Skip to contentSkip to conversation

vhost-vdpa: mark vhost device invalid to reflect vdpa device unregistration

RFC PATCH In review [email protected] 3 participants 4 messages Jul 4, 2021 – Jul 5, 2021

Subsystems (5)
lore
Share & tools
lore
#0 lore raw

RFC PATCH 1/1 vhost-vdpa: mark vhost device invalid to reflect vdpa device unregistration

From: Gautam Dawar <[email protected]>

As mentioned in Bug 213179, any malicious user-space application can render a module registering a vDPA device to hang forever. This will typically surface when vdpa_device_unregister() is called from the function responsible for module unload, leading rmmod commands to not return.

This patch unblocks the caller module by continuing with the clean-up but after marking the vhost device as unavailable. For future requests from user-space application, the vhost device availability is checked first and if it has gone unavailable, such requests are denied.

---
 drivers/vhost/vdpa.c | 45 ++++++++++++++++++++++++++++++++++++++------
 1 file changed, 39 insertions(+), 6 deletions(-)
drivers/vhost/vdpa.c +39 -6
index e4b7d26649d8..623bc7f0c0ca 100644
@@ -47,6 +47,7 @@ struct vhost_vdpa {
 	int minor;
 	struct eventfd_ctx *config_ctx;
 	int in_batch;
+	int dev_invalid;
 	struct vdpa_iova_range range;
 };
 
@@ -61,6 +62,11 @@ static void handle_vq_kick(struct vhost_work *work)
 	struct vhost_vdpa *v = container_of(vq->dev, struct vhost_vdpa, vdev);
 	const struct vdpa_config_ops *ops = v->vdpa->config;
 
+	if (v->dev_invalid) {
+		dev_info(&v->dev,
+			 "%s: vhost_vdpa device unavailable\n", __func__);
+		return;
+	}
 	ops->kick_vq(v->vdpa, vq - v->vqs);
 }
 
@@ -120,6 +126,11 @@ static void vhost_vdpa_reset(struct vhost_vdpa *v)
 {
 	struct vdpa_device *vdpa = v->vdpa;
 
+	if (v->dev_invalid) {
+		dev_info(&v->dev,
+			 "%s: vhost_vdpa device unavailable\n", __func__);
+		return;
+	}
 	vdpa_reset(vdpa);
 	v->in_batch = 0;
 }
@@ -367,6 +378,11 @@ static long vhost_vdpa_vring_ioctl(struct vhost_vdpa *v, unsigned int cmd,
 	u32 idx;
 	long r;
 
+	if (v->dev_invalid) {
+		dev_info(&v->dev,
+			 "%s: vhost_vdpa device unavailable\n", __func__);
+		return -ENODEV;
+	}
 	r = get_user(idx, (u32 __user *)argp);
 	if (r < 0)
 		return r;
@@ -450,6 +466,11 @@ static long vhost_vdpa_unlocked_ioctl(struct file *filep,
 		return 0;
 	}
 
+	if (v->dev_invalid) {
+		dev_info(&v->dev,
+			 "%s: vhost_vdpa device unavailable\n", __func__);
+		return -ENODEV;
+	}
 	mutex_lock(&d->mutex);
 
 	switch (cmd) {
@@ -745,8 +766,13 @@ static int vhost_vdpa_process_iotlb_msg(struct vhost_dev *dev,
 	const struct vdpa_config_ops *ops = vdpa->config;
 	int r = 0;
 
-	mutex_lock(&dev->mutex);
+	if (v->dev_invalid) {
+		dev_info(&v->dev,
+			 "%s: vhost_vdpa device unavailable\n", __func__);
+		return -ENODEV;
+	}
 
+	mutex_lock(&dev->mutex);
 	r = vhost_dev_check_owner(dev);
 	if (r)
 		goto unlock;
@@ -949,6 +975,11 @@ static vm_fault_t vhost_vdpa_fault(struct vm_fault *vmf)
 	struct vm_area_struct *vma = vmf->vma;
 	u16 index = vma->vm_pgoff;
 
+	if (v->dev_invalid) {
+		dev_info(&v->dev,
+			 "%s: vhost_vdpa device unavailable\n", __func__);
+		return VM_FAULT_NOPAGE;
+	}
 	notify = ops->get_vq_notification(vdpa, index);
 
