[v2] vfio: fix deadlock between group lock and kvm lock

  After 51cdc8bc120e, we have another deadlock scenario between the
kvm->lock and the vfio group_lock with two different codepaths acquiring
the locks in different order.  Specifically in vfio_open_device, vfio
holds the vfio group_lock when issuing device->ops->open_device but some
drivers (like vfio-ap) need to acquire kvm->lock during their open_device
routine;  Meanwhile, kvm_vfio_release will acquire the kvm->lock first
before calling vfio_file_set_kvm which will acquire the vfio group_lock.

To resolve this, let's remove the need for the vfio group_lock from the
kvm_vfio_release codepath.  This is done by introducing a new spinlock to
protect modifications to the vfio group kvm pointer, and acquiring a kvm
ref from within vfio while holding this spinlock, with the reference held
until the last close for the device in question.

Fixes: 51cdc8bc120e ("kvm/vfio: Fix potential deadlock on vfio group_lock")
Reported-by: Anthony Krowiak <akrowiak@linux.ibm.com>
Suggested-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Matthew Rosato <mjrosato@linux.ibm.com>
---
Changes from v1:
* use spin_lock instead of spin_lock_irqsave (Jason)
* clear device->kvm_put as part of vfio_kvm_put_kvm (Yi)
* Re-arrange code to avoid referencing the group contents from within
  vfio_main (Kevin) which meant moving most of the code in this patch 
  to group.c along with getting/dropping of the dev_set lock
---
 drivers/vfio/group.c     | 90 +++++++++++++++++++++++++++++++++++++---
 drivers/vfio/vfio.h      |  1 +
 drivers/vfio/vfio_main.c | 11 ++---
 include/linux/vfio.h     |  2 +-
 4 files changed, 91 insertions(+), 13 deletions(-)
  

On Wed,  1 Feb 2023 14:20:10 -0500
Matthew Rosato <mjrosato@linux.ibm.com> wrote:

> After 51cdc8bc120e, we have another deadlock scenario between the
> kvm->lock and the vfio group_lock with two different codepaths acquiring
> the locks in different order.  Specifically in vfio_open_device, vfio
> holds the vfio group_lock when issuing device->ops->open_device but some
> drivers (like vfio-ap) need to acquire kvm->lock during their open_device
> routine;  Meanwhile, kvm_vfio_release will acquire the kvm->lock first
> before calling vfio_file_set_kvm which will acquire the vfio group_lock.
> 
> To resolve this, let's remove the need for the vfio group_lock from the
> kvm_vfio_release codepath.  This is done by introducing a new spinlock to
> protect modifications to the vfio group kvm pointer, and acquiring a kvm
> ref from within vfio while holding this spinlock, with the reference held
> until the last close for the device in question.
> 
> Fixes: 51cdc8bc120e ("kvm/vfio: Fix potential deadlock on vfio group_lock")
> Reported-by: Anthony Krowiak <akrowiak@linux.ibm.com>
> Suggested-by: Jason Gunthorpe <jgg@nvidia.com>
> Signed-off-by: Matthew Rosato <mjrosato@linux.ibm.com>
> ---
> Changes from v1:
> * use spin_lock instead of spin_lock_irqsave (Jason)
> * clear device->kvm_put as part of vfio_kvm_put_kvm (Yi)
> * Re-arrange code to avoid referencing the group contents from within
>   vfio_main (Kevin) which meant moving most of the code in this patch 
>   to group.c along with getting/dropping of the dev_set lock
> ---
>  drivers/vfio/group.c     | 90 +++++++++++++++++++++++++++++++++++++---
>  drivers/vfio/vfio.h      |  1 +
>  drivers/vfio/vfio_main.c | 11 ++---
>  include/linux/vfio.h     |  2 +-
>  4 files changed, 91 insertions(+), 13 deletions(-)
> 
> diff --git a/drivers/vfio/group.c b/drivers/vfio/group.c
> index bb24b2f0271e..52f434861294 100644
> --- a/drivers/vfio/group.c
> +++ b/drivers/vfio/group.c
> @@ -13,6 +13,9 @@
>  #include <linux/vfio.h>
>  #include <linux/iommufd.h>
>  #include <linux/anon_inodes.h>
> +#ifdef CONFIG_HAVE_KVM
> +#include <linux/kvm_host.h>
> +#endif
>  #include "vfio.h"
>  
>  static struct vfio {
> @@ -154,6 +157,55 @@ static int vfio_group_ioctl_set_container(struct vfio_group *group,
>  	return ret;
>  }
>  
> +#ifdef CONFIG_HAVE_KVM
> +static bool vfio_kvm_get_kvm_safe(struct vfio_device *device, struct kvm *kvm)

I'm tempted to name these vfio_device_get_kvm_safe() and only pass the
vfio_device, where of course we can get the kvm pointer from the group
internally.

> +{
> +	void (*pfn)(struct kvm *kvm);
> +	bool (*fn)(struct kvm *kvm);
> +	bool ret;
> +

We should assert_lockdep_held(&device->dev_set->lock) in both of these
since that seems to be what's protecting device->kvm and
device->put_kvm.

If we change as above to get the kvm pointer from the group within this
function, we can also move the kvm_ref_lock here, which seems to
simplify the caller quite a bit.

