[3/3] drm/amdgpu: wire up the can_remove() callback

Message ID 20240202222603.141240-3-hamza.mahfooz@amd.com
State New
Headers
Series [1/3] driver core: bus: introduce can_remove() |

Commit Message

Hamza Mahfooz Feb. 2, 2024, 10:25 p.m. UTC
  Removing an amdgpu device that still has user space references allocated
to it causes undefined behaviour. So, implement amdgpu_pci_can_remove()
and disallow devices that still have files allocated to them from being
unbound.

Cc: stable@vger.kernel.org
Signed-off-by: Hamza Mahfooz <hamza.mahfooz@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 17 +++++++++++++++++
 1 file changed, 17 insertions(+)
  

Comments

Bjorn Helgaas Feb. 2, 2024, 10:41 p.m. UTC | #1
[+cc Bartosz]

On Fri, Feb 02, 2024 at 05:25:56PM -0500, Hamza Mahfooz wrote:
> Removing an amdgpu device that still has user space references allocated
> to it causes undefined behaviour. So, implement amdgpu_pci_can_remove()
> and disallow devices that still have files allocated to them from being
> unbound.

Maybe this would help for things that are completely built-in or
soldered down, but nothing can prevent a user from physically pulling
a card or cable, so I don't think this is a generic solution to the
problem of dangling user space references.

Maybe Bartosz's recent LPC talk is relevant:
https://lpc.events/event/17/contributions/1627/

> Cc: stable@vger.kernel.org
> Signed-off-by: Hamza Mahfooz <hamza.mahfooz@amd.com>
> ---
>  drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 17 +++++++++++++++++
>  1 file changed, 17 insertions(+)
> 
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
> index cc69005f5b46..cfa64f3c5be5 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
> @@ -2323,6 +2323,22 @@ static int amdgpu_pci_probe(struct pci_dev *pdev,
>  	return ret;
>  }
>  
> +static bool amdgpu_pci_can_remove(struct pci_dev *pdev)
> +{
> +	struct drm_device *dev = pci_get_drvdata(pdev);
> +
> +	mutex_lock(&dev->filelist_mutex);
> +
> +	if (!list_empty(&dev->filelist)) {
> +		mutex_unlock(&dev->filelist_mutex);
> +		return false;
> +	}
> +
> +	mutex_unlock(&dev->filelist_mutex);
> +
> +	return true;
> +}
> +
>  static void
>  amdgpu_pci_remove(struct pci_dev *pdev)
>  {
> @@ -2929,6 +2945,7 @@ static struct pci_driver amdgpu_kms_pci_driver = {
>  	.name = DRIVER_NAME,
>  	.id_table = pciidlist,
>  	.probe = amdgpu_pci_probe,
> +	.can_remove = amdgpu_pci_can_remove,
>  	.remove = amdgpu_pci_remove,
>  	.shutdown = amdgpu_pci_shutdown,
>  	.driver.pm = &amdgpu_pm_ops,
> -- 
> 2.43.0
>
  
Greg KH Feb. 2, 2024, 11:40 p.m. UTC | #2
On Fri, Feb 02, 2024 at 05:25:56PM -0500, Hamza Mahfooz wrote:
> Removing an amdgpu device that still has user space references allocated
> to it causes undefined behaviour.

Then fix that please.  There should not be anything special about your
hardware that all of the tens of thousands of other devices can't handle
today.

What happens when I yank your device out of a system with a pci hotplug
bus?  You can't prevent that either, so this should not be any different
at all.

sorry, but please, just fix your driver.

greg k-h
  
Greg KH Feb. 2, 2024, 11:41 p.m. UTC | #3
On Fri, Feb 02, 2024 at 05:25:56PM -0500, Hamza Mahfooz wrote:
> Removing an amdgpu device that still has user space references allocated
> to it causes undefined behaviour. So, implement amdgpu_pci_can_remove()
> and disallow devices that still have files allocated to them from being
> unbound.
> 
> Cc: stable@vger.kernel.org
> Signed-off-by: Hamza Mahfooz <hamza.mahfooz@amd.com>
> ---
>  drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 17 +++++++++++++++++
>  1 file changed, 17 insertions(+)
> 
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
> index cc69005f5b46..cfa64f3c5be5 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
> @@ -2323,6 +2323,22 @@ static int amdgpu_pci_probe(struct pci_dev *pdev,
>  	return ret;
>  }
>  
> +static bool amdgpu_pci_can_remove(struct pci_dev *pdev)
> +{
> +	struct drm_device *dev = pci_get_drvdata(pdev);
> +
> +	mutex_lock(&dev->filelist_mutex);
> +
> +	if (!list_empty(&dev->filelist)) {
> +		mutex_unlock(&dev->filelist_mutex);
> +		return false;
> +	}
> +
> +	mutex_unlock(&dev->filelist_mutex);
> +
> +	return true;

Also, to be pedantic, this will not work as right after you returned
"true" here, userspace could open a file, causing the same issue you are
trying to prevent to have happen, happen.

