[PATCH net] sctp: avoid livelock while updating retransmit path

sctp_assoc_update_retran_path() walks the association transport list from
the current retransmit path's successor and stops once it reaches the
current retransmit path again. However, the loop skips transports in
SCTP_UNCONFIRMED state before checking for the wraparound condition.

This makes the loop non-terminating when the association contains only
UNCONFIRMED transports at that point and asoc->peer.retran_path is also
UNCONFIRMED. One way to reach that state is through ASCONF wildcard
DEL-IP processing after an unconfirmed address is selected as the
primary transport. sctp_assoc_del_nonprimary_peers() then removes the
other transports one by one; when removing the current retran_path,
sctp_assoc_rm_peer() calls sctp_assoc_update_retran_path() before
unlinking it. If the remaining candidate and the current retran_path are
both UNCONFIRMED, the loop repeatedly continues before it can observe
that it has completed a full pass.

The same reproducer that exercised the bug fixed by commit 9b2854f86f0b
("sctp: don't free the ASCONF's own transport in DEL-IP processing") can
still trigger this CPU stall after that fix is applied. With the UAF
prevented, the ASCONF processing no longer dereferences the freed
transport, but it can still reach the retransmit-path update described
above and spin in the all-UNCONFIRMED case.

Fix this by remembering whether the current transport is the original
retran_path, still considering it as a candidate when it is not
UNCONFIRMED, and then breaking after the candidate logic. This preserves
the existing fallback semantics while making the full-pass termination
independent of the transport state.

Also restore the NULL guard around the retran_path assignment. In the
all-UNCONFIRMED case there is no eligible replacement transport, and
installing NULL would leave later retransmit-path users and the debug
print with a NULL path.

Fixes: 4c47af4d5eb2 ("net: sctp: rework multihoming retransmission path selection to rfc4960")
Signed-off-by: Yiqi Sun <redacted>
---
 net/sctp/associola.c | 19 +++++++++++--------
 1 file changed, 11 insertions(+), 8 deletions(-)

diff --git a/net/sctp/associola.c b/net/sctp/associola.c
index c0512c827d0f..6f19eb0b01e2 100644
--- a/net/sctp/associola.c
+++ b/net/sctp/associola.c

@@ -1272,6 +1272,7 @@ void sctp_assoc_update_retran_path(struct sctp_association *asoc)
 {
 	struct sctp_transport *trans = asoc->peer.retran_path;
 	struct sctp_transport *trans_next = NULL;
+	bool last = false;
 
 	/* We're done as we only have the one and only path. */
 	if (asoc->peer.transport_count == 1)

@@ -1289,18 +1290,20 @@ void sctp_assoc_update_retran_path(struct sctp_association *asoc)
 		/* Manually skip the head element. */
 		if (&trans->transports == &asoc->peer.transport_addr_list)
 			continue;
-		if (trans->state == SCTP_UNCONFIRMED)
-			continue;
-		trans_next = sctp_trans_elect_best(trans, trans_next);
-		/* Active is good enough for immediate return. */
-		if (trans_next->state == SCTP_ACTIVE)
-			break;
+		last = trans == asoc->peer.retran_path;
+		if (trans->state != SCTP_UNCONFIRMED) {
+			trans_next = sctp_trans_elect_best(trans, trans_next);
+			/* Active is good enough for immediate return. */
+			if (trans_next->state == SCTP_ACTIVE)
+				break;
+		}
 		/* We've reached the end, time to update path. */
-		if (trans == asoc->peer.retran_path)
+		if (last)
 			break;
 	}
 
-	asoc->peer.retran_path = trans_next;
+	if (trans_next)
+		asoc->peer.retran_path = trans_next;
 
 	pr_debug("%s: association:%p updated new path to addr:%pISpc\n",
 		 __func__, asoc, &asoc->peer.retran_path->ipaddr.sa);

-- 
2.34.1

On Thu, Aug 27, 2026 at 3:50 AM Yiqi Sun [off-list ref] wrote:

sctp_assoc_update_retran_path() walks the association transport list from
the current retransmit path's successor and stops once it reaches the
current retransmit path again. However, the loop skips transports in
SCTP_UNCONFIRMED state before checking for the wraparound condition.

