Skip to contentSkip to conversation

vfio/iommu_type1: acquire iommu lock in vfio_iommu_type1_release()

PATCH In review Sidhartha Kumar 2 participants 4 messages Jun 7, 2023 – Jun 8, 2023

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 vfio/iommu_type1: acquire iommu lock in vfio_iommu_type1_release()

From vfio_iommu_type1_release() there is a code path:

vfio_iommu_unmap_unpin_all()
vfio_remove_dma()
vfio_unmap_unpin()
unmap_unpin_slow()
vfio_unpin_pages_remote()
vfio_find_vpfn()

This path is taken without acquiring the iommu lock so it could lead to a race condition in the traversal of the pfn_list rb tree. The lack of the iommu lock in vfio_iommu_type1_release() was confirmed by adding a

WARN_ON(!mutex_is_locked(&iommu->lock))

which was reported in dmesg. Fix this potential race by adding a iommu lock and release in vfio_iommu_type1_release().

---
 drivers/vfio/vfio_iommu_type1.c | 2 ++
 1 file changed, 2 insertions(+)
drivers/vfio/vfio_iommu_type1.c +2 -0
index 306e6f1d1c70e..7d2fea1b483dc 100644
@@ -2601,7 +2601,9 @@ static void vfio_iommu_type1_release(void *iommu_data)
 		kfree(group);
 	}
 
+	mutex_lock(&iommu->lock);
 	vfio_iommu_unmap_unpin_all(iommu);
+	mutex_unlock(&iommu->lock);
 
 	list_for_each_entry_safe(domain, domain_tmp,
 				 &iommu->domain_list, next) {
Signature · 1 line
2.40.1
#1 lore raw
Quoted 11 lines · On Wed, 7 Jun 2023 12:07:52 -0700 Sidhartha Kumar…

From vfio_iommu_type1_release() there is a code path:

vfio_iommu_unmap_unpin_all()
vfio_remove_dma()
vfio_unmap_unpin()
unmap_unpin_slow()
vfio_unpin_pages_remote()
vfio_find_vpfn()

This path is taken without acquiring the iommu lock so it could lead to a race condition in the traversal of the pfn_list rb tree.

What's the competing thread for the race, vfio_remove_dma() tests:

	WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list));

The fix is not unreasonable, but is this a theoretical fix upstream that's tickled by some downstream additions, or are we actually competing against page pinning by an mdev driver after the container is released? Thanks,

Alex

Quoted 28 lines

The lack of
the iommu lock in vfio_iommu_type1_release() was confirmed by adding a

WARN_ON(!mutex_is_locked(&iommu->lock))

which was reported in dmesg. Fix this potential race by adding a iommu lock and release in vfio_iommu_type1_release().

---
 drivers/vfio/vfio_iommu_type1.c | 2 ++
 1 file changed, 2 insertions(+)
drivers/vfio/vfio_iommu_type1.c +2 -0
index 306e6f1d1c70e..7d2fea1b483dc 100644
@@ -2601,7 +2601,9 @@ static void vfio_iommu_type1_release(void *iommu_data)
 		kfree(group);
 	}
 
+	mutex_lock(&iommu->lock);
 	vfio_iommu_unmap_unpin_all(iommu);
+	mutex_unlock(&iommu->lock);
 
 	list_for_each_entry_safe(domain, domain_tmp,
 				 &iommu->domain_list, next) {
#2 lore raw
Quoted 15 lines · Alex Williamson
Quoted 11 lines · On Wed, 7 Jun 2023 12:07:52 -0700 Sidhartha Kumar…

From vfio_iommu_type1_release() there is a code path:

vfio_iommu_unmap_unpin_all()
vfio_remove_dma()
vfio_unmap_unpin()
unmap_unpin_slow()
vfio_unpin_pages_remote()
vfio_find_vpfn()

This path is taken without acquiring the iommu lock so it could lead to a race condition in the traversal of the pfn_list rb tree.