So even if we wanted to do this, which again, we do not, this isn't even
a solution for it because it will still cause you problems.

greg k-h
  
Daniel Vetter Feb. 6, 2024, 2:29 p.m. UTC | #4
On Fri, Feb 02, 2024 at 03:40:03PM -0800, Greg Kroah-Hartman wrote:
> On Fri, Feb 02, 2024 at 05:25:56PM -0500, Hamza Mahfooz wrote:
> > Removing an amdgpu device that still has user space references allocated
> > to it causes undefined behaviour.
> 
> Then fix that please.  There should not be anything special about your
> hardware that all of the tens of thousands of other devices can't handle
> today.
> 
> What happens when I yank your device out of a system with a pci hotplug
> bus?  You can't prevent that either, so this should not be any different
> at all.
> 
> sorry, but please, just fix your driver.

fwiw Christian König from amd already rejected this too, I have no idea
why this was submitted since the very elaborate plan I developed with a
bunch of amd folks was to fix the various lifetime lolz we still have in
drm. We unfortunately export the world of internal objects to userspace as
uabi objects with dma_buf, dma_fence and everything else, but it's all
fixable and we have the plan even documented:

https://dri.freedesktop.org/docs/drm/gpu/drm-uapi.html#device-hot-unplug

So yeah anything that isn't that plan of record is very much no-go for drm
drivers. Unless we change that plan of course, but that needs a
documentation patch first and a big discussion.

Aside from an absolute massive pile of kernel-internal refcounting bugs
the really big one we agreed on after a lot of discussion is that SIGBUS
on dma-buf mmaps is no-go for drm drivers, because it would break way too
much userspace in ways which are simply not fixable (since sig handlers
are shared in a process, which means the gl/vk driver cannot use it).

Otherwise it's bog standard "fix the kernel bugs" work, just a lot of it.

Cheers, Sima
  
Christian König Feb. 6, 2024, 6:42 p.m. UTC | #5
Am 06.02.24 um 15:29 schrieb Daniel Vetter:
> On Fri, Feb 02, 2024 at 03:40:03PM -0800, Greg Kroah-Hartman wrote:
>> On Fri, Feb 02, 2024 at 05:25:56PM -0500, Hamza Mahfooz wrote:
>>> Removing an amdgpu device that still has user space references allocated
>>> to it causes undefined behaviour.
>> Then fix that please.  There should not be anything special about your
>> hardware that all of the tens of thousands of other devices can't handle
>> today.
>>
>> What happens when I yank your device out of a system with a pci hotplug
>> bus?  You can't prevent that either, so this should not be any different
>> at all.
>>
>> sorry, but please, just fix your driver.
> fwiw Christian König from amd already rejected this too, I have no idea
> why this was submitted

Well that was my fault.

I commented on an internal bug tracker that when sysfs bind/undbind is a 
different code path from PCI remove/re-scan we could try to reject it.

Turned out it isn't a different code path.

>   since the very elaborate plan I developed with a
> bunch of amd folks was to fix the various lifetime lolz we still have in
> drm. We unfortunately export the world of internal objects to userspace as
> uabi objects with dma_buf, dma_fence and everything else, but it's all
> fixable and we have the plan even documented:
>
> https://dri.freedesktop.org/docs/drm/gpu/drm-uapi.html#device-hot-unplug
>
> So yeah anything that isn't that plan of record is very much no-go for drm
> drivers. Unless we change that plan of course, but that needs a
> documentation patch first and a big discussion.
>
> Aside from an absolute massive pile of kernel-internal refcounting bugs
> the really big one we agreed on after a lot of discussion is that SIGBUS
> on dma-buf mmaps is no-go for drm drivers, because it would break way too
> much userspace in ways which are simply not fixable (since sig handlers
> are shared in a process, which means the gl/vk driver cannot use it).
>
> Otherwise it's bog standard "fix the kernel bugs" work, just a lot of it.

