[PATCH] xfrm: iptfs: fix deadlock in iptfs_destroy_state

[PATCH] xfrm: iptfs: fix deadlock in iptfs_destroy_state

[Dudu Lu]

iptfs_destroy_state() acquires x->lock (spin_lock_bh) and then calls
hrtimer_cancel(&xtfs->iptfs_timer). The timer callback
iptfs_delay_timer() also acquires x->lock (spin_lock). If the timer
fires on another CPU during destroy, hrtimer_cancel() waits for the
callback to complete, but the callback is blocked trying to acquire
the same lock — a classic ABBA deadlock.

The same pattern exists for drop_timer: destroy holds drop_lock and
calls hrtimer_cancel(&xtfs->drop_timer), while iptfs_drop_timer()
also acquires drop_lock.

Fix by cancelling the timers before acquiring the locks. The timer
callbacks check for state validity, so a late cancel is safe. The
queue splice is still done under the lock for consistency.

Fixes: 4b3faf610cc6 ("xfrm: iptfs: add new iptfs xfrm mode impl")
Signed-off-by: Dudu Lu <phx0fer@gmail.com>
---
 net/xfrm/xfrm_iptfs.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/net/xfrm/xfrm_iptfs.c b/net/xfrm/xfrm_iptfs.c
index 97bc979e55ba..11291b87158c 100644
--- a/net/xfrm/xfrm_iptfs.c
+++ b/net/xfrm/xfrm_iptfs.c
@@ -2708,8 +2708,10 @@ static void iptfs_destroy_state(struct xfrm_state *x)
 	if (!xtfs)
 		return;
 
-	spin_lock_bh(&xtfs->x->lock);
 	hrtimer_cancel(&xtfs->iptfs_timer);
+	hrtimer_cancel(&xtfs->drop_timer);
+
+	spin_lock_bh(&xtfs->x->lock);
 	__skb_queue_head_init(&list);
 	skb_queue_splice_init(&xtfs->queue, &list);
 	spin_unlock_bh(&xtfs->x->lock);
@@ -2717,9 +2719,7 @@ static void iptfs_destroy_state(struct xfrm_state *x)
 	while ((skb = __skb_dequeue(&list)))
 		kfree_skb(skb);
 
-	spin_lock_bh(&xtfs->drop_lock);
-	hrtimer_cancel(&xtfs->drop_timer);
-	spin_unlock_bh(&xtfs->drop_lock);
+	/* drop_timer already cancelled above */
 
 	if (xtfs->ra_newskb)
 		kfree_skb(xtfs->ra_newskb);
-- 
2.39.3 (Apple Git-145)

Re: [PATCH] xfrm: iptfs: fix deadlock in iptfs_destroy_state

[Simon Horman]

From: 'Simon Horman' <horms@kernel.org>

This is an AI-generated review of your patch. The human sending this
email has considered the AI review valid, or at least plausible.
Full review at: https://sashiko.dev
---
xfrm: iptfs: fix deadlock in iptfs_destroy_state

This commit resolves an ABBA deadlock in iptfs_destroy_state() by moving
the cancellation of iptfs_timer and drop_timer outside of their
respective spinlocks.

> diff --git a/net/xfrm/xfrm_iptfs.c b/net/xfrm/xfrm_iptfs.c
> index 97bc979e55baf..11291b87158cb 100644
> --- a/net/xfrm/xfrm_iptfs.c
> +++ b/net/xfrm/xfrm_iptfs.c
> @@ -2708,8 +2708,10 @@ static void iptfs_destroy_state(struct xfrm_state *x)
>  	if (!xtfs)
>  		return;
>  
> -	spin_lock_bh(&xtfs->x->lock);
>  	hrtimer_cancel(&xtfs->iptfs_timer);
> +	hrtimer_cancel(&xtfs->drop_timer);
> +
> +	spin_lock_bh(&xtfs->x->lock);
>  	__skb_queue_head_init(&list);
>  	skb_queue_splice_init(&xtfs->queue, &list);
>  	spin_unlock_bh(&xtfs->x->lock);

Is it safe to call hrtimer_cancel() and skb_queue_splice_init() on xtfs
here if it was created via iptfs_clone_state() but failed initialization?

If iptfs_clone_state() uses kmemdup() to shallow-copy the xtfs structure
from the original state, the copy includes xtfs->queue and the hrtimer
structs (iptfs_timer and drop_timer). Reinitialization only happens later
in iptfs_init_state() -> __iptfs_init_state().

If x->type->init_state() fails during __xfrm_init_state(), the state is
destroyed via xfrm_state_put() before iptfs_init_state() is called.

On this destruction path, does calling hrtimer_cancel() result in a
regression where it operates on the copied timers? If the original timer
was queued, the copied timer has the HRTIMER_STATE_ENQUEUED bit set, but
its rb_node is not natively in the rbtree. Calling hrtimer_cancel() might
call rb_erase_cached() using the cloned node's pointers (which point to
the original node's parent/children).

Additionally, since xtfs->queue is a shallow copy, its next/prev pointers
would point to the original state's queue. Could splicing it here lead to
a regression by modifying the original state's skb queue? If the original
queue was empty, it points to itself, and this code might dequeue it and
call kfree_skb() on an address inside orig->mode_data.