> +	pfn = symbol_get(kvm_put_kvm);
> +	if (WARN_ON(!pfn))
> +		return false;
> +
> +	fn = symbol_get(kvm_get_kvm_safe);
> +	if (WARN_ON(!fn)) {
> +		symbol_put(kvm_put_kvm);
> +		return false;
> +	}
> +
> +	ret = fn(kvm);
> +	if (ret)
> +		device->put_kvm = pfn;
> +	else
> +		symbol_put(kvm_put_kvm);
> +
> +	symbol_put(kvm_get_kvm_safe);
> +
> +	return ret;
> +}
> +
> +static void vfio_kvm_put_kvm(struct vfio_device *device)
> +{
> +	if (WARN_ON(!device->kvm || !device->put_kvm))
> +		return;

It simplifies the caller if we can use this even in the !device->kvm
case.

> +
> +	device->put_kvm(device->kvm);
> +	device->put_kvm = NULL;
> +	symbol_put(kvm_put_kvm);
> +}
> +
> +#else
> +static bool vfio_kvm_get_kvm_safe(struct vfio_device *device, struct kvm *kvm)
> +{
> +	return false;
> +}
> +
> +static void vfio_kvm_put_kvm(struct vfio_device *device)
> +{
> +}
> +#endif
> +
>  static int vfio_device_group_open(struct vfio_device *device)
>  {
>  	int ret;
> @@ -164,14 +216,32 @@ static int vfio_device_group_open(struct vfio_device *device)
>  		goto out_unlock;
>  	}
>  
> +	mutex_lock(&device->dev_set->lock);
> +
>  	/*
> -	 * Here we pass the KVM pointer with the group under the lock.  If the
> -	 * device driver will use it, it must obtain a reference and release it
> -	 * during close_device.
> +	 * Before the first device open, get the KVM pointer currently
> +	 * associated with the group (if there is one) and obtain a reference
> +	 * now that will be held until the open_count reaches 0 again.  Save
> +	 * the pointer in the device for use by drivers.
>  	 */
> +	if (device->open_count == 0) {
> +		spin_lock(&device->group->kvm_ref_lock);
> +		if (device->group->kvm &&
> +		    vfio_kvm_get_kvm_safe(device, device->group->kvm))
> +			device->kvm = device->group->kvm;
> +		spin_unlock(&device->group->kvm_ref_lock);
> +	}
> +
>  	ret = vfio_device_open(device, device->group->iommufd,
>  			       device->group->kvm);

We're using device->group->kvm outside of kvm_ref_lock here, it should
be using device->kvm.

>  
> +	if (ret && device->kvm && device->open_count == 0) {

Slightly redundant, if device->open_count == 0 here, we can infer ret
is non-zero.

I fiddled with it a little further, see if you like anything from the
version below and incorporate what you do.  Thanks,

Alex

diff --git a/drivers/vfio/group.c b/drivers/vfio/group.c
index bb24b2f0271e..5121a34e1489 100644
--- a/drivers/vfio/group.c
+++ b/drivers/vfio/group.c
@@ -13,6 +13,9 @@
 #include <linux/vfio.h>
 #include <linux/iommufd.h>
 #include <linux/anon_inodes.h>
+#ifdef CONFIG_HAVE_KVM
+#include <linux/kvm_host.h>
+#endif
 #include "vfio.h"
 
 static struct vfio {
@@ -154,6 +157,64 @@ static int vfio_group_ioctl_set_container(struct vfio_group *group,
 	return ret;
 }
 
+#ifdef CONFIG_HAVE_KVM
+static void vfio_device_get_kvm_safe(struct vfio_device *device)
+{
+	void (*pfn)(struct kvm *kvm);
+	bool (*fn)(struct kvm *kvm);
+	bool ret;
+
+	lockdep_assert_held(&device->dev_set->lock);
+
+	spin_lock(&device->group->kvm_ref_lock);
+	if (!device->group->kvm)
+		goto unlock;
+
+	pfn = symbol_get(kvm_put_kvm);
+	if (WARN_ON(!pfn))
+		goto unlock;
+
+	fn = symbol_get(kvm_get_kvm_safe);
+	if (WARN_ON(!fn)) {
+		symbol_put(kvm_put_kvm);
+		goto unlock;
+	}
+
+	ret = fn(device->group->kvm);
+	symbol_put(kvm_get_kvm_safe);
+	if (!ret) {
+		symbol_put(kvm_put_kvm);
+		goto unlock;
+	}
+
+	device->put_kvm = pfn;
+	device->kvm = device->group->kvm;
+unlock:
+	spin_unlock(&device->group->kvm_ref_lock);
+}
+
+static void vfio_device_put_kvm(struct vfio_device *device)
+{
+	lockdep_assert_held(&device->dev_set->lock);
+
+	if (!device->kvm)
+		return;
+
+	if (WARN_ON(!device->put_kvm))
+		goto clear;
+
+	device->put_kvm(device->kvm);
+	device->put_kvm = NULL;
+	symbol_put(kvm_put_kvm);
+
+clear:
+	device->kvm = NULL;
+}
+#else
+static void vfio_device_get_kvm_safe(struct vfio_device *device) {}
+static void vfio_device_put_kvm(struct vfio_device *device) {}
+#endif
+
 static int vfio_device_group_open(struct vfio_device *device)
 {
 	int ret;
@@ -164,13 +225,23 @@ static int vfio_device_group_open(struct vfio_device *device)
 		goto out_unlock;
 	}
 