Quoted 9 lines

What's the competing thread for the race, vfio_remove_dma() tests:

	WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list));

The fix is not unreasonable, but is this a theoretical fix upstream that's tickled by some downstream additions, or are we actually competing against page pinning by an mdev driver after the container is released? Thanks,

Hello,

In a stress test of starting and stopping multiple VMs for a few days we observed a memory leak that occurs after a few days. These guests have their memory pinned via the pin_user_pages_remote() call in vaddr_get_pfns(). From examining the vfio/iommu_type1 code this potential race condition was noticed, but we have not root caused this race to be the cause of the memory leak.

Thanks,
Sidhartha Kumar

Quoted 31 lines

Alex

Quoted 28 lines

The lack of
the iommu lock in vfio_iommu_type1_release() was confirmed by adding a

WARN_ON(!mutex_is_locked(&iommu->lock))

which was reported in dmesg. Fix this potential race by adding a iommu lock and release in vfio_iommu_type1_release().

---
  drivers/vfio/vfio_iommu_type1.c | 2 ++
  1 file changed, 2 insertions(+)
drivers/vfio/vfio_iommu_type1.c +2 -0
index 306e6f1d1c70e..7d2fea1b483dc 100644
@@ -2601,7 +2601,9 @@ static void vfio_iommu_type1_release(void *iommu_data)
  		kfree(group);
  	}
 
+	mutex_lock(&iommu->lock);
  	vfio_iommu_unmap_unpin_all(iommu);
+	mutex_unlock(&iommu->lock);
 
  	list_for_each_entry_safe(domain, domain_tmp,
  				 &iommu->domain_list, next) {
#3 lore raw
Quoted 26 lines · On Thu, 8 Jun 2023 13:16:12 -0700 Sidhartha Kumar <[email protected]>…
Quoted 15 lines · Alex Williamson
Quoted 11 lines · On Wed, 7 Jun 2023 12:07:52 -0700 Sidhartha Kumar…

From vfio_iommu_type1_release() there is a code path:

vfio_iommu_unmap_unpin_all()
vfio_remove_dma()
vfio_unmap_unpin()
unmap_unpin_slow()
vfio_unpin_pages_remote()
vfio_find_vpfn()

This path is taken without acquiring the iommu lock so it could lead to a race condition in the traversal of the pfn_list rb tree.

Quoted 9 lines

What's the competing thread for the race, vfio_remove_dma() tests:

	WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list));

The fix is not unreasonable, but is this a theoretical fix upstream that's tickled by some downstream additions, or are we actually competing against page pinning by an mdev driver after the container is released? Thanks,

Quoted 8 lines

Hello,

In a stress test of starting and stopping multiple VMs for a few days we observed a memory leak that occurs after a few days. These guests have their memory pinned via the pin_user_pages_remote() call in vaddr_get_pfns(). From examining the vfio/iommu_type1 code this potential race condition was noticed, but we have not root caused this race to be the cause of the memory leak.

Ok, I think we're going to need to wait for that root cause. It's trivial to hold the lock here, but we're imminently freeing the object hosting that lock, so we have bigger problems than threads stepping on each other in the pfn_list if this lock were to actually resolve anything. It's not an oversight that we don't hold the lock here, there should be no outstanding references to the iommu object at this point. Thanks,

Alex

Quoted 30 lines
Quoted 29 lines
Quoted 28 lines

The lack of
the iommu lock in vfio_iommu_type1_release() was confirmed by adding a

WARN_ON(!mutex_is_locked(&iommu->lock))

which was reported in dmesg. Fix this potential race by adding a iommu lock and release in vfio_iommu_type1_release().

---
  drivers/vfio/vfio_iommu_type1.c | 2 ++
  1 file changed, 2 insertions(+)
drivers/vfio/vfio_iommu_type1.c +2 -0
index 306e6f1d1c70e..7d2fea1b483dc 100644
@@ -2601,7 +2601,9 @@ static void vfio_iommu_type1_release(void *iommu_data)
  		kfree(group);
  	}
 
+	mutex_lock(&iommu->lock);
  	vfio_iommu_unmap_unpin_all(iommu);
+	mutex_unlock(&iommu->lock);
 
  	list_for_each_entry_safe(domain, domain_tmp,
  				 &iommu->domain_list, next) {