Re: [PATCH] tpm: xen-tpmfront: fix use-after-free in backend_changed()

From: Jürgen Groß

Date: Thu Nov 19 2026 - 10:36:14 EST


On 19.11.26 16:20, Jarkko Sakkinen wrote:
On Thu, Nov 19, 2026 at 11:20:33AM +0000, Suyuan Wang via B4 Relay wrote:
From: Suyuan Wang <stevewong@xxxxxxxxxxx>

When the backend moves to Closing or Closed, backend_changed()
unregisters the device and then keeps using it:

device_unregister(&dev->dev);
xenbus_frontend_closed(dev);

The xenbus watch callback that calls backend_changed() does not hold a
reference to the device. The tpm_chip registered by xen-tpmfront has dev
as its parent and holds a reference to it, but tpmfront_remove() drops
that reference when it unregisters the chip. So device_unregister() can
release the last reference and free the xenbus_device, and
xenbus_frontend_closed(dev) then accesses freed memory.

Take a temporary reference before unregistering and drop it after
xenbus_frontend_closed(). This keeps dev alive for the last access and
keeps the existing order, in which TPM operations are stopped before
the frontend is marked Closed.

Found with a static analysis tool and checked by hand.

Fixes: e2683957fb26 ("drivers/tpm: add xen tpmfront interface")
Cc: stable@xxxxxxxxxxxxxxx
Co-developed-by: Kaixuan Li <kaixuan.li@xxxxxxxxxx>
Signed-off-by: Kaixuan Li <kaixuan.li@xxxxxxxxxx>
Signed-off-by: Suyuan Wang <stevewong@xxxxxxxxxxx>
---
Compile tested on x86_64 (W=1, no new warnings). Not run with a Xen vTPM
backend.
---
drivers/char/tpm/xen-tpmfront.c | 3 +++
1 file changed, 3 insertions(+)

diff --git a/drivers/char/tpm/xen-tpmfront.c b/drivers/char/tpm/xen-tpmfront.c
index 3c0cce57c4ae..1be925cca535 100644
--- a/drivers/char/tpm/xen-tpmfront.c
+++ b/drivers/char/tpm/xen-tpmfront.c
@@ -397,8 +397,11 @@ static void backend_changed(struct xenbus_device *dev,
case XenbusStateClosing:
case XenbusStateClosed:
+ /* device_unregister() can drop the last reference to dev */
+ get_device(&dev->dev);
device_unregister(&dev->dev);
xenbus_frontend_closed(dev);
+ put_device(&dev->dev);

Not a NAK but makes me wonder why this has not hit the wall ever.

I.e. "feeling confused".
Is there any vtpm backend other than the stubdom based one?

If not, then it is quite easy: this backend never switches state to
XenbusStateClosing or XenbusStateClosed.


Juergen

Attachment: OpenPGP_signature.asc
Description: OpenPGP digital signature