+	mutex_lock(&device->dev_set->lock);
+
 	/*
-	 * Here we pass the KVM pointer with the group under the lock.  If the
-	 * device driver will use it, it must obtain a reference and release it
-	 * during close_device.
+	 * Before the first device open, get the KVM pointer currently
+	 * associated with the group (if there is one) and obtain a reference
+	 * now that will be held until the open_count reaches 0 again.  Save
+	 * the pointer in the device for use by drivers.
 	 */
-	ret = vfio_device_open(device, device->group->iommufd,
-			       device->group->kvm);
+	if (device->open_count == 0)
+		vfio_device_get_kvm_safe(device);
+
+	ret = vfio_device_open(device, device->group->iommufd, device->kvm);
+
+	if (device->open_count == 0)
+		vfio_device_put_kvm(device);
+
+	mutex_unlock(&device->dev_set->lock);
 
 out_unlock:
 	mutex_unlock(&device->group->group_lock);
@@ -180,7 +251,14 @@ static int vfio_device_group_open(struct vfio_device *device)
 void vfio_device_group_close(struct vfio_device *device)
 {
 	mutex_lock(&device->group->group_lock);
+	mutex_lock(&device->dev_set->lock);
+
 	vfio_device_close(device, device->group->iommufd);
+
+	if (device->open_count == 0)
+		vfio_device_put_kvm(device);
+
+	mutex_unlock(&device->dev_set->lock);
 	mutex_unlock(&device->group->group_lock);
 }
 
@@ -450,6 +528,7 @@ static struct vfio_group *vfio_group_alloc(struct iommu_group *iommu_group,
 
 	refcount_set(&group->drivers, 1);
 	mutex_init(&group->group_lock);
+	spin_lock_init(&group->kvm_ref_lock);
 	INIT_LIST_HEAD(&group->device_list);
 	mutex_init(&group->device_lock);
 	group->iommu_group = iommu_group;
@@ -803,9 +882,9 @@ void vfio_file_set_kvm(struct file *file, struct kvm *kvm)
 	if (!vfio_file_is_group(file))
 		return;
 
-	mutex_lock(&group->group_lock);
+	spin_lock(&group->kvm_ref_lock);
 	group->kvm = kvm;
-	mutex_unlock(&group->group_lock);
+	spin_unlock(&group->kvm_ref_lock);
 }
 EXPORT_SYMBOL_GPL(vfio_file_set_kvm);
 
diff --git a/drivers/vfio/vfio.h b/drivers/vfio/vfio.h
index f8219a438bfb..20c6bc249cb8 100644
--- a/drivers/vfio/vfio.h
+++ b/drivers/vfio/vfio.h
@@ -74,6 +74,7 @@ struct vfio_group {
 	struct file			*opened_file;
 	struct blocking_notifier_head	notifier;
 	struct iommufd_ctx		*iommufd;
+	spinlock_t			kvm_ref_lock;
 };
 
 int vfio_device_set_group(struct vfio_device *device,
diff --git a/drivers/vfio/vfio_main.c b/drivers/vfio/vfio_main.c
index 5177bb061b17..14dbf781ea8c 100644
--- a/drivers/vfio/vfio_main.c
+++ b/drivers/vfio/vfio_main.c
@@ -361,7 +361,6 @@ static int vfio_device_first_open(struct vfio_device *device,
 	if (ret)
 		goto err_module_put;
 
-	device->kvm = kvm;
 	if (device->ops->open_device) {
 		ret = device->ops->open_device(device);
 		if (ret)
@@ -370,7 +369,6 @@ static int vfio_device_first_open(struct vfio_device *device,
 	return 0;
 
 err_unuse_iommu:
-	device->kvm = NULL;
 	if (iommufd)
 		vfio_iommufd_unbind(device);
 	else
@@ -387,7 +385,6 @@ static void vfio_device_last_close(struct vfio_device *device,
 
 	if (device->ops->close_device)
 		device->ops->close_device(device);
-	device->kvm = NULL;
 	if (iommufd)
 		vfio_iommufd_unbind(device);
 	else
@@ -400,14 +397,14 @@ int vfio_device_open(struct vfio_device *device,
 {
 	int ret = 0;
 
-	mutex_lock(&device->dev_set->lock);
+	lockdep_assert_held(&device->dev_set->lock);
+
 	device->open_count++;
 	if (device->open_count == 1) {
 		ret = vfio_device_first_open(device, iommufd, kvm);
 		if (ret)
 			device->open_count--;
 	}
-	mutex_unlock(&device->dev_set->lock);
 
 	return ret;
 }
@@ -415,12 +412,12 @@ int vfio_device_open(struct vfio_device *device,
 void vfio_device_close(struct vfio_device *device,
 		       struct iommufd_ctx *iommufd)
 {
-	mutex_lock(&device->dev_set->lock);
+	lockdep_assert_held(&device->dev_set->lock);
+
 	vfio_assert_device_open(device);
 	if (device->open_count == 1)
 		vfio_device_last_close(device, iommufd);
 	device->open_count--;
-	mutex_unlock(&device->dev_set->lock);
 }
 
 /*
diff --git a/include/linux/vfio.h b/include/linux/vfio.h
index 35be78e9ae57..87ff862ff555 100644
--- a/include/linux/vfio.h
+++ b/include/linux/vfio.h
@@ -46,7 +46,6 @@ struct vfio_device {
 	struct vfio_device_set *dev_set;
 	struct list_head dev_set_list;
 	unsigned int migration_flags;
-	/* Driver must reference the kvm during open_device or never touch it */
 	struct kvm *kvm;
 
 	/* Members below here are private, not for driver use */
@@ -58,6 +57,7 @@ struct vfio_device {
 	struct list_head group_next;
 	struct list_head iommu_entry;
 	struct iommufd_access *iommufd_access;
+	void (*put_kvm)(struct kvm *kvm);
 #if IS_ENABLED(CONFIG_IOMMUFD)
 	struct iommufd_device *iommufd_device;
 	struct iommufd_ctx *iommufd_ictx;
  

