On 8/19/24 4:24 AM, Ma Ke wrote:

cxgbi_check_route() dont release the reference acquired by ip_dev_find() which introducing a reference count leak. We could remedy this by insuring the reference is released.ip_dev_find().

Cc: stable@vger.kernel.org Fixes: 9ba682f01e2f ("[SCSI] libcxgbi: common library for cxgb3i and cxgb4i") Signed-off-by: Ma Ke make24@iscas.ac.cn

---

drivers/scsi/cxgbi/libcxgbi.c | 1 + 1 file changed, 1 insertion(+)

diff --git a/drivers/scsi/cxgbi/libcxgbi.c b/drivers/scsi/cxgbi/libcxgbi.c index bf75940f2be1..6b0f1e8dac40 100644 --- a/drivers/scsi/cxgbi/libcxgbi.c +++ b/drivers/scsi/cxgbi/libcxgbi.c @@ -670,6 +670,7 @@ cxgbi_check_route(struct sockaddr *dst_addr, int ifindex) "route to %pI4 :%u, ndev p#%d,%s, cdev 0x%p.\n", &daddr->sin_addr.s_addr, ntohs(daddr->sin_port), port, ndev->name, cdev);

• dev_put(ndev);

After we call ip_dev_find can we hit one of the error path gotos like the test for IFF_UP? If so you would leak the reference in those paths.

The ndev above looks like it could come from the ip_dev_find call when IFF_LOOPBACK is set or it could come from dst_neigh_lookup call. For the dst_neigh_lookup case we don't want to do a dev_put do we? If you do, then we are leaking the reference in the goto error paths.

csk = cxgbi_sock_create(cdev); if (!csk) {