[PATCH] drm/amdgpu: fix PASID task_info lookup race

Fan Wu posted 1 patch 1 month ago
There is a newer version of this series
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 40 ++++++++++++++++----------
1 file changed, 25 insertions(+), 15 deletions(-)
[PATCH] drm/amdgpu: fix PASID task_info lookup race
Posted by Fan Wu 1 month ago
The amdgpu_vm_get_task_info_pasid() function previously called
amdgpu_vm_get_vm_from_pasid() which returns a raw VM pointer after
releasing the pasids xarray lock. The caller then dereferences
vm->task_info without any lifetime protection.

Race condition:

    CPU 0 (lookup)                      CPU 1 (release)
    ------------------                  ------------------
    amdgpu_vm_get_task_info_pasid()
      xa_lock()
      vm = xa_load(pasids)
      xa_unlock()
                                        amdgpu_vm_fini()
                                          xa_erase_irq(pasids)
                                          // teardown continues
                                        kfree(fpriv)
                                        // VM freed (embedded in fpriv)
      vm->task_info  // potential UAF

This can leave the VM pointer dangling because struct amdgpu_vm is
embedded in struct amdgpu_fpriv which is freed via kfree(fpriv) in
amdgpu_file_release_kms() after amdgpu_vm_fini() returns.

Fix this by acquiring the task_info reference while holding the
xarray lock. This avoids the window where the VM could be freed
between the lookup and the dereference.

Cache vm->task_info in a local variable before attempting to take a
reference, which keeps the lookup straightforward inside the locked
section. Use kref_get_unless_zero() to safely handle the case where
task_info's refcount is already being decremented to zero by another
thread in the teardown path.

Note: An RCU-based approach was considered but is not currently
feasible because: (1) the pasids xarray is initialized without
XA_FLAGS_RCU, and (2) struct amdgpu_fpriv is freed with kfree()
rather than kfree_rcu(). A future refactoring could enable RCU
if needed for performance.

Also remove the unsafe helper function amdgpu_vm_get_vm_from_pasid()
to prevent future misuse.

Fixes: b8f67b9ddf4f ("drm/amdgpu: change vm->task_info handling")
Signed-off-by: Fan Wu <fanwu01@zju.edu.cn>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 40 ++++++++++++++++----------
 1 file changed, 25 insertions(+), 15 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
index f2beb980e3c3..7e8621c9b661 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
@@ -2468,19 +2468,6 @@ static void amdgpu_vm_destroy_task_info(struct kref *kref)
 	kfree(ti);
 }
 
-static inline struct amdgpu_vm *
-amdgpu_vm_get_vm_from_pasid(struct amdgpu_device *adev, u32 pasid)
-{
-	struct amdgpu_vm *vm;
-	unsigned long flags;
-
-	xa_lock_irqsave(&adev->vm_manager.pasids, flags);
-	vm = xa_load(&adev->vm_manager.pasids, pasid);
-	xa_unlock_irqrestore(&adev->vm_manager.pasids, flags);
-
-	return vm;
-}
-
 /**
  * amdgpu_vm_put_task_info - reference down the vm task_info ptr
  *
@@ -2527,8 +2514,31 @@ amdgpu_vm_get_task_info_vm(struct amdgpu_vm *vm)
 struct amdgpu_task_info *
 amdgpu_vm_get_task_info_pasid(struct amdgpu_device *adev, u32 pasid)
 {
-	return amdgpu_vm_get_task_info_vm(
-			amdgpu_vm_get_vm_from_pasid(adev, pasid));
+	struct amdgpu_vm *vm;
+	unsigned long flags;
+	struct amdgpu_task_info *ti = NULL;
+
+	/*
+	 * Acquire the task_info reference while holding the pasids xarray
+	 * lock to prevent a race with amdgpu_vm_fini() which removes the
+	 * PASID mapping before freeing the VM (embedded in struct amdgpu_fpriv).
+	 * Without this, the VM could be freed between xa_load() return and
+	 * the task_info dereference.
+	 */
+	xa_lock_irqsave(&adev->vm_manager.pasids, flags);
+	vm = xa_load(&adev->vm_manager.pasids, pasid);
+	if (vm) {
+		/*
+		 * Cache vm->task_info in a local variable before
+		 * attempting to take a reference.
+		 */
+		ti = vm->task_info;
+		if (ti && !kref_get_unless_zero(&ti->refcount))
+			ti = NULL;
+	}
+	xa_unlock_irqrestore(&adev->vm_manager.pasids, flags);
+
+	return ti;
 }
 