> From: Alex Williamson <alex.williamson@redhat.com>
> Sent: Thursday, February 2, 2023 7:28 AM
> 
> On Wed,  1 Feb 2023 14:20:10 -0500
> Matthew Rosato <mjrosato@linux.ibm.com> wrote:
> 
> > After 51cdc8bc120e, we have another deadlock scenario between the
> > kvm->lock and the vfio group_lock with two different codepaths acquiring
> > the locks in different order.  Specifically in vfio_open_device, vfio
> > holds the vfio group_lock when issuing device->ops->open_device but
> some
> > drivers (like vfio-ap) need to acquire kvm->lock during their open_device
> > routine;  Meanwhile, kvm_vfio_release will acquire the kvm->lock first
> > before calling vfio_file_set_kvm which will acquire the vfio group_lock.
> >
> > To resolve this, let's remove the need for the vfio group_lock from the
> > kvm_vfio_release codepath.  This is done by introducing a new spinlock to
> > protect modifications to the vfio group kvm pointer, and acquiring a kvm
> > ref from within vfio while holding this spinlock, with the reference held
> > until the last close for the device in question.
> >
> > Fixes: 51cdc8bc120e ("kvm/vfio: Fix potential deadlock on vfio group_lock")
> > Reported-by: Anthony Krowiak <akrowiak@linux.ibm.com>
> > Suggested-by: Jason Gunthorpe <jgg@nvidia.com>
> > Signed-off-by: Matthew Rosato <mjrosato@linux.ibm.com>
> > ---
> > Changes from v1:
> > * use spin_lock instead of spin_lock_irqsave (Jason)
> > * clear device->kvm_put as part of vfio_kvm_put_kvm (Yi)
> > * Re-arrange code to avoid referencing the group contents from within
> >   vfio_main (Kevin) which meant moving most of the code in this patch
> >   to group.c along with getting/dropping of the dev_set lock
> > ---
> >  drivers/vfio/group.c     | 90
> +++++++++++++++++++++++++++++++++++++---
> >  drivers/vfio/vfio.h      |  1 +
> >  drivers/vfio/vfio_main.c | 11 ++---
> >  include/linux/vfio.h     |  2 +-
> >  4 files changed, 91 insertions(+), 13 deletions(-)
> >
> > diff --git a/drivers/vfio/group.c b/drivers/vfio/group.c
> > index bb24b2f0271e..52f434861294 100644
> > --- a/drivers/vfio/group.c
> > +++ b/drivers/vfio/group.c
> > @@ -13,6 +13,9 @@
> >  #include <linux/vfio.h>
> >  #include <linux/iommufd.h>
> >  #include <linux/anon_inodes.h>
> > +#ifdef CONFIG_HAVE_KVM
> > +#include <linux/kvm_host.h>
> > +#endif
> >  #include "vfio.h"
> >
> >  static struct vfio {
> > @@ -154,6 +157,55 @@ static int vfio_group_ioctl_set_container(struct
> vfio_group *group,
> >  	return ret;
> >  }
> >
> > +#ifdef CONFIG_HAVE_KVM
> > +static bool vfio_kvm_get_kvm_safe(struct vfio_device *device, struct
> kvm *kvm)
> 
> I'm tempted to name these vfio_device_get_kvm_safe() and only pass the
> vfio_device, where of course we can get the kvm pointer from the group
> internally.
> 
> > +{
> > +	void (*pfn)(struct kvm *kvm);
> > +	bool (*fn)(struct kvm *kvm);
> > +	bool ret;
> > +
> 
> We should assert_lockdep_held(&device->dev_set->lock) in both of these
> since that seems to be what's protecting device->kvm and
> device->put_kvm.
> 
> If we change as above to get the kvm pointer from the group within this
> function, we can also move the kvm_ref_lock here, which seems to
> simplify the caller quite a bit.
> 
> > +	pfn = symbol_get(kvm_put_kvm);
> > +	if (WARN_ON(!pfn))
> > +		return false;
> > +
> > +	fn = symbol_get(kvm_get_kvm_safe);
> > +	if (WARN_ON(!fn)) {
> > +		symbol_put(kvm_put_kvm);
> > +		return false;
> > +	}
> > +
> > +	ret = fn(kvm);
> > +	if (ret)
> > +		device->put_kvm = pfn;
> > +	else
> > +		symbol_put(kvm_put_kvm);
> > +
> > +	symbol_put(kvm_get_kvm_safe);
> > +
> > +	return ret;
> > +}
> > +
> > +static void vfio_kvm_put_kvm(struct vfio_device *device)
> > +{
> > +	if (WARN_ON(!device->kvm || !device->put_kvm))
> > +		return;
> 
> It simplifies the caller if we can use this even in the !device->kvm
> case.
> 
> > +
> > +	device->put_kvm(device->kvm);
> > +	device->put_kvm = NULL;
> > +	symbol_put(kvm_put_kvm);
> > +}
> > +
> > +#else
> > +static bool vfio_kvm_get_kvm_safe(struct vfio_device *device, struct
> kvm *kvm)
> > +{
> > +	return false;
> > +}
> > +
> > +static void vfio_kvm_put_kvm(struct vfio_device *device)
> > +{
> > +}
> > +#endif
> > +
> >  static int vfio_device_group_open(struct vfio_device *device)
> >  {
> >  	int ret;
> > @@ -164,14 +216,32 @@ static int vfio_device_group_open(struct
> vfio_device *device)
> >  		goto out_unlock;
> >  	}
> >
> > +	mutex_lock(&device->dev_set->lock);
> > +
> >  	/*
> > -	 * Here we pass the KVM pointer with the group under the lock.  If
> the
> > -	 * device driver will use it, it must obtain a reference and release it
> > -	 * during close_device.
> > +	 * Before the first device open, get the KVM pointer currently
> > +	 * associated with the group (if there is one) and obtain a reference
> > +	 * now that will be held until the open_count reaches 0 again.  Save
> > +	 * the pointer in the device for use by drivers.
> >  	 */
> > +	if (device->open_count == 0) {
> > +		spin_lock(&device->group->kvm_ref_lock);
> > +		if (device->group->kvm &&
> > +		    vfio_kvm_get_kvm_safe(device, device->group->kvm))
> > +			device->kvm = device->group->kvm;
> > +		spin_unlock(&device->group->kvm_ref_lock);
> > +	}
> > +
> >  	ret = vfio_device_open(device, device->group->iommufd,
> >  			       device->group->kvm);
> 
> We're using device->group->kvm outside of kvm_ref_lock here, it should
> be using device->kvm.