This makes the loop non-terminating when the association contains only
UNCONFIRMED transports at that point and asoc->peer.retran_path is also
UNCONFIRMED. One way to reach that state is through ASCONF wildcard
DEL-IP processing after an unconfirmed address is selected as the
primary transport. sctp_assoc_del_nonprimary_peers() then removes the
other transports one by one; when removing the current retran_path,
sctp_assoc_rm_peer() calls sctp_assoc_update_retran_path() before
unlinking it. If the remaining candidate and the current retran_path are
both UNCONFIRMED, the loop repeatedly continues before it can observe
that it has completed a full pass.

The same reproducer that exercised the bug fixed by commit 9b2854f86f0b
("sctp: don't free the ASCONF's own transport in DEL-IP processing") can
still trigger this CPU stall after that fix is applied. With the UAF
prevented, the ASCONF processing no longer dereferences the freed
transport, but it can still reach the retransmit-path update described
above and spin in the all-UNCONFIRMED case.

Please share the PoC with maintainers.

After removing the continue, I think you can keep using
if (trans == asoc->peer.retran_path) here without 'last' needed.

Thanks.

        }

-       asoc->peer.retran_path = trans_next;
+       if (trans_next)
+               asoc->peer.retran_path = trans_next;

        pr_debug("%s: association:%p updated new path to addr:%pISpc\n",
                 __func__, asoc, &asoc->peer.retran_path->ipaddr.sa);
--
2.34.1

On Thu, Aug 27, 2026 at 3:50 AM, Xin Long wrote:

After removing the continue, I think you can keep using
if (trans == asoc->peer.retran_path) here without 'last' needed.

Yes. The v1 'last' variable was redundant once the SCTP_UNCONFIRMED
path no longer uses continue. Drop it in this revision and retain the
original wraparound comparison after the candidate-selection block.

The reproducer is attached.

sctp_assoc_update_retran_path() walks the association transport list
from the current retransmit path's successor and stops once it reaches
the current retransmit path again. However, the loop skips transports in
SCTP_UNCONFIRMED state before checking for the wraparound condition.

This makes the loop non-terminating when the association contains only
UNCONFIRMED transports at that point and asoc->peer.retran_path is also
UNCONFIRMED. One way to reach that state is through ASCONF wildcard
DEL-IP processing after an unconfirmed address is selected as the
primary transport. sctp_assoc_del_nonprimary_peers() then removes the
other transports one by one; when removing the current retran_path,
sctp_assoc_rm_peer() calls sctp_assoc_update_retran_path() before
unlinking it. If the remaining candidate and the current retran_path are
both UNCONFIRMED, the loop repeatedly continues before it can observe
that it has completed a full pass.

Fix this by considering a transport only when it is not UNCONFIRMED,
then checking whether the walk has returned to retran_path. This makes
the full-pass termination independent of the transport state while
preserving the existing fallback selection semantics.

Also restore the NULL guard around the retran_path assignment. In the
all-UNCONFIRMED case there is no eligible replacement transport, and
installing NULL would leave later retransmit-path users and the debug
print with a NULL path.

Fixes: 4c47af4d5eb2 ("net: sctp: rework multihoming retransmission path selection to rfc4960")
Signed-off-by: Yiqi Sun <redacted>
---
Changes in v2:
- Drop the redundant 'last' variable as suggested by Xin Long.
- Link to v1: https://lore.kernel.org/r/20260827075006.3979566-1-sunyiqixm@gmail.com/
---
 net/sctp/associola.c | 15 ++++++++-------
 1 file changed, 8 insertions(+), 7 deletions(-)

diff --git a/net/sctp/associola.c b/net/sctp/associola.c
index c0512c827d0f..4521be3bd85a 100644
--- a/net/sctp/associola.c
+++ b/net/sctp/associola.c