 static int amdgpu_vm_create_task_info(struct amdgpu_vm *vm)
-- 
2.34.1
Re: [PATCH] drm/amdgpu: fix PASID task_info lookup race
Posted by Christian König 1 month ago
On 3/9/26 17:04, Fan Wu wrote:
> The amdgpu_vm_get_task_info_pasid() function previously called
> amdgpu_vm_get_vm_from_pasid() which returns a raw VM pointer after
> releasing the pasids xarray lock. The caller then dereferences
> vm->task_info without any lifetime protection.
> 
> Race condition:
> 
>     CPU 0 (lookup)                      CPU 1 (release)
>     ------------------                  ------------------
>     amdgpu_vm_get_task_info_pasid()
>       xa_lock()
>       vm = xa_load(pasids)
>       xa_unlock()
>                                         amdgpu_vm_fini()
>                                           xa_erase_irq(pasids)
>                                           // teardown continues
>                                         kfree(fpriv)
>                                         // VM freed (embedded in fpriv)
>       vm->task_info  // potential UAF
> 
> This can leave the VM pointer dangling because struct amdgpu_vm is
> embedded in struct amdgpu_fpriv which is freed via kfree(fpriv) in
> amdgpu_file_release_kms() after amdgpu_vm_fini() returns.
> 
> Fix this by acquiring the task_info reference while holding the
> xarray lock. This avoids the window where the VM could be freed
> between the lookup and the dereference.
> 
> Cache vm->task_info in a local variable before attempting to take a
> reference, which keeps the lookup straightforward inside the locked
> section. Use kref_get_unless_zero() to safely handle the case where
> task_info's refcount is already being decremented to zero by another
> thread in the teardown path.
> 
> Note: An RCU-based approach was considered but is not currently
> feasible because: (1) the pasids xarray is initialized without
> XA_FLAGS_RCU, and (2) struct amdgpu_fpriv is freed with kfree()
> rather than kfree_rcu(). A future refactoring could enable RCU
> if needed for performance.
> 
> Also remove the unsafe helper function amdgpu_vm_get_vm_from_pasid()
> to prevent future misuse.
> 
> Fixes: b8f67b9ddf4f ("drm/amdgpu: change vm->task_info handling")
> Signed-off-by: Fan Wu <fanwu01@zju.edu.cn>
> ---
>  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 40 ++++++++++++++++----------
>  1 file changed, 25 insertions(+), 15 deletions(-)
> 
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> index f2beb980e3c3..7e8621c9b661 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> @@ -2468,19 +2468,6 @@ static void amdgpu_vm_destroy_task_info(struct kref *kref)
>         kfree(ti);
>  }
> 
> -static inline struct amdgpu_vm *
> -amdgpu_vm_get_vm_from_pasid(struct amdgpu_device *adev, u32 pasid)
> -{
> -       struct amdgpu_vm *vm;
> -       unsigned long flags;
> -
> -       xa_lock_irqsave(&adev->vm_manager.pasids, flags);
> -       vm = xa_load(&adev->vm_manager.pasids, pasid);
> -       xa_unlock_irqrestore(&adev->vm_manager.pasids, flags);
> -
> -       return vm;
> -}
> -
>  /**
>   * amdgpu_vm_put_task_info - reference down the vm task_info ptr
>   *
> @@ -2527,8 +2514,31 @@ amdgpu_vm_get_task_info_vm(struct amdgpu_vm *vm)
>  struct amdgpu_task_info *
>  amdgpu_vm_get_task_info_pasid(struct amdgpu_device *adev, u32 pasid)
>  {
> -       return amdgpu_vm_get_task_info_vm(
> -                       amdgpu_vm_get_vm_from_pasid(adev, pasid));
> +       struct amdgpu_vm *vm;
> +       unsigned long flags;
> +       struct amdgpu_task_info *ti = NULL;
> +
> +       /*
> +        * Acquire the task_info reference while holding the pasids xarray
> +        * lock to prevent a race with amdgpu_vm_fini() which removes the
> +        * PASID mapping before freeing the VM (embedded in struct amdgpu_fpriv).
> +        * Without this, the VM could be freed between xa_load() return and
> +        * the task_info dereference.