Existing code set device->kvm in the vfio_device_first_open() which is
called by vfio_device_open(). After above change, seems not necessary
to pass kvm pointer into the call chain. Isn't it?

Regards,
Yi Liu
  

> From: Liu, Yi L <yi.l.liu@intel.com>
> Sent: Thursday, February 2, 2023 11:47 AM
> > >  	ret = vfio_device_open(device, device->group->iommufd,
> > >  			       device->group->kvm);
> >
> > We're using device->group->kvm outside of kvm_ref_lock here, it should
> > be using device->kvm.
> 
> Existing code set device->kvm in the vfio_device_first_open() which is
> called by vfio_device_open(). After above change, seems not necessary
> to pass kvm pointer into the call chain. Isn't it?
> 

Looks so.
  

> From: Alex Williamson <alex.williamson@redhat.com>
> Sent: Thursday, February 2, 2023 7:28 AM
> >
> > +#ifdef CONFIG_HAVE_KVM
> > +static bool vfio_kvm_get_kvm_safe(struct vfio_device *device, struct kvm
> *kvm)
> 
> I'm tempted to name these vfio_device_get_kvm_safe() and only pass the
> vfio_device, where of course we can get the kvm pointer from the group
> internally.
> 

I have a different thought. In the end the cdev series also need the similar
safe get/put logic then it's better to keep it in vfio_main.c called by
the group/cdev path individually.
  

On Thu, 2 Feb 2023 03:46:59 +0000
"Liu, Yi L" <yi.l.liu@intel.com> wrote:

