[PATCH v2] KVM: guest_memfd: Elaborate on how release() vs. get_pfn() is safe against UAF

[PATCH v2] KVM: guest_memfd: Elaborate on how release() vs. get_pfn() is safe against UAF

[Sean Christopherson]

Add more context and information to the comment in kvm_gmem_release() that
explains why there's no synchronization on RCU _or_ kvm->srcu.  Point (b)
from commit 67b43038ce14 ("KVM: guest_memfd: Remove RCU-protected attribute
from slot->gmem.file")

      b) kvm->srcu ensures that kvm_gmem_unbind() and freeing of a memslot
         occur after the memslot is no longer visible to kvm_gmem_get_pfn().

is especially difficult to fully grok, particularly in light of commit
ae431059e75d ("KVM: guest_memfd: Remove bindings on memslot deletion when
gmem is dying"), which addressed a race between unbind() and release().

See the extended on-list discussion[*] for more details about exactly what
KVM guards against, and how.

No functional change intended.

Link: https://lore.kernel.org/all/CAEvNRgGmyd1yqQXsnz5hWRZpBZUs%3DpiEWbEaqP9%2Bcz9ZqEMQ6g@mail.gmail.com [*]
Cc: Yan Zhao <yan.y.zhao@intel.com>
Cc: Vishal Annapurve <vannapurve@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
---

v2: Explain how this all works in even gorier detail. [Yan]

v1: https://lore.kernel.org/all/20251113232229.1698886-1-seanjc@google.com

 virt/kvm/guest_memfd.c | 50 +++++++++++++++++++++++++++++++++++++-----
 1 file changed, 44 insertions(+), 6 deletions(-)

diff --git a/virt/kvm/guest_memfd.c b/virt/kvm/guest_memfd.c
index b596486d184c..7f1c6a0f8039 100644
--- a/virt/kvm/guest_memfd.c
+++ b/virt/kvm/guest_memfd.c
@@ -300,17 +300,55 @@ static int kvm_gmem_release(struct inode *inode, struct file *file)
 	 * dereferencing the slot for existing bindings needs to be protected
 	 * against memslot updates, specifically so that unbind doesn't race
 	 * and free the memslot (kvm_gmem_get_file() will return NULL).
-	 *
-	 * Since .release is called only when the reference count is zero,
-	 * after which file_ref_get() and get_file_active() fail,
-	 * kvm_gmem_get_pfn() cannot be using the file concurrently.
-	 * file_ref_put() provides a full barrier, and get_file_active() the
-	 * matching acquire barrier.
 	 */
 	mutex_lock(&kvm->slots_lock);
 
 	filemap_invalidate_lock(inode->i_mapping);
 
+	/*
+	 * Note!  synchronize_srcu() is _not_ needed after nullifying memslot
+	 * bindings as slot->gmem.file cannot be set back to a non-null value
+	 * without the memslot first being deleted.  I.e. this relies on the
+	 * synchronize_srcu_expedited() in kvm_swap_active_memslots() to ensure
+	 * kvm_gmem_get_pfn() (which runs with kvm->srcu held for read) can't
+	 * grab a reference to slot->gmem.file even if the struct file object
+	 * is reallocated.
+	 *
+	 * file_ref_put() provides a full barrier, and __get_file_rcu() the
+	 * matching acquire barrier, to ensure that kvm_gmem_get_file() (via
+	 * __get_file_rcu()) sees refcount==0 or fails the "file reloaded"
+	 * check (file != NULL due to nullifying the file pointer here).
+	 *
+	 * Unlike most other users of get_file_rcu(), where callers don't care
+	 * if they race with a write, only that they have a reference to _a_
+	 * live file, kvm_gmem_get_pfn() needs to get the exact file that is
+	 * associated with the memslot.  Without the aforementioned SRCU
+	 * synchronization, the following could happen:
+	 *
+	 *  CPU0				CPU1
+	 *  kvm_gmem_get_pfn()
+	 *    f = X (from slot->gmem.file)
+	 *					kvm_gmem_release())
+	 *					  slot->gmem.file = NULL
+	 *
+	 *					kvm_set_memory_region()
+	 *					  slot deleted
+	 *
+	 *					kvm_set_memory_region()
+	 *					  slot created
+	 *					  slot->gmem.file = f (alloc the same object)
+	 *
+	 *  get_file_active()
+	 *    file = f
+	 *    file_reloaded = f
+	 *
+	 * <KVM does weird things with an old memslot+file>
+	 *
+	 * Obviously KVM would be broken in many places if the synchronization
+	 * were omitted, but it's important to note that get_file_active() does
+	 * NOT guarantee a reference to the correct file was obtained, only
+	 * that the file doesn't point at a reallocated object.
+	 */
 	xa_for_each(&f->bindings, index, slot)
 		WRITE_ONCE(slot->gmem.file, NULL);
 

base-commit: 76671054f9a1ff6abb976583cd8da37650acdc97
-- 
2.55.0.887.g758fc8c411-goog

Re: [PATCH v2] KVM: guest_memfd: Elaborate on how release() vs. get_pfn() is safe against UAF

[sashiko-bot]

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Pre-existing issues:
- [Critical] Bypassing filemap_invalidate_lock in kvm_gmem_unbind when the file is dying allows concurrent xarray modifications, causing Use-After-Free and leaked mmu_invalidate_in_progress.
--

commit 22b749f62017c828a40c33364eefe3474406907f
Author: Sean Christopherson <seanjc@google.com>

KVM: guest_memfd: Elaborate on how release() vs. get_pfn() is safe against UAF

This commit updates the comment in kvm_gmem_release() to elaborate on the
synchronization between release() and get_pfn(). It explains why neither
RCU nor kvm->srcu synchronization is necessary when nullifying memslot
bindings, detailing the barriers used to prevent a use-after-free.

