Netdev List
 help / color / mirror / Atom feed
From: Minxi Hou <houminxi@gmail.com>
To: netdev@vger.kernel.org
Cc: Aaron Conole <aconole@redhat.com>,
	Eelco Chaudron <echaudro@redhat.com>,
	Ilya Maximets <i.maximets@ovn.org>,
	"David S . Miller" <davem@davemloft.net>,
	Eric Dumazet <edumazet@kernel.org>,
	Jakub Kicinski <kuba@kernel.org>, Paolo Abeni <pabeni@redhat.com>,
	Simon Horman <horms@kernel.org>, Shuah Khan <shuah@kernel.org>,
	linux-kselftest@vger.kernel.org, stable@vger.kernel.org,
	Minxi Hou <houminxi@gmail.com>
Subject: [PATCH net v5 1/2] net: openvswitch: do not record user features when the change fails
Date: Wed, 18 Nov 2026 01:54:34 -0500	[thread overview]
Message-ID: <20261118065435.2367014-2-houminxi@gmail.com> (raw)
In-Reply-To: <20261118065435.2367014-1-houminxi@gmail.com>

ovs_dp_change() stores the new features in dp->user_features before it
updates the per-cpu upcall port IDs, and that update fails when the
per-cpu-pids attribute is empty or its length is not a multiple of four.
The request returns -EINVAL with the new features already recorded, while
the code that enables or disables the tc skb extension key never ran.

The datapath then claims TC_RECIRC_SHARING without holding the reference
the bit stands for. A later request that clears the bit drops a
reference that was never taken, and a datapath destroyed with the bit
clear never drops the one it did take. The key is shared with the other
users of recirc sharing, so the count can reach zero while one of them
still reads it, and a disable on a count of zero warns from
static_key_dec_not_one(), which panics a host that runs with
panic_on_warn.

The request only needs CAP_NET_ADMIN, which a user namespace that owns a
network namespace provides. datapath_policy has no entry for
OVS_DP_ATTR_PER_CPU_PIDS and the operation skips strict validation, so an
empty attribute reaches the check that fails.

Record the features last, in one setter that also takes and drops the
reference, and use it from every place that changes or clears the bit.
Two of those places leaked a reference: a datapath that fails after
ovs_dp_change() enabled the key was freed without a disable, and
ovs_dp_reset_user_features() cleared the bit without one. The reset runs
when an outdated user space instance sends OVS_DP_CMD_NEW for a name that
already exists, and after it the datapath's packets no longer share the tc
skb extension.

Fixes: 35d39fecbc24 ("net/sched: Enable tc skb ext allocation on chain miss only when needed")
Cc: stable@vger.kernel.org
__dp_destroy() clears the bit before the ports are detached and before the
datapath leaves the list, so an upcall can still be reading the other bits
from the receive path. queue_userspace_packet() decides from
OVS_DP_F_UNALIGNED whether to build a nonlinear skb, and pad_packet() reads
the same field again, so a store that clears every bit can change that
decision between the two reads. Clear only the recirc bit, which is the one
this path dropped before.

The store goes through WRITE_ONCE() because the receive path reads the
field without ovs_mutex.

The receive path read the field twice as well, once to size the head room
and once to decide whether to pad, so a set that changed the bit in
between could make it pad a nonlinear skb. It reads the field once now and
uses that value for both.

Signed-off-by: Minxi Hou <houminxi@gmail.com>
---
 net/openvswitch/datapath.c | 58 +++++++++++++++++++++++++++-----------
 net/openvswitch/datapath.h |  7 ++++-
 2 files changed, 47 insertions(+), 18 deletions(-)

diff --git a/net/openvswitch/datapath.c b/net/openvswitch/datapath.c
index b2a0b09fd4a5..7c1af21f74df 100644
--- a/net/openvswitch/datapath.c
+++ b/net/openvswitch/datapath.c
@@ -439,9 +439,9 @@ static size_t upcall_msg_size(const struct dp_upcall_info *upcall_info,
 	return size;
 }
 