> > From: Alex Williamson <alex.williamson@redhat.com>
> > Sent: Thursday, February 2, 2023 7:28 AM
> > 
> > On Wed,  1 Feb 2023 14:20:10 -0500
> > Matthew Rosato <mjrosato@linux.ibm.com> wrote:
> >   
> > > After 51cdc8bc120e, we have another deadlock scenario between the
> > > kvm->lock and the vfio group_lock with two different codepaths acquiring
> > > the locks in different order.  Specifically in vfio_open_device, vfio
> > > holds the vfio group_lock when issuing device->ops->open_device but  
> > some  
> > > drivers (like vfio-ap) need to acquire kvm->lock during their open_device
> > > routine;  Meanwhile, kvm_vfio_release will acquire the kvm->lock first
> > > before calling vfio_file_set_kvm which will acquire the vfio group_lock.
> > >
> > > To resolve this, let's remove the need for the vfio group_lock from the
> > > kvm_vfio_release codepath.  This is done by introducing a new spinlock to
> > > protect modifications to the vfio group kvm pointer, and acquiring a kvm
> > > ref from within vfio while holding this spinlock, with the reference held
> > > until the last close for the device in question.
> > >
> > > Fixes: 51cdc8bc120e ("kvm/vfio: Fix potential deadlock on vfio group_lock")
> > > Reported-by: Anthony Krowiak <akrowiak@linux.ibm.com>
> > > Suggested-by: Jason Gunthorpe <jgg@nvidia.com>
> > > Signed-off-by: Matthew Rosato <mjrosato@linux.ibm.com>
> > > ---
> > > Changes from v1:
> > > * use spin_lock instead of spin_lock_irqsave (Jason)
> > > * clear device->kvm_put as part of vfio_kvm_put_kvm (Yi)
> > > * Re-arrange code to avoid referencing the group contents from within
> > >   vfio_main (Kevin) which meant moving most of the code in this patch
> > >   to group.c along with getting/dropping of the dev_set lock
> > > ---
> > >  drivers/vfio/group.c     | 90  
> > +++++++++++++++++++++++++++++++++++++---  
> > >  drivers/vfio/vfio.h      |  1 +
> > >  drivers/vfio/vfio_main.c | 11 ++---
> > >  include/linux/vfio.h     |  2 +-
> > >  4 files changed, 91 insertions(+), 13 deletions(-)
> > >
> > > diff --git a/drivers/vfio/group.c b/drivers/vfio/group.c
> > > index bb24b2f0271e..52f434861294 100644
> > > --- a/drivers/vfio/group.c
> > > +++ b/drivers/vfio/group.c
> > > @@ -13,6 +13,9 @@
> > >  #include <linux/vfio.h>
> > >  #include <linux/iommufd.h>
> > >  #include <linux/anon_inodes.h>
> > > +#ifdef CONFIG_HAVE_KVM
> > > +#include <linux/kvm_host.h>
> > > +#endif
> > >  #include "vfio.h"
> > >
> > >  static struct vfio {
> > > @@ -154,6 +157,55 @@ static int vfio_group_ioctl_set_container(struct  
> > vfio_group *group,  
> > >  	return ret;
> > >  }
> > >
> > > +#ifdef CONFIG_HAVE_KVM
> > > +static bool vfio_kvm_get_kvm_safe(struct vfio_device *device, struct  
> > kvm *kvm)
> > 
> > I'm tempted to name these vfio_device_get_kvm_safe() and only pass the
> > vfio_device, where of course we can get the kvm pointer from the group
> > internally.
> >   
> > > +{
> > > +	void (*pfn)(struct kvm *kvm);
> > > +	bool (*fn)(struct kvm *kvm);
> > > +	bool ret;
> > > +  
> > 
> > We should assert_lockdep_held(&device->dev_set->lock) in both of these
> > since that seems to be what's protecting device->kvm and
> > device->put_kvm.
> > 
> > If we change as above to get the kvm pointer from the group within this
> > function, we can also move the kvm_ref_lock here, which seems to
> > simplify the caller quite a bit.
> >   
> > > +	pfn = symbol_get(kvm_put_kvm);
> > > +	if (WARN_ON(!pfn))
> > > +		return false;
> > > +
> > > +	fn = symbol_get(kvm_get_kvm_safe);
> > > +	if (WARN_ON(!fn)) {
> > > +		symbol_put(kvm_put_kvm);
> > > +		return false;
> > > +	}
> > > +
> > > +	ret = fn(kvm);
> > > +	if (ret)
> > > +		device->put_kvm = pfn;
> > > +	else
> > > +		symbol_put(kvm_put_kvm);
> > > +
> > > +	symbol_put(kvm_get_kvm_safe);
> > > +
> > > +	return ret;
> > > +}
> > > +
> > > +static void vfio_kvm_put_kvm(struct vfio_device *device)
> > > +{
> > > +	if (WARN_ON(!device->kvm || !device->put_kvm))
> > > +		return;  
> > 
> > It simplifies the caller if we can use this even in the !device->kvm
> > case.
> >   
> > > +
> > > +	device->put_kvm(device->kvm);
> > > +	device->put_kvm = NULL;
> > > +	symbol_put(kvm_put_kvm);
> > > +}
> > > +
> > > +#else
> > > +static bool vfio_kvm_get_kvm_safe(struct vfio_device *device, struct  
> > kvm *kvm)  
> > > +{
> > > +	return false;
> > > +}
> > > +
> > > +static void vfio_kvm_put_kvm(struct vfio_device *device)
> > > +{
> > > +}
> > > +#endif
> > > +
> > >  static int vfio_device_group_open(struct vfio_device *device)
> > >  {
> > >  	int ret;
> > > @@ -164,14 +216,32 @@ static int vfio_device_group_open(struct  
> > vfio_device *device)  
> > >  		goto out_unlock;
> > >  	}
> > >
> > > +	mutex_lock(&device->dev_set->lock);
> > > +
> > >  	/*
> > > -	 * Here we pass the KVM pointer with the group under the lock.  If  
> > the  
> > > -	 * device driver will use it, it must obtain a reference and release it
> > > -	 * during close_device.
> > > +	 * Before the first device open, get the KVM pointer currently
> > > +	 * associated with the group (if there is one) and obtain a reference
> > > +	 * now that will be held until the open_count reaches 0 again.  Save
> > > +	 * the pointer in the device for use by drivers.
> > >  	 */
> > > +	if (device->open_count == 0) {
> > > +		spin_lock(&device->group->kvm_ref_lock);
> > > +		if (device->group->kvm &&
> > > +		    vfio_kvm_get_kvm_safe(device, device->group->kvm))
> > > +			device->kvm = device->group->kvm;
> > > +		spin_unlock(&device->group->kvm_ref_lock);
> > > +	}
> > > +
> > >  	ret = vfio_device_open(device, device->group->iommufd,
> > >  			       device->group->kvm);  
> > 
> > We're using device->group->kvm outside of kvm_ref_lock here, it should
> > be using device->kvm.  
> 
> Existing code set device->kvm in the vfio_device_first_open() which is
> called by vfio_device_open(). After above change, seems not necessary
> to pass kvm pointer into the call chain. Isn't it?

Yes, we can get it from the device.  I didn't check how much this
bloats the patch though.  As a fix, it might make sense to save that
refactoring for a follow-on patch.  Thanks,