 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
@@ -1091,11 +1122,13 @@ static void vhost_vdpa_remove(struct vdpa_device *vdpa)
 		opened = atomic_cmpxchg(&v->opened, 0, 1);
 		if (!opened)
 			break;
-		wait_for_completion_timeout(&v->completion,
-					    msecs_to_jiffies(1000));
-		dev_warn_once(&v->dev,
-			      "%s waiting for /dev/%s to be closed\n",
-			      __func__, dev_name(&v->dev));
+		if (!wait_for_completion_timeout(&v->completion,
+					    msecs_to_jiffies(1000))) {
+			dev_warn(&v->dev,
+				 "%s /dev/%s in use, continue..\n",
+				 __func__, dev_name(&v->dev));
+			break;
+		}
 	} while (1);
 
 	put_device(&v->dev);

+ v->dev_invalid = true;

Signature · 1 line
2.30.1
#1 lore raw
Quoted 7 lines · On Mon, Jul 05, 2021 at 02:22:04AM +0530, [email protected] wrote:

From: Gautam Dawar <[email protected]>

As mentioned in Bug 213179, any malicious user-space application can render a module registering a vDPA device to hang forever. This will typically surface when vdpa_device_unregister() is called from the function responsible for module unload, leading rmmod commands to not return.

Quoted 6 lines

This patch unblocks the caller module by continuing with the clean-up but after marking the vhost device as unavailable. For future requests from user-space application, the vhost device availability is checked first and if it has gone unavailable, such requests are denied.

I don't seem mappings handled below. Did I miss it?

Quoted 97 lines
---
 drivers/vhost/vdpa.c | 45 ++++++++++++++++++++++++++++++++++++++------
 1 file changed, 39 insertions(+), 6 deletions(-)
drivers/vhost/vdpa.c +32 -3
index e4b7d26649d8..623bc7f0c0ca 100644
@@ -47,6 +47,7 @@ struct vhost_vdpa {
 	int minor;
 	struct eventfd_ctx *config_ctx;
 	int in_batch;
+	int dev_invalid;
 	struct vdpa_iova_range range;
 };
 
@@ -61,6 +62,11 @@ static void handle_vq_kick(struct vhost_work *work)
 	struct vhost_vdpa *v = container_of(vq->dev, struct vhost_vdpa, vdev);
 	const struct vdpa_config_ops *ops = v->vdpa->config;
 
+	if (v->dev_invalid) {
+		dev_info(&v->dev,
+			 "%s: vhost_vdpa device unavailable\n", __func__);
+		return;
+	}
 	ops->kick_vq(v->vdpa, vq - v->vqs);
 }
 
@@ -120,6 +126,11 @@ static void vhost_vdpa_reset(struct vhost_vdpa *v)
 {
 	struct vdpa_device *vdpa = v->vdpa;
 
+	if (v->dev_invalid) {
+		dev_info(&v->dev,
+			 "%s: vhost_vdpa device unavailable\n", __func__);
+		return;
+	}
 	vdpa_reset(vdpa);
 	v->in_batch = 0;
 }
@@ -367,6 +378,11 @@ static long vhost_vdpa_vring_ioctl(struct vhost_vdpa *v, unsigned int cmd,
 	u32 idx;
 	long r;
 
+	if (v->dev_invalid) {
+		dev_info(&v->dev,
+			 "%s: vhost_vdpa device unavailable\n", __func__);
+		return -ENODEV;
+	}
 	r = get_user(idx, (u32 __user *)argp);
 	if (r < 0)
 		return r;
@@ -450,6 +466,11 @@ static long vhost_vdpa_unlocked_ioctl(struct file *filep,
 		return 0;
 	}
 
+	if (v->dev_invalid) {
+		dev_info(&v->dev,
+			 "%s: vhost_vdpa device unavailable\n", __func__);
+		return -ENODEV;
+	}
 	mutex_lock(&d->mutex);
 
 	switch (cmd) {
@@ -745,8 +766,13 @@ static int vhost_vdpa_process_iotlb_msg(struct vhost_dev *dev,
 	const struct vdpa_config_ops *ops = vdpa->config;
 	int r = 0;
 
-	mutex_lock(&dev->mutex);
+	if (v->dev_invalid) {
+		dev_info(&v->dev,
+			 "%s: vhost_vdpa device unavailable\n", __func__);
+		return -ENODEV;
+	}
 
+	mutex_lock(&dev->mutex);
 	r = vhost_dev_check_owner(dev);
 	if (r)
 		goto unlock;
@@ -949,6 +975,11 @@ static vm_fault_t vhost_vdpa_fault(struct vm_fault *vmf)
 	struct vm_area_struct *vma = vmf->vma;
 	u16 index = vma->vm_pgoff;
 