-static void pad_packet(struct datapath *dp, struct sk_buff *skb)
+static void pad_packet(struct sk_buff *skb, u32 features)
 {
-	if (!(dp->user_features & OVS_DP_F_UNALIGNED)) {
+	if (!(features & OVS_DP_F_UNALIGNED)) {
 		size_t plen = NLA_ALIGN(skb->len) - skb->len;
 
 		if (plen > 0)
@@ -461,6 +461,7 @@ static int queue_userspace_packet(struct datapath *dp, struct sk_buff *skb,
 	size_t msg_size;
 	size_t skb_len;
 	unsigned int hlen;
+	u32 features;
 	int err, dp_ifindex;
 	u64 hash;
 
@@ -494,11 +495,16 @@ static int queue_userspace_packet(struct datapath *dp, struct sk_buff *skb,
 	    (err = skb_csum_hwoffload_help(skb, 0)))
 		goto out;
 
+	/* The features can be changed while this runs, so read them once and
+	 * use that value for both the head room and the padding.
+	 */
+	features = READ_ONCE(dp->user_features);
+
 	/* Older versions of OVS user space enforce alignment of the last
 	 * Netlink attribute to NLA_ALIGNTO which would require extensive
 	 * padding logic. Only perform zerocopy if padding is not required.
 	 */
-	if (dp->user_features & OVS_DP_F_UNALIGNED)
+	if (features & OVS_DP_F_UNALIGNED)
 		hlen = min(skb_zerocopy_headlen(skb), cutlen);
 	else
 		hlen = skb_len;
@@ -598,7 +604,7 @@ static int queue_userspace_packet(struct datapath *dp, struct sk_buff *skb,
 		goto out;
 
 	/* Pad OVS_PACKET_ATTR_PACKET if linear copy was performed */
-	pad_packet(dp, user_skb);
+	pad_packet(user_skb, features);
 
 	((struct nlmsghdr *) user_skb->data)->nlmsg_len = user_skb->len;
 
@@ -1704,6 +1710,24 @@ static struct datapath *lookup_datapath(struct net *net,
 	return dp ? dp : ERR_PTR(-ENODEV);
 }
 
+/* Called with ovs_mutex. The receive path reads the field without it, so
+ * the store is single and goes through WRITE_ONCE(). The store happens
+ * before the reference is taken or dropped, the order ovs_dp_change() used
+ * before this helper existed.
+ */
+static void ovs_dp_set_user_features(struct datapath *dp, u32 features)
+{
+	bool was = dp->user_features & OVS_DP_F_TC_RECIRC_SHARING;
+	bool now = features & OVS_DP_F_TC_RECIRC_SHARING;
+
+	WRITE_ONCE(dp->user_features, features);
+
+	if (now && !was)
+		tc_skb_ext_tc_enable();
+	else if (!now && was)
+		tc_skb_ext_tc_disable();
+}
+
 static void ovs_dp_reset_user_features(struct sk_buff *skb,
 				       struct genl_info *info)
 {
@@ -1716,7 +1740,7 @@ static void ovs_dp_reset_user_features(struct sk_buff *skb,
 
 	pr_warn("%s: Dropping previously announced user features\n",
 		ovs_dp_name(dp));
-	dp->user_features = 0;
+	ovs_dp_set_user_features(dp, 0);
 }
 
 static int ovs_dp_set_upcall_portids(struct datapath *dp,
@@ -1773,7 +1797,7 @@ u32 ovs_dp_get_upcall_portid(const struct datapath *dp, uint32_t cpu_id)
 
 static int ovs_dp_change(struct datapath *dp, struct nlattr *a[])
 {
-	u32 user_features = 0, old_features = dp->user_features;
+	u32 user_features = 0;
 	int err;
 
 	if (a[OVS_DP_ATTR_USER_FEATURES]) {
@@ -1801,9 +1825,7 @@ static int ovs_dp_change(struct datapath *dp, struct nlattr *a[])
 			return err;
 	}
 
-	dp->user_features = user_features;
-
-	if (dp->user_features & OVS_DP_F_DISPATCH_UPCALL_PER_CPU &&
+	if (user_features & OVS_DP_F_DISPATCH_UPCALL_PER_CPU &&
 	    a[OVS_DP_ATTR_PER_CPU_PIDS]) {
 		/* Upcall Netlink Port IDs have been updated */
 		err = ovs_dp_set_upcall_portids(dp,
@@ -1812,12 +1834,7 @@ static int ovs_dp_change(struct datapath *dp, struct nlattr *a[])
 			return err;
 	}
 