That the VM is freed is irrelevant, the point is that we need to grab the reference to the task info before we drop that one.

> +        */
> +       xa_lock_irqsave(&adev->vm_manager.pasids, flags);
> +       vm = xa_load(&adev->vm_manager.pasids, pasid);
> +       if (vm) {
> +               /*
> +                * Cache vm->task_info in a local variable before
> +                * attempting to take a reference.
> +                */

Please drop that comment, taking the task info into a local variable is actually superflous.

> +               ti = vm->task_info;
> +               if (ti && !kref_get_unless_zero(&ti->refcount))

That is unecessary as wel, the task info is dropped after the VM is removed from pasid mapping.

So just using kref_get() is sufficient.

Regards,
Christian.

> +                       ti = NULL;
> +       }
> +       xa_unlock_irqrestore(&adev->vm_manager.pasids, flags);
> +
> +       return ti;
>  }
> 
>  static int amdgpu_vm_create_task_info(struct amdgpu_vm *vm)
> --
> 2.34.1
>
[PATCH v2] drm/amdgpu: fix PASID task_info lookup race
Posted by Fan Wu 1 month ago
amdgpu_vm_get_task_info_pasid() currently looks up the VM from the
PASID xarray, drops the xarray lock, and only then grabs the task_info
reference through amdgpu_vm_get_task_info_vm().

Take the task_info reference directly while holding the PASID xarray
lock instead. This keeps the lookup and reference acquisition in the
same critical section.

The task_info is dropped only after the VM is removed from the PASID
mapping, so a regular kref_get() is sufficient here.

Also remove the now unnecessary amdgpu_vm_get_vm_from_pasid() helper.

Fixes: b8f67b9ddf4f ("drm/amdgpu: change vm->task_info handling")
Signed-off-by: Fan Wu <fanwu01@zju.edu.cn>
---
v2:
- grab the task_info reference directly under the PASID xarray lock
- drop the extra local-variable comment
- use kref_get() instead of kref_get_unless_zero()
- simplify the changelog

 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 35 +++++++++++++++-----------
 1 file changed, 20 insertions(+), 15 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
index f2beb980e3c3..706ca8dd65d2 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
@@ -2468,19 +2468,6 @@ static void amdgpu_vm_destroy_task_info(struct kref *kref)
 	kfree(ti);
 }
 
-static inline struct amdgpu_vm *
-amdgpu_vm_get_vm_from_pasid(struct amdgpu_device *adev, u32 pasid)
-{
-	struct amdgpu_vm *vm;
-	unsigned long flags;
-
-	xa_lock_irqsave(&adev->vm_manager.pasids, flags);
-	vm = xa_load(&adev->vm_manager.pasids, pasid);
-	xa_unlock_irqrestore(&adev->vm_manager.pasids, flags);
-
-	return vm;
-}
-
 /**
  * amdgpu_vm_put_task_info - reference down the vm task_info ptr
  *
@@ -2527,8 +2514,26 @@ amdgpu_vm_get_task_info_vm(struct amdgpu_vm *vm)
 struct amdgpu_task_info *
 amdgpu_vm_get_task_info_pasid(struct amdgpu_device *adev, u32 pasid)
 {
-	return amdgpu_vm_get_task_info_vm(
-			amdgpu_vm_get_vm_from_pasid(adev, pasid));
+	struct amdgpu_vm *vm;
+	unsigned long flags;
+	struct amdgpu_task_info *ti = NULL;
+
+	/*
+	 * Acquire the task_info reference while holding the pasids xarray
+	 * lock to prevent a race with amdgpu_vm_fini() which removes the
+	 * PASID mapping before freeing the VM (embedded in struct amdgpu_fpriv).
+	 * Without this, the VM could be freed between xa_load() return and
+	 * the task_info dereference.
+	 */
+	xa_lock_irqsave(&adev->vm_manager.pasids, flags);
+	vm = xa_load(&adev->vm_manager.pasids, pasid);
+	if (vm && vm->task_info) {
+		ti = vm->task_info;
+		kref_get(&ti->refcount);
+	}
+	xa_unlock_irqrestore(&adev->vm_manager.pasids, flags);
+
+	return ti;
 }
 
 static int amdgpu_vm_create_task_info(struct amdgpu_vm *vm)
-- 
2.34.1