@@ -1289,18 +1289,19 @@ void sctp_assoc_update_retran_path(struct sctp_association *asoc)
 		/* Manually skip the head element. */
 		if (&trans->transports == &asoc->peer.transport_addr_list)
 			continue;
-		if (trans->state == SCTP_UNCONFIRMED)
-			continue;
-		trans_next = sctp_trans_elect_best(trans, trans_next);
-		/* Active is good enough for immediate return. */
-		if (trans_next->state == SCTP_ACTIVE)
-			break;
+		if (trans->state != SCTP_UNCONFIRMED) {
+			trans_next = sctp_trans_elect_best(trans, trans_next);
+			/* Active is good enough for immediate return. */
+			if (trans_next->state == SCTP_ACTIVE)
+				break;
+		}
 		/* We've reached the end, time to update path. */
 		if (trans == asoc->peer.retran_path)
 			break;
 	}
 
-	asoc->peer.retran_path = trans_next;
+	if (trans_next)
+		asoc->peer.retran_path = trans_next;
 
 	pr_debug("%s: association:%p updated new path to addr:%pISpc\n",
 		 __func__, asoc, &asoc->peer.retran_path->ipaddr.sa);

-- 
2.34.1

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

Critical: 0 · High: 0 · Medium: 1 · Low: 0

- [Medium] The permanent part of the commit message (everything above the
  `---` line) is not a changelog.

--

sctp: avoid livelock while updating retransmit path

sctp_assoc_update_retran_path() could loop forever when every remaining
transport, including the current retran_path, was in SCTP_UNCONFIRMED
state, because the state check ran before the wraparound test.  The patch
moves the candidate selection under a state test so the wraparound
comparison is always reached, and restores the NULL guard before assigning
asoc->peer.retran_path.

Link: https://lore.kernel.org/r/20260827075006.3979566-1-sunyiqixm@gmail.com/

The code change itself looks fine to me; my only comments are about the
commit message.

sctp: avoid livelock while updating retransmit path

--jv3lfyjkrnvkybth
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline

On Thu, Aug 27, 2026 at 3:50 AM, Xin Long wrote:

quoted

After removing the continue, I think you can keep using
if (trans == asoc->peer.retran_path) here without 'last' needed.

Yes. The v1 'last' variable was redundant once the SCTP_UNCONFIRMED
path no longer uses continue. Drop it in this revision and retain the
original wraparound comparison after the candidate-selection block.

The reproducer is attached.

[Severity: Medium]
Was this text meant to end up above the --- separator?  Everything above

--- is what git am records permanently, so the MIME boundary and headers:

    --jv3lfyjkrnvkybth
    Content-Type: text/plain; charset=us-ascii
    Content-Disposition: inline

get committed verbatim into the changelog.

The reply to the reviewer and the v1/v2 discussion ("The v1 'last'
variable was redundant once the SCTP_UNCONFIRMED path no longer uses
continue. Drop it in this revision") also read as thread material rather
than changelog material, and the same information is already recorded in
the "Changes in v2" block below the --- marker.

Can the statement "The reproducer is attached." be dropped or replaced?
Nothing is attached to the patch, and there is no selftest or
Reported-by/Closes tag pointing at one, so a reader of
git log net/sctp/associola.c has no way to reach the described ASCONF
wildcard DEL-IP scenario.

Would you mind respinning with only the last three paragraphs (the
livelock description, the fix description, and the NULL guard note) plus
the Fixes and Signed-off-by tags above the --- line?

-- 
Sashiko AI review · https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20260902025206.phbpyxmpf4zrtdpx%40sunyiqi-llm-kernel

On Tue, Sep 1, 2026 at 10:53 PM Yiqi Sun [off-list ref] wrote:

Hi Yiqi,

Could you repost the patch in a separate thread without including
the discussion below in the commit message?


    On Thu, Aug 27, 2026 at 3:50 AM, Xin Long wrote:
    > After removing the continue, I think you can keep using
    > if (trans == asoc->peer.retran_path) here without 'last' needed.

    Yes. The v1 'last' variable was redundant once the SCTP_UNCONFIRMED
    path no longer uses continue. Drop it in this revision and retain the
    original wraparound comparison after the candidate-selection block.

Thanks.