diff mbox series

[v10,3/6] iommufd: Add iommufd_access_change_ioas(_id) helpers

Message ID ad75a1f7f0b4d5b6d35238b4ae7b41db1410110c.1690488745.git.nicolinc@nvidia.com (mailing list archive)
State New
Headers show
Series Add IO page table replacement support | expand

Commit Message

Nicolin Chen July 27, 2023, 8:24 p.m. UTC 
The complication of the mutex and refcount will be amplified after we
introduce the replace support for access. So, add a preparatory change
of a constitutive helper iommufd_access_change_ioas() and its wrapper
iommufd_access_change_ioas_id(). They can simply take care of existing
iommufd_access_attach() and iommufd_access_detach(), with a less risk
of race condition.

Also, update the unprotect routine in iommufd_access_destroy_object()
to calling the new iommufd_access_change_ioas() helper.

Suggested-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Nicolin Chen <nicolinc@nvidia.com>
---
 drivers/iommu/iommufd/device.c | 123 +++++++++++++++++++++------------
 1 file changed, 80 insertions(+), 43 deletions(-)

Comments

Tian, Kevin July 28, 2023, 4:23 a.m. UTC | #1 
> From: Nicolin Chen <nicolinc@nvidia.com>
> Sent: Friday, July 28, 2023 4:25 AM
> 
> +static int iommufd_access_change_ioas(struct iommufd_access *access,
> +				      struct iommufd_ioas *new_ioas)
> +{
> +	u32 iopt_access_list_id = access->iopt_access_list_id;
> +	struct iommufd_ioas *cur_ioas = access->ioas;
> +	int rc;
> +
> +	lockdep_assert_held(&access->ioas_lock);
> +
> +	/* We are racing with a concurrent detach, bail */
> +	if (cur_ioas != access->ioas_unpin)
> +		return -EBUSY;
> +
> +	if (IS_ERR(new_ioas))
> +		return PTR_ERR(new_ioas);

iommufd_access_change_ioas_id() already checks errors.

> +
>  void iommufd_access_destroy_object(struct iommufd_object *obj)
>  {
>  	struct iommufd_access *access =
>  		container_of(obj, struct iommufd_access, obj);
> 
> -	if (access->ioas) {
> -		iopt_remove_access(&access->ioas->iopt, access,
> -				   access->iopt_access_list_id);
> -		refcount_dec(&access->ioas->obj.users);
> -		access->ioas = NULL;
> -	}
> +	mutex_lock(&access->ioas_lock);
> +	if (access->ioas)
> +		WARN_ON(iommufd_access_change_ioas(access, NULL));
> +	mutex_unlock(&access->ioas_lock);
>  	iommufd_ctx_put(access->ictx);
>  }

this changes the behavior of destroy. Previously it always removes
the access w/o detecting race while now it will give up and throw
out a warning. While I'm fine with this change from bisec p.o.v.
it might be good to split this into a separate patch.

>  void iommufd_access_detach(struct iommufd_access *access)
>  {
> -	struct iommufd_ioas *cur_ioas = access->ioas;
> +	int rc;
> 
>  	mutex_lock(&access->ioas_lock);
> -	if (WARN_ON(!access->ioas))
> -		goto out;
> -	/*
> -	 * Set ioas to NULL to block any further iommufd_access_pin_pages().
> -	 * iommufd_access_unpin_pages() can continue using access-
> >ioas_unpin.
> -	 */
> -	access->ioas = NULL;
> -
> -	if (access->ops->unmap) {
> +	if (WARN_ON(!access->ioas)) {
>  		mutex_unlock(&access->ioas_lock);
> -		access->ops->unmap(access->data, 0, ULONG_MAX);
> -		mutex_lock(&access->ioas_lock);
> +		return;
>  	}
> -	iopt_remove_access(&cur_ioas->iopt, access,
> -			   access->iopt_access_list_id);
> -	refcount_dec(&cur_ioas->obj.users);
> -out:
> -	access->ioas_unpin = NULL;
> +	rc = iommufd_access_change_ioas(access, NULL);
> +	WARN_ON(rc);

'rc' can be removed. 