+	if (v->dev_invalid) {
+		dev_info(&v->dev,
+			 "%s: vhost_vdpa device unavailable\n", __func__);
+		return VM_FAULT_NOPAGE;
+	}
 	notify = ops->get_vq_notification(vdpa, index);
 
 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
@@ -1091,11 +1122,13 @@ static void vhost_vdpa_remove(struct vdpa_device *vdpa)
 		opened = atomic_cmpxchg(&v->opened, 0, 1);
 		if (!opened)
 			break;
-		wait_for_completion_timeout(&v->completion,
-					    msecs_to_jiffies(1000));

drivers/vhost/vdpa.c static void vhost_vdpa_remove(struct vdpa_device *vdpa) · line 1131

Quoted 10 lines
drivers/vhost/vdpa.c +7 -3
-		dev_warn_once(&v->dev,
-			      "%s waiting for /dev/%s to be closed\n",
-			      __func__, dev_name(&v->dev));
+		if (!wait_for_completion_timeout(&v->completion,
+					    msecs_to_jiffies(1000))) {
+			dev_warn(&v->dev,
+				 "%s /dev/%s in use, continue..\n",
+				 __func__, dev_name(&v->dev));
+			break;
+		}

When you have an arbitrary timeout you know something's not entirely robust ...

Quoted 6 lines
drivers/vhost/vdpa.c +1 -0
 	} while (1);
 
 	put_device(&v->dev);
+	v->dev_invalid = true;
Signature · 1 line
2.30.1
#2 lore raw
In reply to [email protected]
Quoted 11 lines · 在 2021/7/5 上午4:52, [email protected] 写道:
  	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
drivers/vhost/vdpa.c +1 -5
@@ -1091,11 +1122,13 @@ static void vhost_vdpa_remove(struct vdpa_device *vdpa)
  		opened = atomic_cmpxchg(&v->opened, 0, 1);
  		if (!opened)
  			break;
-		wait_for_completion_timeout(&v->completion,
-					    msecs_to_jiffies(1000));
-		dev_warn_once(&v->dev,
-			      "%s waiting for/dev/%s to be closed\n",
-			      __func__, dev_name(&v->dev));
+		if (!wait_for_completion_timeout(&v->completion,

drivers/vhost/vdpa.c static void vhost_vdpa_remove(struct vdpa_device *vdpa)

Quoted 10 lines
drivers/vhost/vdpa.c +7 -0
+					    msecs_to_jiffies(1000))) {
+			dev_warn(&v->dev,
+				 "%s/dev/%s in use, continue..\n",
+				 __func__, dev_name(&v->dev));
+			break;
+		}
  	} while (1);
 
  	put_device(&v->dev);
+	v->dev_invalid = true;

Besides the mapping handling mentioned by Michael. I think this can lead use-after-free. put_device may release the memory.

Another fundamental issue, vDPA is the parent of vhost-vDPA device. I'm not sure the device core can allow the parent to go away first.

Thanks

#3 lore raw
Quoted 27 lines · Jason Wang
Quoted 11 lines · 在 2021/7/5 上午4:52, [email protected] 写道:

      vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);

drivers/vhost/vdpa.c +0 -5
@@ -1091,11 +1122,13 @@ static void vhost_vdpa_remove(struct 
 vdpa_device *vdpa)
           opened = atomic_cmpxchg(&v->opened, 0, 1);
           if (!opened)
               break;
-        wait_for_completion_timeout(&v->completion,
-                        msecs_to_jiffies(1000));
-        dev_warn_once(&v->dev,
-                  "%s waiting for/dev/%s to be closed\n",
-                  __func__, dev_name(&v->dev));

drivers/vhost/vdpa.c static void vhost_vdpa_remove(struct

Quoted 10 lines
drivers/vhost/vdpa.c +8 -0
+        if (!wait_for_completion_timeout(&v->completion,
+                        msecs_to_jiffies(1000))) {
+            dev_warn(&v->dev,
+                 "%s/dev/%s in use, continue..\n",
+                 __func__, dev_name(&v->dev));
+            break;
+        }
       } while (1);
         put_device(&v->dev);
+    v->dev_invalid = true;

Besides the mapping handling mentioned by Michael. I think this can lead use-after-free. put_device may release the memory.

Quoted 1 line

Another fundamental issue, vDPA is the parent of vhost-vDPA device. I'm not sure the device core can allow the parent to go away first.

Or this probably means you need couple the fd loosely with the vhost-vDPA device.

Thanks

Quoted 4 lines

Thanks