dev.c
> > > @@ -186,8 +186,13 @@ static int host1x_probe(struct
> > > platform_device
> > > *pdev)
> > >  			return -ENOMEM;
> > >  
> > >  		err = iommu_attach_device(host->domain, &pdev-
> > > > dev);
> > > 
> > > -		if (err)
> > > +		if (err == -ENODEV) {
> > > +			iommu_domain_free(host->domain);
> > > +			host->domain = NULL;
> > > +			goto iommu_skip;
> > > +		} else if (err) {
> > >  			goto fail_free_domain;
> > > +		}
> > >  
> > >  		geometry = &host->domain->geometry;
> > >  
> > > @@ -198,6 +203,7 @@ static int host1x_probe(struct platform_device
> > > *pdev)
> > >  		host->iova_end = geometry->aperture_end;
> > >  	}
> > >  
> > > +iommu_skip:
> > >  	err = host1x_channel_list_init(host);
> > >  	if (err) {
> > >  		dev_err(&pdev->dev, "failed to initialize channel
> > > list\n");
> 
> Please note that this does no longer cleanly apply after Mikko's 'gpu:
> host1x: Refactor channel allocation code' commit from June 15 already
> applied to current -next. Other than that

Good to know! I intended this patch for fixing 4.12 stable, so I'm not
even entirely sure it should land in next. The patch that fixes the
device-tree to actually enable the iommu should be merged in priority.

If that is the case and anyone wants this patch rebased to next, feel
free to let me know and I'll send a rebased version.

> Tested-by: Marcel Ziswiler <marcel.ziswiler@toradex.com>
> Tested-on: Apalis TK1, Apalis T30, Beaver, Colibri T30, Jetson-TK1
> 
> Finally graphics working again, thanks guys!

Yay! :)
Mikko Perttunen July 11, 2017, 5:09 p.m.
On 07/11/2017 05:44 PM, Paul Kocialkowski wrote:
> On Tue, 2017-07-11 at 14:37 +0000, Marcel Ziswiler wrote:
>> On Tue, 2017-07-11 at 11:49 +0300, Paul Kocialkowski wrote:
>>> On Mon, 2017-07-10 at 21:33 +0200, Paul Kocialkowski wrote:
>>>> When there is no device to attach to the IOMMU domain, as may be
>>>> the
>>>> case when the device-tree does not contain the proper iommu node,
>>>> it
>>>> is
>>>> best to keep going without IOMMU support rather than failing.
>>>> This allows the driver to probe and function instead of taking
>>>> down
>>>> all of the tegra drm driver, leading to missing display support.
>>>
>>> Fixes: 404bfb78daf3 ("gpu: host1x: Add IOMMU support")
>>>
>>>> Signed-off-by: Paul Kocialkowski <contact-W9ppeneeCTY@public.gmane
>>>> .
>>>> org>
>>>> ---
>>>>   drivers/gpu/host1x/dev.c | 8 +++++++-
>>>>   1 file changed, 7 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/drivers/gpu/host1x/dev.c b/drivers/gpu/host1x/dev.c
>>>> index ac65f52850a6..f296738d0de8 100644
>>>> --- a/drivers/gpu/host1x/dev.c
>>>> +++ b/drivers/gpu/host1x/dev.c
>>>> @@ -186,8 +186,13 @@ static int host1x_probe(struct
>>>> platform_device
>>>> *pdev)
>>>>   			return -ENOMEM;
>>>>   
>>>>   		err = iommu_attach_device(host->domain, &pdev-
>>>>> dev);
>>>>
>>>> -		if (err)
>>>> +		if (err == -ENODEV) {
>>>> +			iommu_domain_free(host->domain);
>>>> +			host->domain = NULL;
>>>> +			goto iommu_skip;
>>>> +		} else if (err) {
>>>>   			goto fail_free_domain;
>>>> +		}
>>>>   
>>>>   		geometry = &host->domain->geometry;
>>>>   
>>>> @@ -198,6 +203,7 @@ static int host1x_probe(struct platform_device
>>>> *pdev)
>>>>   		host->iova_end = geometry->aperture_end;
>>>>   	}
>>>>   
>>>> +iommu_skip:
>>>>   	err = host1x_channel_list_init(host);
>>>>   	if (err) {
>>>>   		dev_err(&pdev->dev, "failed to initialize channel
>>>> list\n");
>>
>> Please note that this does no longer cleanly apply after Mikko's 'gpu:
>> host1x: Refactor channel allocation code' commit from June 15 already
>> applied to current -next. Other than that
> 
> Good to know! I intended this patch for fixing 4.12 stable, so I'm not
> even entirely sure it should land in next. The patch that fixes the
> device-tree to actually enable the iommu should be merged in priority.
> 
> If that is the case and anyone wants this patch rebased to next, feel
> free to let me know and I'll send a rebased version.

We should definitely have this for non-stable kernels too - it would be 
a bit strange if the driver worked if IOMMU support wasn't built into 
the kernel but failed if it was :) It can also be helpful for debugging 
to be able to disable IOMMU just for host1x.

> 
>> Tested-by: Marcel Ziswiler <marcel.ziswiler@toradex.com>
>> Tested-on: Apalis TK1, Apalis T30, Beaver, Colibri T30, Jetson-TK1
>>
>> Finally graphics working again, thanks guys!
> 
> Yay! :)
> 
--
To unsubscribe from this list: send the line "unsubscribe linux-tegra" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Jon Hunter July 18, 2017, 9:09 a.m.
Hi Mikko,

On 11/07/17 07:43, Mikko Perttunen wrote:
> Thanks for the patch, didn't consider this case. I really need to get
> together some system to automatically test on multiple platforms.. :)
We already have the infrastructure in place to do this, however, at the
moment we are just making sure the platforms boot. Something like this
we should be able to catch, even if we don't explicitly test display/hdmi.

I was thinking that we could parse the /dev and /sys file-systems to
ensure that devices for a given platform are present after boot. I am
not sure if there is a good way to do this, but if you have any
thoughts, then I am willing to add more testing.

What is interesting about this case, was that there were no specific
error messages that indicated this had broken AFAICT.

Cheers
Jon

Patch hide | download patch | download mbox

diff --git a/drivers/gpu/host1x/dev.c b/drivers/gpu/host1x/dev.c
index ac65f52850a6..f296738d0de8 100644
--- a/drivers/gpu/host1x/dev.c
+++ b/drivers/gpu/host1x/dev.c
@@ -186,8 +186,13 @@  static int host1x_probe(struct platform_device *pdev)
 			return -ENOMEM;
 
 		err = iommu_attach_device(host->domain, &pdev->dev);
-		if (err)
+		if (err == -ENODEV) {
+			iommu_domain_free(host->domain);
+			host->domain = NULL;
+			goto iommu_skip;
+		} else if (err) {
 			goto fail_free_domain;
+		}
 
 		geometry = &host->domain->geometry;
 
@@ -198,6 +203,7 @@  static int host1x_probe(struct platform_device *pdev)
 		host->iova_end = geometry->aperture_end;
 	}
 
+iommu_skip:
 	err = host1x_channel_list_init(host);
 	if (err) {
 		dev_err(&pdev->dev, "failed to initialize channel list\n");