Alex
  

> -----Original Message-----
> From: Alex Williamson <alex.williamson@redhat.com>
> Sent: Thursday, February 2, 2023 12:15 PM
> To: Liu, Yi L <yi.l.liu@intel.com>
> Cc: Matthew Rosato <mjrosato@linux.ibm.com>; pbonzini@redhat.com;
> jgg@nvidia.com; cohuck@redhat.com; farman@linux.ibm.com;
> pmorel@linux.ibm.com; borntraeger@linux.ibm.com;
> frankja@linux.ibm.com; imbrenda@linux.ibm.com; david@redhat.com;
> akrowiak@linux.ibm.com; jjherne@linux.ibm.com; pasic@linux.ibm.com;
> zhenyuw@linux.intel.com; Wang, Zhi A <zhi.a.wang@intel.com>;
> Christopherson,, Sean <seanjc@google.com>; Tian, Kevin
> <kevin.tian@intel.com>; linux-s390@vger.kernel.org; kvm@vger.kernel.org;
> intel-gvt-dev@lists.freedesktop.org; intel-gfx@lists.freedesktop.org; linux-
> kernel@vger.kernel.org
> Subject: Re: [PATCH v2] vfio: fix deadlock between group lock and kvm lock
> 
> On Thu, 2 Feb 2023 03:46:59 +0000
> "Liu, Yi L" <yi.l.liu@intel.com> wrote:
> 
> > > From: Alex Williamson <alex.williamson@redhat.com>
> > > Sent: Thursday, February 2, 2023 7:28 AM
> > >
> > > On Wed,  1 Feb 2023 14:20:10 -0500
> > > Matthew Rosato <mjrosato@linux.ibm.com> wrote:
> > >
> > > > After 51cdc8bc120e, we have another deadlock scenario between the
> > > > kvm->lock and the vfio group_lock with two different codepaths
> acquiring
> > > > the locks in different order.  Specifically in vfio_open_device, vfio
> > > > holds the vfio group_lock when issuing device->ops->open_device but
> > > some
> > > > drivers (like vfio-ap) need to acquire kvm->lock during their
> open_device
> > > > routine;  Meanwhile, kvm_vfio_release will acquire the kvm->lock first
> > > > before calling vfio_file_set_kvm which will acquire the vfio group_lock.
> > > >
> > > > To resolve this, let's remove the need for the vfio group_lock from the
> > > > kvm_vfio_release codepath.  This is done by introducing a new
> spinlock to
> > > > protect modifications to the vfio group kvm pointer, and acquiring a
> kvm
> > > > ref from within vfio while holding this spinlock, with the reference held
> > > > until the last close for the device in question.
> > > >
> > > > Fixes: 51cdc8bc120e ("kvm/vfio: Fix potential deadlock on vfio
> group_lock")
> > > > Reported-by: Anthony Krowiak <akrowiak@linux.ibm.com>
> > > > Suggested-by: Jason Gunthorpe <jgg@nvidia.com>
> > > > Signed-off-by: Matthew Rosato <mjrosato@linux.ibm.com>
> > > > ---
> > > > Changes from v1:
> > > > * use spin_lock instead of spin_lock_irqsave (Jason)
> > > > * clear device->kvm_put as part of vfio_kvm_put_kvm (Yi)
> > > > * Re-arrange code to avoid referencing the group contents from
> within
> > > >   vfio_main (Kevin) which meant moving most of the code in this patch
> > > >   to group.c along with getting/dropping of the dev_set lock
> > > > ---
> > > >  drivers/vfio/group.c     | 90
> > > +++++++++++++++++++++++++++++++++++++---
> > > >  drivers/vfio/vfio.h      |  1 +
> > > >  drivers/vfio/vfio_main.c | 11 ++---
> > > >  include/linux/vfio.h     |  2 +-
> > > >  4 files changed, 91 insertions(+), 13 deletions(-)
> > > >
> > > > diff --git a/drivers/vfio/group.c b/drivers/vfio/group.c
> > > > index bb24b2f0271e..52f434861294 100644
> > > > --- a/drivers/vfio/group.c
> > > > +++ b/drivers/vfio/group.c
> > > > @@ -13,6 +13,9 @@
> > > >  #include <linux/vfio.h>
> > > >  #include <linux/iommufd.h>
> > > >  #include <linux/anon_inodes.h>
> > > > +#ifdef CONFIG_HAVE_KVM
> > > > +#include <linux/kvm_host.h>
> > > > +#endif
> > > >  #include "vfio.h"
> > > >
> > > >  static struct vfio {
> > > > @@ -154,6 +157,55 @@ static int
> vfio_group_ioctl_set_container(struct
> > > vfio_group *group,
> > > >  	return ret;
> > > >  }
> > > >
> > > > +#ifdef CONFIG_HAVE_KVM
> > > > +static bool vfio_kvm_get_kvm_safe(struct vfio_device *device,
> struct
> > > kvm *kvm)
> > >
> > > I'm tempted to name these vfio_device_get_kvm_safe() and only pass
> the
> > > vfio_device, where of course we can get the kvm pointer from the
> group
> > > internally.
> > >
> > > > +{
> > > > +	void (*pfn)(struct kvm *kvm);
> > > > +	bool (*fn)(struct kvm *kvm);
> > > > +	bool ret;
> > > > +
> > >
> > > We should assert_lockdep_held(&device->dev_set->lock) in both of
> these
> > > since that seems to be what's protecting device->kvm and
> > > device->put_kvm.
> > >
> > > If we change as above to get the kvm pointer from the group within this
> > > function, we can also move the kvm_ref_lock here, which seems to
> > > simplify the caller quite a bit.
> > >
> > > > +	pfn = symbol_get(kvm_put_kvm);
> > > > +	if (WARN_ON(!pfn))
> > > > +		return false;
> > > > +
> > > > +	fn = symbol_get(kvm_get_kvm_safe);
> > > > +	if (WARN_ON(!fn)) {
> > > > +		symbol_put(kvm_put_kvm);
> > > > +		return false;
> > > > +	}
> > > > +
> > > > +	ret = fn(kvm);
> > > > +	if (ret)
> > > > +		device->put_kvm = pfn;
> > > > +	else
> > > > +		symbol_put(kvm_put_kvm);
> > > > +
> > > > +	symbol_put(kvm_get_kvm_safe);
> > > > +
> > > > +	return ret;
> > > > +}
> > > > +
> > > > +static void vfio_kvm_put_kvm(struct vfio_device *device)
> > > > +{
> > > > +	if (WARN_ON(!device->kvm || !device->put_kvm))
> > > > +		return;
> > >
> > > It simplifies the caller if we can use this even in the !device->kvm
> > > case.
> > >
> > > > +
> > > > +	device->put_kvm(device->kvm);
> > > > +	device->put_kvm = NULL;
> > > > +	symbol_put(kvm_put_kvm);
> > > > +}
> > > > +
> > > > +#else
> > > > +static bool vfio_kvm_get_kvm_safe(struct vfio_device *device,
> struct
> > > kvm *kvm)
> > > > +{
> > > > +	return false;
> > > > +}
> > > > +
> > > > +static void vfio_kvm_put_kvm(struct vfio_device *device)
> > > > +{
> > > > +}
> > > > +#endif
> > > > +
> > > >  static int vfio_device_group_open(struct vfio_device *device)
> > > >  {
> > > >  	int ret;
> > > > @@ -164,14 +216,32 @@ static int vfio_device_group_open(struct
> > > vfio_device *device)
> > > >  		goto out_unlock;
> > > >  	}
> > > >
> > > > +	mutex_lock(&device->dev_set->lock);
> > > > +
> > > >  	/*
> > > > -	 * Here we pass the KVM pointer with the group under the lock.  If
> > > the
> > > > -	 * device driver will use it, it must obtain a reference and release it
> > > > -	 * during close_device.
> > > > +	 * Before the first device open, get the KVM pointer currently
> > > > +	 * associated with the group (if there is one) and obtain a reference
> > > > +	 * now that will be held until the open_count reaches 0 again.  Save
> > > > +	 * the pointer in the device for use by drivers.
> > > >  	 */
> > > > +	if (device->open_count == 0) {
> > > > +		spin_lock(&device->group->kvm_ref_lock);
> > > > +		if (device->group->kvm &&
> > > > +		    vfio_kvm_get_kvm_safe(device, device->group->kvm))
> > > > +			device->kvm = device->group->kvm;
> > > > +		spin_unlock(&device->group->kvm_ref_lock);
> > > > +	}
> > > > +
> > > >  	ret = vfio_device_open(device, device->group->iommufd,
> > > >  			       device->group->kvm);
> > >
> > > We're using device->group->kvm outside of kvm_ref_lock here, it
> should
> > > be using device->kvm.
> >
> > Existing code set device->kvm in the vfio_device_first_open() which is
> > called by vfio_device_open(). After above change, seems not necessary
> > to pass kvm pointer into the call chain. Isn't it?
> 
> Yes, we can get it from the device.  I didn't check how much this
> bloats the patch though.  As a fix, it might make sense to save that
> refactoring for a follow-on patch.  Thanks,