-	if ((dp->user_features & OVS_DP_F_TC_RECIRC_SHARING) &&
-	    !(old_features & OVS_DP_F_TC_RECIRC_SHARING))
-		tc_skb_ext_tc_enable();
-	else if (!(dp->user_features & OVS_DP_F_TC_RECIRC_SHARING) &&
-		 (old_features & OVS_DP_F_TC_RECIRC_SHARING))
-		tc_skb_ext_tc_disable();
+	ovs_dp_set_user_features(dp, user_features);
 
 	return 0;
 }
@@ -1939,6 +1956,9 @@ static int ovs_dp_cmd_new(struct sk_buff *skb, struct genl_info *info)
 err_destroy_portids:
 	kfree(rcu_dereference_raw(dp->upcall_portids));
 err_unlock_and_destroy_meters:
+	/* The datapath is freed here and never reaches __dp_destroy(). */
+	ovs_dp_set_user_features(dp, dp->user_features &
+				     ~OVS_DP_F_TC_RECIRC_SHARING);
 	ovs_unlock();
 	ovs_meters_exit(dp);
 err_destroy_ports:
@@ -1961,8 +1981,12 @@ static void __dp_destroy(struct datapath *dp)
 	struct flow_table *table = &dp->table;
 	int i;
 
-	if (dp->user_features & OVS_DP_F_TC_RECIRC_SHARING)
-		tc_skb_ext_tc_disable();
+	/* Only the recirc bit goes here. The datapath is still on the list and
+	 * its ports are still attached, so an upcall can be reading the other
+	 * bits from the receive path while this runs.
+	 */
+	ovs_dp_set_user_features(dp, dp->user_features &
+				     ~OVS_DP_F_TC_RECIRC_SHARING);
 
 	for (i = 0; i < DP_VPORT_HASH_BUCKETS; i++) {
 		struct vport *vport;
diff --git a/net/openvswitch/datapath.h b/net/openvswitch/datapath.h
index 7a7afaeceee5..55db1e635c74 100644
--- a/net/openvswitch/datapath.h
+++ b/net/openvswitch/datapath.h
@@ -76,7 +76,12 @@ struct dp_nlsk_pids {
  * ovs_mutex and RCU.
  * @stats_percpu: Per-CPU datapath statistics.
  * @net: Reference to net namespace.
- * @user_features: Bitmap of enabled %OVS_DP_F_* features.
+ * @user_features: Bitmap of enabled %OVS_DP_F_* features.  A datapath
+ * that has %OVS_DP_F_TC_RECIRC_SHARING set holds one reference to the tc
+ * skb extension.  ovs_dp_set_user_features() keeps the bit and the
+ * reference in step, and every change goes through it with ovs_mutex
+ * held.  ovs_dp_change() calls it after the per-cpu pid update, which can
+ * fail, so a request that fails leaves the features alone.
  * @max_headroom: The maximum headroom of all vports in this datapath; it will
  * be used by all the internal vports in this dp.
  * @meter_tbl: Meter table.
-- 
2.56.1


  reply	other threads:[~2026-11-18  6:54 UTC|newest]

Thread overview: 3+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2026-11-18  6:54 [PATCH net v5 0/2] net: openvswitch: fix a feature leak on a failed change Minxi Hou
2026-11-18  6:54 ` Minxi Hou [this message]
2026-11-18  6:54 ` [PATCH net v5 2/2] selftests/net/openvswitch: check that a failed set keeps the features Minxi Hou

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20261118065435.2367014-2-houminxi@gmail.com \
    --to=houminxi@gmail.com \
    --cc=aconole@redhat.com \
    --cc=davem@davemloft.net \
    --cc=echaudro@redhat.com \
    --cc=edumazet@kernel.org \
    --cc=horms@kernel.org \
    --cc=i.maximets@ovn.org \
    --cc=kuba@kernel.org \
    --cc=linux-kselftest@vger.kernel.org \
    --cc=netdev@vger.kernel.org \
    --cc=pabeni@redhat.com \
    --cc=shuah@kernel.org \
    --cc=stable@vger.kernel.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox