[PATCH net] net/tls: serialize context removal in tls_device_down()

[PATCH net] net/tls: serialize context removal in tls_device_down()

[Jaeyeong Lee]

tls_device_down() moves each context to tls_device_down_list while
holding tls_device_lock. It then drops its reference and removes the
context after releasing the lock.

tls_device_queue_ctx_destruction() removes contexts from the same list
under tls_device_lock. If it removes an adjacent context while
tls_device_down() is deleting its node, the two list updates can
interleave. One update can then target a context freed by the other path,
causing a use-after-free write.

Drop the reference and remove the context in the existing
tls_device_lock critical section. Keep tls_device_free_ctx() after
releasing the lock.

This issue was found with a static analysis tool.

Fixes: f6336724a4d4 ("net/tls: Remove the context from the list in tls_device_down")
Cc: stable@vger.kernel.org
Signed-off-by: Jaeyeong Lee <lee@jaeyeong.cc>
---
Reproduced with a deterministic KUnit harness under KASAN in a 2-vCPU
x86_64 QEMU guest. The unpatched kernel reported:

  BUG: KASAN: slab-use-after-free in tls_device_racy_list_del
  Write of size 8

The patched tls_device.o was compile-tested against current net/main with
KASAN and DEBUG_LIST enabled.

 net/tls/tls_device.c | 10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)

diff --git a/net/tls/tls_device.c b/net/tls/tls_device.c
index 4cdfbb57ecff0..fa7170cd4fd10 100644
--- a/net/tls/tls_device.c
+++ b/net/tls/tls_device.c
@@ -1350,6 +1350,8 @@ static int tls_device_down(struct net_device *netdev)
 	spin_unlock_irqrestore(&tls_device_lock, flags);
 
 	list_for_each_entry_safe(ctx, tmp, &list, list)	{
+		bool free_ctx;
+
 		/* Stop offloaded TX and switch to the fallback.
 		 * tls_is_skb_tx_device_offloaded will return false.
 		 */
@@ -1387,17 +1389,19 @@ static int tls_device_down(struct net_device *netdev)
 		 */
 		spin_lock_irqsave(&tls_device_lock, flags);
 		list_move_tail(&ctx->list, &tls_device_down_list);
+		/* Serialize the final removal with sk_destruct. */
+		free_ctx = refcount_dec_and_test(&ctx->refcount);
+		if (free_ctx)
+			list_del(&ctx->list);
 		spin_unlock_irqrestore(&tls_device_lock, flags);
 
 		/* Device contexts for RX and TX will be freed in on sk_destruct
 		 * by tls_device_free_ctx. rx_conf and tx_conf stay in TLS_HW.
-		 * Now release the ref taken above.
 		 */
-		if (refcount_dec_and_test(&ctx->refcount)) {
+		if (free_ctx) {
 			/* sk_destruct ran after tls_device_down took a ref, and
 			 * it returned early. Complete the destruction here.
 			 */
-			list_del(&ctx->list);
 			tls_device_free_ctx(ctx);
 		}
 	}
-- 
2.43.0