Just "WARN_ON(iommufd_access_change_ioas(access, NULL));"

otherwise looks good to me,

Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Nicolin Chen July 28, 2023, 4:36 a.m. UTC | #2 
On Fri, Jul 28, 2023 at 04:23:03AM +0000, Tian, Kevin wrote:
> > From: Nicolin Chen <nicolinc@nvidia.com>
> > Sent: Friday, July 28, 2023 4:25 AM
> >
> > +static int iommufd_access_change_ioas(struct iommufd_access *access,
> > +                                   struct iommufd_ioas *new_ioas)
> > +{
> > +     u32 iopt_access_list_id = access->iopt_access_list_id;
> > +     struct iommufd_ioas *cur_ioas = access->ioas;
> > +     int rc;
> > +
> > +     lockdep_assert_held(&access->ioas_lock);
> > +
> > +     /* We are racing with a concurrent detach, bail */
> > +     if (cur_ioas != access->ioas_unpin)
> > +             return -EBUSY;
> > +
> > +     if (IS_ERR(new_ioas))
> > +             return PTR_ERR(new_ioas);
> 
> iommufd_access_change_ioas_id() already checks errors.

I've thought about that: given that iommufd_access_change_ioas
is a standalone API, though it's not used anywhere else at the
moment, it might be safer to have this check again. Otherwise,
we would need a line of comments saying that "caller must make
sure that the input new_ioas is not holding an error code" or
so?

> > +
> >  void iommufd_access_destroy_object(struct iommufd_object *obj)
> >  {
> >       struct iommufd_access *access =
> >               container_of(obj, struct iommufd_access, obj);
> >
> > -     if (access->ioas) {
> > -             iopt_remove_access(&access->ioas->iopt, access,
> > -                                access->iopt_access_list_id);
> > -             refcount_dec(&access->ioas->obj.users);
> > -             access->ioas = NULL;
> > -     }
> > +     mutex_lock(&access->ioas_lock);
> > +     if (access->ioas)
> > +             WARN_ON(iommufd_access_change_ioas(access, NULL));
> > +     mutex_unlock(&access->ioas_lock);
> >       iommufd_ctx_put(access->ictx);
> >  }
> 
> this changes the behavior of destroy. Previously it always removes
> the access w/o detecting race while now it will give up and throw
> out a warning.

You mean the -EBUSY case? That's a good catch..

> While I'm fine with this change from bisec p.o.v.
> it might be good to split this into a separate patch.

Yea, I can do that.

> >  void iommufd_access_detach(struct iommufd_access *access)
> >  {
> > -     struct iommufd_ioas *cur_ioas = access->ioas;
> > +     int rc;
> >
> >       mutex_lock(&access->ioas_lock);
> > -     if (WARN_ON(!access->ioas))
> > -             goto out;
> > -     /*
> > -      * Set ioas to NULL to block any further iommufd_access_pin_pages().
> > -      * iommufd_access_unpin_pages() can continue using access-
> > >ioas_unpin.
> > -      */
> > -     access->ioas = NULL;
> > -
> > -     if (access->ops->unmap) {
> > +     if (WARN_ON(!access->ioas)) {
> >               mutex_unlock(&access->ioas_lock);
> > -             access->ops->unmap(access->data, 0, ULONG_MAX);
> > -             mutex_lock(&access->ioas_lock);
> > +             return;
> >       }
> > -     iopt_remove_access(&cur_ioas->iopt, access,
> > -                        access->iopt_access_list_id);
> > -     refcount_dec(&cur_ioas->obj.users);
> > -out:
> > -     access->ioas_unpin = NULL;
> > +     rc = iommufd_access_change_ioas(access, NULL);
> > +     WARN_ON(rc);
> 
> 'rc' can be removed.
> 
> Just "WARN_ON(iommufd_access_change_ioas(access, NULL));"

Will do that in v11.

> otherwise looks good to me,
> 
> Reviewed-by: Kevin Tian <kevin.tian@intel.com>

Thanks!
Nic
Tian, Kevin July 28, 2023, 4:41 a.m. UTC | #3 
> From: Nicolin Chen <nicolinc@nvidia.com>
> Sent: Friday, July 28, 2023 12:37 PM
> 
> On Fri, Jul 28, 2023 at 04:23:03AM +0000, Tian, Kevin wrote:
> > > From: Nicolin Chen <nicolinc@nvidia.com>
> > > Sent: Friday, July 28, 2023 4:25 AM
> > >
> > > +static int iommufd_access_change_ioas(struct iommufd_access *access,
> > > +                                   struct iommufd_ioas *new_ioas)
> > > +{
> > > +     u32 iopt_access_list_id = access->iopt_access_list_id;
> > > +     struct iommufd_ioas *cur_ioas = access->ioas;
> > > +     int rc;
> > > +
> > > +     lockdep_assert_held(&access->ioas_lock);
> > > +
> > > +     /* We are racing with a concurrent detach, bail */
> > > +     if (cur_ioas != access->ioas_unpin)
> > > +             return -EBUSY;
> > > +
> > > +     if (IS_ERR(new_ioas))
> > > +             return PTR_ERR(new_ioas);
> >
> > iommufd_access_change_ioas_id() already checks errors.
> 
> I've thought about that: given that iommufd_access_change_ioas
> is a standalone API, though it's not used anywhere else at the
> moment, it might be safer to have this check again. Otherwise,
> we would need a line of comments saying that "caller must make
> sure that the input new_ioas is not holding an error code" or
> so?
> 