65 lines diff file. 😊 follow-on path works well.

diff --git a/drivers/vfio/group.c b/drivers/vfio/group.c
index bb24b2f0271e..9e04e55c838f 100644
--- a/drivers/vfio/group.c
+++ b/drivers/vfio/group.c
@@ -169,8 +169,7 @@ static int vfio_device_group_open(struct vfio_device *device)
 	 * device driver will use it, it must obtain a reference and release it
 	 * during close_device.
 	 */
-	ret = vfio_device_open(device, device->group->iommufd,
-			       device->group->kvm);
+	ret = vfio_device_open(device, device->group->iommufd);
 
 out_unlock:
 	mutex_unlock(&device->group->group_lock);
diff --git a/drivers/vfio/vfio.h b/drivers/vfio/vfio.h
index f8219a438bfb..4ece6cb4cf2e 100644
--- a/drivers/vfio/vfio.h
+++ b/drivers/vfio/vfio.h
@@ -19,7 +19,7 @@ struct vfio_container;
 void vfio_device_put_registration(struct vfio_device *device);
 bool vfio_device_try_get_registration(struct vfio_device *device);
 int vfio_device_open(struct vfio_device *device,
-		     struct iommufd_ctx *iommufd, struct kvm *kvm);
+		     struct iommufd_ctx *iommufd);
 void vfio_device_close(struct vfio_device *device,
 		       struct iommufd_ctx *iommufd);
 
diff --git a/drivers/vfio/vfio_main.c b/drivers/vfio/vfio_main.c
index 5177bb061b17..45a7d6d38e2e 100644
--- a/drivers/vfio/vfio_main.c
+++ b/drivers/vfio/vfio_main.c
@@ -345,7 +345,7 @@ static bool vfio_assert_device_open(struct vfio_device *device)
 }
 
 static int vfio_device_first_open(struct vfio_device *device,
-				  struct iommufd_ctx *iommufd, struct kvm *kvm)
+				  struct iommufd_ctx *iommufd)
 {
 	int ret;
 
@@ -361,7 +361,6 @@ static int vfio_device_first_open(struct vfio_device *device,
 	if (ret)
 		goto err_module_put;
 
-	device->kvm = kvm;
 	if (device->ops->open_device) {
 		ret = device->ops->open_device(device);
 		if (ret)
@@ -396,14 +395,14 @@ static void vfio_device_last_close(struct vfio_device *device,
 }
 
 int vfio_device_open(struct vfio_device *device,
-		     struct iommufd_ctx *iommufd, struct kvm *kvm)
+		     struct iommufd_ctx *iommufd)
 {
 	int ret = 0;
 
 	mutex_lock(&device->dev_set->lock);
 	device->open_count++;
 	if (device->open_count == 1) {
-		ret = vfio_device_first_open(device, iommufd, kvm);
+		ret = vfio_device_first_open(device, iommufd);
 		if (ret)
 			device->open_count--;
 	}
  

On 2/1/23 11:10 PM, Tian, Kevin wrote:
>> From: Alex Williamson <alex.williamson@redhat.com>
>> Sent: Thursday, February 2, 2023 7:28 AM
>>>
>>> +#ifdef CONFIG_HAVE_KVM
>>> +static bool vfio_kvm_get_kvm_safe(struct vfio_device *device, struct kvm
>> *kvm)
>>
>> I'm tempted to name these vfio_device_get_kvm_safe() and only pass the
>> vfio_device, where of course we can get the kvm pointer from the group
>> internally.
>>
> 
> I have a different thought. In the end the cdev series also need the similar
> safe get/put logic then it's better to keep it in vfio_main.c called by
> the group/cdev path individually.

Ah, I hadn't considered the cdev series - OK, I can move the functions back into vfio_main and externalize both via drivers/vfio/vfio.h so they can be called from group.c for this fix and then available to vfio_main.c already for cdev.