Ignoring a few memory leaks because of messed up refcounting we actually 
got that working quite nicely.

At least hot unplug / hot add seems to be working rather reliable in our 
internal testing.

So it can't be that messed up.

Regards,
Christian.

>
> Cheers, Sima
  
Daniel Vetter Feb. 9, 2024, 11 a.m. UTC | #6
On Tue, Feb 06, 2024 at 07:42:49PM +0100, Christian König wrote:
> Am 06.02.24 um 15:29 schrieb Daniel Vetter:
> > On Fri, Feb 02, 2024 at 03:40:03PM -0800, Greg Kroah-Hartman wrote:
> > > On Fri, Feb 02, 2024 at 05:25:56PM -0500, Hamza Mahfooz wrote:
> > > > Removing an amdgpu device that still has user space references allocated
> > > > to it causes undefined behaviour.
> > > Then fix that please.  There should not be anything special about your
> > > hardware that all of the tens of thousands of other devices can't handle
> > > today.
> > > 
> > > What happens when I yank your device out of a system with a pci hotplug
> > > bus?  You can't prevent that either, so this should not be any different
> > > at all.
> > > 
> > > sorry, but please, just fix your driver.
> > fwiw Christian König from amd already rejected this too, I have no idea
> > why this was submitted
> 
> Well that was my fault.
> 
> I commented on an internal bug tracker that when sysfs bind/undbind is a
> different code path from PCI remove/re-scan we could try to reject it.
> 
> Turned out it isn't a different code path.

Yeah it's exactly the same code, and removing the sysfs stuff means we
cant test hotunplug without physical hotunplugging stuff anymore. So
really not great - if one is buggy so is the other, and sysfs allows us to
control the timing a lot better to hit specific issues.
-Sima

> >   since the very elaborate plan I developed with a
> > bunch of amd folks was to fix the various lifetime lolz we still have in
> > drm. We unfortunately export the world of internal objects to userspace as
> > uabi objects with dma_buf, dma_fence and everything else, but it's all
> > fixable and we have the plan even documented:
> > 
> > https://dri.freedesktop.org/docs/drm/gpu/drm-uapi.html#device-hot-unplug
> > 
> > So yeah anything that isn't that plan of record is very much no-go for drm
> > drivers. Unless we change that plan of course, but that needs a
> > documentation patch first and a big discussion.
> > 
> > Aside from an absolute massive pile of kernel-internal refcounting bugs
> > the really big one we agreed on after a lot of discussion is that SIGBUS
> > on dma-buf mmaps is no-go for drm drivers, because it would break way too
> > much userspace in ways which are simply not fixable (since sig handlers
> > are shared in a process, which means the gl/vk driver cannot use it).
> > 
> > Otherwise it's bog standard "fix the kernel bugs" work, just a lot of it.
> 
> Ignoring a few memory leaks because of messed up refcounting we actually got
> that working quite nicely.
> 
> At least hot unplug / hot add seems to be working rather reliable in our
> internal testing.
> 
> So it can't be that messed up.
> 
> Regards,
> Christian.
> 
> > 
> > Cheers, Sima
>
  

Patch

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
index cc69005f5b46..cfa64f3c5be5 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
@@ -2323,6 +2323,22 @@  static int amdgpu_pci_probe(struct pci_dev *pdev,
 	return ret;
 }
 
+static bool amdgpu_pci_can_remove(struct pci_dev *pdev)
+{
+	struct drm_device *dev = pci_get_drvdata(pdev);
+
+	mutex_lock(&dev->filelist_mutex);
+
+	if (!list_empty(&dev->filelist)) {
+		mutex_unlock(&dev->filelist_mutex);
+		return false;
+	}
+
+	mutex_unlock(&dev->filelist_mutex);
+
+	return true;
+}
+
 static void
 amdgpu_pci_remove(struct pci_dev *pdev)
 {
@@ -2929,6 +2945,7 @@  static struct pci_driver amdgpu_kms_pci_driver = {
 	.name = DRIVER_NAME,
 	.id_table = pciidlist,
 	.probe = amdgpu_pci_probe,
+	.can_remove = amdgpu_pci_can_remove,
 	.remove = amdgpu_pci_remove,
 	.shutdown = amdgpu_pci_shutdown,
 	.driver.pm = &amdgpu_pm_ops,