I don't think it's a common practice for the caller to pass in
an error pointer when it already knows it's an error...
Nicolin Chen July 28, 2023, 4:44 a.m. UTC | #4 
On Fri, Jul 28, 2023 at 04:41:18AM +0000, Tian, Kevin wrote:
> > From: Nicolin Chen <nicolinc@nvidia.com>
> > Sent: Friday, July 28, 2023 12:37 PM
> >
> > On Fri, Jul 28, 2023 at 04:23:03AM +0000, Tian, Kevin wrote:
> > > > From: Nicolin Chen <nicolinc@nvidia.com>
> > > > Sent: Friday, July 28, 2023 4:25 AM
> > > >
> > > > +static int iommufd_access_change_ioas(struct iommufd_access *access,
> > > > +                                   struct iommufd_ioas *new_ioas)
> > > > +{
> > > > +     u32 iopt_access_list_id = access->iopt_access_list_id;
> > > > +     struct iommufd_ioas *cur_ioas = access->ioas;
> > > > +     int rc;
> > > > +
> > > > +     lockdep_assert_held(&access->ioas_lock);
> > > > +
> > > > +     /* We are racing with a concurrent detach, bail */
> > > > +     if (cur_ioas != access->ioas_unpin)
> > > > +             return -EBUSY;
> > > > +
> > > > +     if (IS_ERR(new_ioas))
> > > > +             return PTR_ERR(new_ioas);
> > >
> > > iommufd_access_change_ioas_id() already checks errors.
> >
> > I've thought about that: given that iommufd_access_change_ioas
> > is a standalone API, though it's not used anywhere else at the
> > moment, it might be safer to have this check again. Otherwise,
> > we would need a line of comments saying that "caller must make
> > sure that the input new_ioas is not holding an error code" or
> > so?
> >
> 
> I don't think it's a common practice for the caller to pass in
> an error pointer when it already knows it's an error...

OK. I will just drop it then.
diff mbox series

Patch

diff --git a/drivers/iommu/iommufd/device.c b/drivers/iommu/iommufd/device.c
index 7a3e8660b902..e79cbedd8626 100644
--- a/drivers/iommu/iommufd/device.c
+++ b/drivers/iommu/iommufd/device.c
@@ -684,17 +684,82 @@  void iommufd_device_detach(struct iommufd_device *idev)
 }
 EXPORT_SYMBOL_NS_GPL(iommufd_device_detach, IOMMUFD);
 
+/*
+ * On success, it will refcount_inc() at a valid new_ioas and refcount_dec() at
+ * a valid cur_ioas (access->ioas). A caller passing in a valid new_ioas should
+ * call iommufd_put_object() if it does an iommufd_get_object() for a new_ioas.
+ */
+static int iommufd_access_change_ioas(struct iommufd_access *access,
+				      struct iommufd_ioas *new_ioas)
+{
+	u32 iopt_access_list_id = access->iopt_access_list_id;
+	struct iommufd_ioas *cur_ioas = access->ioas;
+	int rc;
+
+	lockdep_assert_held(&access->ioas_lock);
+
+	/* We are racing with a concurrent detach, bail */
+	if (cur_ioas != access->ioas_unpin)
+		return -EBUSY;
+
+	if (IS_ERR(new_ioas))
+		return PTR_ERR(new_ioas);
+
+	if (cur_ioas == new_ioas)
+		return 0;
+
+	/*
+	 * Set ioas to NULL to block any further iommufd_access_pin_pages().
+	 * iommufd_access_unpin_pages() can continue using access->ioas_unpin.
+	 */
+	access->ioas = NULL;
+
+	if (new_ioas) {
+		rc = iopt_add_access(&new_ioas->iopt, access);
+		if (rc) {
+			access->ioas = cur_ioas;
+			return rc;
+		}
+		refcount_inc(&new_ioas->obj.users);
+	}
+
+	if (cur_ioas) {
+		if (access->ops->unmap) {
+			mutex_unlock(&access->ioas_lock);
+			access->ops->unmap(access->data, 0, ULONG_MAX);
+			mutex_lock(&access->ioas_lock);
+		}
+		iopt_remove_access(&cur_ioas->iopt, access, iopt_access_list_id);
+		refcount_dec(&cur_ioas->obj.users);
+	}
+
+	access->ioas = new_ioas;
+	access->ioas_unpin = new_ioas;
+
+	return 0;
+}
+
+static int iommufd_access_change_ioas_id(struct iommufd_access *access, u32 id)
+{
+	struct iommufd_ioas *ioas = iommufd_get_ioas(access->ictx, id);
+	int rc;
+
+	if (IS_ERR(ioas))
+		return PTR_ERR(ioas);
+	rc = iommufd_access_change_ioas(access, ioas);
+	iommufd_put_object(&ioas->obj);
+	return rc;
+}
+
 void iommufd_access_destroy_object(struct iommufd_object *obj)
 {
 	struct iommufd_access *access =
 		container_of(obj, struct iommufd_access, obj);
 