Link: https://lore.kernel.org/all/CAEvNRgGmyd1yqQXsnz5hWRZpBZUs%3DpiEWbEaqP9%2Bcz9ZqEMQ6g@mail.gmail.com [*]

> diff --git a/virt/kvm/guest_memfd.c b/virt/kvm/guest_memfd.c
> index b596486d184ca..7f1c6a0f8039d 100644
> --- a/virt/kvm/guest_memfd.c
> +++ b/virt/kvm/guest_memfd.c
[ ... ]
> @@ -300,17 +300,55 @@ static int kvm_gmem_release(struct inode *inode, struct file *file)
>  	 * dereferencing the slot for existing bindings needs to be protected
>  	 * against memslot updates, specifically so that unbind doesn't race
>  	 * and free the memslot (kvm_gmem_get_file() will return NULL).
>  	 */
>  	mutex_lock(&kvm->slots_lock);
>  
>  	filemap_invalidate_lock(inode->i_mapping);
>  
> +	/*
> +	 * Note!  synchronize_srcu() is _not_ needed after nullifying memslot
[ ... ]
> +	 * Obviously KVM would be broken in many places if the synchronization
> +	 * were omitted, but it's important to note that get_file_active() does
> +	 * NOT guarantee a reference to the correct file was obtained, only
> +	 * that the file doesn't point at a reallocated object.
> +	 */
>  	xa_for_each(&f->bindings, index, slot)
>  		WRITE_ONCE(slot->gmem.file, NULL);
>  

[Severity: Critical]
This is a pre-existing issue, but could bypassing filemap_invalidate_lock in
kvm_gmem_unbind() when the file is dying allow concurrent xarray modifications
that result in a use-after-free?

In kvm_gmem_unbind(), when the file is dying:

	if (!file) {
		__kvm_gmem_unbind(slot, slot->gmem.file->private_data);
		return;
	}

Because kvm_gmem_error_folio() handles memory failure on guest_memfd pages by
holding filemap_invalidate_lock_shared(mapping) and iterates f->bindings
without holding kvm->srcu or kvm->slots_lock, it relies entirely on
filemap_invalidate_lock.

If kvm_gmem_unbind() bypasses filemap_invalidate_lock when the underlying
file is dying, it can run concurrently with kvm_gmem_error_folio() and
remove a slot from f->bindings. This allows kvm_set_memory_region() to free
the memslot while kvm_gmem_error_folio() is still holding a pointer to it.

Could this cause a use-after-free during the iteration:

virt/kvm/guest_memfd.c:__kvm_gmem_invalidate_start() {
	xa_for_each_range(&f->bindings, index, slot, start, end - 1) {
		pgoff_t pgoff = slot->gmem.pgoff;
		...
}

Additionally, if the slot is removed before __kvm_gmem_invalidate_end(),
xa_find() returns false, which skips kvm_mmu_invalidate_end() and
permanently leaks mmu_invalidate_in_progress:

virt/kvm/guest_memfd.c:__kvm_gmem_invalidate_end() {
	if (xa_find(&f->bindings, &start, end - 1, XA_PRESENT)) {
		KVM_MMU_LOCK(kvm);
		kvm_mmu_invalidate_end(kvm);
		KVM_MMU_UNLOCK(kvm);
	}
}

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20260826165647.769231-1-seanjc@google.com?part=1

Re: [PATCH v2] KVM: guest_memfd: Elaborate on how release() vs. get_pfn() is safe against UAF

[Sean Christopherson]

On Wed, Aug 26, 2026, sashiko-bot@kernel.org wrote:
> > diff --git a/virt/kvm/guest_memfd.c b/virt/kvm/guest_memfd.c
> > index b596486d184ca..7f1c6a0f8039d 100644
> > --- a/virt/kvm/guest_memfd.c
> > +++ b/virt/kvm/guest_memfd.c
> [ ... ]
> > @@ -300,17 +300,55 @@ static int kvm_gmem_release(struct inode *inode, struct file *file)
> >  	 * dereferencing the slot for existing bindings needs to be protected
> >  	 * against memslot updates, specifically so that unbind doesn't race
> >  	 * and free the memslot (kvm_gmem_get_file() will return NULL).
> >  	 */
> >  	mutex_lock(&kvm->slots_lock);
> >  
> >  	filemap_invalidate_lock(inode->i_mapping);
> >  
> > +	/*
> > +	 * Note!  synchronize_srcu() is _not_ needed after nullifying memslot
> [ ... ]
> > +	 * Obviously KVM would be broken in many places if the synchronization
> > +	 * were omitted, but it's important to note that get_file_active() does
> > +	 * NOT guarantee a reference to the correct file was obtained, only
> > +	 * that the file doesn't point at a reallocated object.
> > +	 */
> >  	xa_for_each(&f->bindings, index, slot)
> >  		WRITE_ONCE(slot->gmem.file, NULL);
> >  
> 
> [Severity: Critical]
> This is a pre-existing issue, but could bypassing filemap_invalidate_lock in
> kvm_gmem_unbind() when the file is dying allow concurrent xarray modifications
> that result in a use-after-free?

Yep, already reported by Sashiko and being fixed:
https://lore.kernel.org/all/20260823-shivank-gmem-fix-split-v1-1-512a29fb8e86@amd.com