-	if (access->ioas) {
-		iopt_remove_access(&access->ioas->iopt, access,
-				   access->iopt_access_list_id);
-		refcount_dec(&access->ioas->obj.users);
-		access->ioas = NULL;
-	}
+	mutex_lock(&access->ioas_lock);
+	if (access->ioas)
+		WARN_ON(iommufd_access_change_ioas(access, NULL));
+	mutex_unlock(&access->ioas_lock);
 	iommufd_ctx_put(access->ictx);
 }
 
@@ -761,60 +826,32 @@  EXPORT_SYMBOL_NS_GPL(iommufd_access_destroy, IOMMUFD);
 
 void iommufd_access_detach(struct iommufd_access *access)
 {
-	struct iommufd_ioas *cur_ioas = access->ioas;
+	int rc;
 
 	mutex_lock(&access->ioas_lock);
-	if (WARN_ON(!access->ioas))
-		goto out;
-	/*
-	 * Set ioas to NULL to block any further iommufd_access_pin_pages().
-	 * iommufd_access_unpin_pages() can continue using access->ioas_unpin.
-	 */
-	access->ioas = NULL;
-
-	if (access->ops->unmap) {
+	if (WARN_ON(!access->ioas)) {
 		mutex_unlock(&access->ioas_lock);
-		access->ops->unmap(access->data, 0, ULONG_MAX);
-		mutex_lock(&access->ioas_lock);
+		return;
 	}
-	iopt_remove_access(&cur_ioas->iopt, access,
-			   access->iopt_access_list_id);
-	refcount_dec(&cur_ioas->obj.users);
-out:
-	access->ioas_unpin = NULL;
+	rc = iommufd_access_change_ioas(access, NULL);
+	WARN_ON(rc);
 	mutex_unlock(&access->ioas_lock);
 }
 EXPORT_SYMBOL_NS_GPL(iommufd_access_detach, IOMMUFD);
 
 int iommufd_access_attach(struct iommufd_access *access, u32 ioas_id)
 {
-	struct iommufd_ioas *new_ioas;
-	int rc = 0;
+	int rc;
 
 	mutex_lock(&access->ioas_lock);
-	if (WARN_ON(access->ioas || access->ioas_unpin)) {
+	if (WARN_ON(access->ioas)) {
 		mutex_unlock(&access->ioas_lock);
 		return -EINVAL;
 	}
 
-	new_ioas = iommufd_get_ioas(access->ictx, ioas_id);
-	if (IS_ERR(new_ioas)) {
-		mutex_unlock(&access->ioas_lock);
-		return PTR_ERR(new_ioas);
-	}
-
-	rc = iopt_add_access(&new_ioas->iopt, access);
-	if (rc) {
-		mutex_unlock(&access->ioas_lock);
-		iommufd_put_object(&new_ioas->obj);
-		return rc;
-	}
-	iommufd_ref_to_users(&new_ioas->obj);
-
-	access->ioas = new_ioas;
-	access->ioas_unpin = new_ioas;
+	rc = iommufd_access_change_ioas_id(access, ioas_id);
 	mutex_unlock(&access->ioas_lock);
-	return 0;
+	return rc;
 }
 EXPORT_SYMBOL_NS_GPL(iommufd_access_attach, IOMMUFD);