net/packet: tpacket_rcv: avoid a producer race condition

PACKET_RX_RING can cause multiple writers to access the same slot if a
fast writer wraps the ring while a slow writer is still copying. This
is particularly likely with few, large, slots (e.g., GSO packets).

Synchronize kernel thread ownership of rx ring slots with a bitmap.

Writers acquire a slot race-free by testing tp_status TP_STATUS_KERNEL
while holding the sk receive queue lock. They release this lock before
copying and set tp_status to TP_STATUS_USER to release to userspace
when done. During copying, another writer may take the lock, also see
TP_STATUS_KERNEL, and start writing to the same slot.

Introduce a new rx_owner_map bitmap with a bit per slot. To acquire a
slot, test and set with the lock held. To release race-free, update
tp_status and owner bit as a transaction, so take the lock again.

This is the one of a variety of discussed options (see Link below):

* instead of a shadow ring, embed the data in the slot itself, such as
in tp_padding. But any test for this field may match a value left by
userspace, causing deadlock.

* avoid the lock on release. This leaves a small race if releasing the
shadow slot before setting TP_STATUS_USER. The below reproducer showed
that this race is not academic. If releasing the slot after tp_status,
the race is more subtle. See the first link for details.

* add a new tp_status TP_KERNEL_OWNED to avoid the transactional store
of two fields. But, legacy applications may interpret all non-zero
tp_status as owned by the user. As libpcap does. So this is possible
only opt-in by newer processes. It can be added as an optional mode.

* embed the struct at the tail of pg_vec to avoid extra allocation.
The implementation proved no less complex than a separate field.

The additional locking cost on release adds contention, no different
than scaling on multicore or multiqueue h/w. In practice, below
reproducer nor small packet tcpdump showed a noticeable change in
perf report in cycles spent in spinlock. Where contention is
problematic, packet sockets support mitigation through PACKET_FANOUT.
And we can consider adding opt-in state TP_KERNEL_OWNED.

Easy to reproduce by running multiple netperf or similar TCP_STREAM
flows concurrently with `tcpdump -B 129 -n greater 60000`.

Based on an earlier patchset by Jon Rosen. See links below.

I believe this issue goes back to the introduction of tpacket_rcv,
which predates git history.

Suggested-by: Jon Rosen <>
Signed-off-by: Willem de Bruijn <>
Signed-off-by: Jon Rosen <>
Signed-off-by: David S. Miller <>
diff --git a/net/packet/af_packet.c b/net/packet/af_packet.c
index e5b0986..29bd405 100644
--- a/net/packet/af_packet.c
+++ b/net/packet/af_packet.c
@@ -2173,6 +2173,7 @@
 	struct timespec64 ts;
 	__u32 ts_status;
 	bool is_drop_n_account = false;
+	unsigned int slot_id = 0;
 	bool do_vnet = false;
 	/* struct tpacket{2,3}_hdr is aligned to a multiple of TPACKET_ALIGNMENT.
@@ -2275,6 +2276,13 @@
 	if (!h.raw)
 		goto drop_n_account;
+	if (po->tp_version <= TPACKET_V2) {
+		slot_id = po->rx_ring.head;
+		if (test_bit(slot_id, po->rx_ring.rx_owner_map))
+			goto drop_n_account;
+		__set_bit(slot_id, po->rx_ring.rx_owner_map);
+	}
 	if (do_vnet &&
 	    virtio_net_hdr_from_skb(skb, h.raw + macoff -
 				    sizeof(struct virtio_net_hdr),
@@ -2380,7 +2388,10 @@
 	if (po->tp_version <= TPACKET_V2) {
+		spin_lock(&sk->sk_receive_queue.lock);
 		__packet_set_status(po, h.raw, status);
+		__clear_bit(slot_id, po->rx_ring.rx_owner_map);
+		spin_unlock(&sk->sk_receive_queue.lock);
 	} else {
@@ -4277,6 +4288,7 @@
 	struct pgv *pg_vec = NULL;
 	struct packet_sock *po = pkt_sk(sk);
+	unsigned long *rx_owner_map = NULL;
 	int was_running, order = 0;
 	struct packet_ring_buffer *rb;
 	struct sk_buff_head *rb_queue;
@@ -4362,6 +4374,12 @@
+			if (!tx_ring) {
+				rx_owner_map = bitmap_alloc(req->tp_frame_nr,
+				if (!rx_owner_map)
+					goto out_free_pg_vec;
+			}
@@ -4391,6 +4409,8 @@
 		err = 0;
 		swap(rb->pg_vec, pg_vec);
+		if (po->tp_version <= TPACKET_V2)
+			swap(rb->rx_owner_map, rx_owner_map);
 		rb->frame_max = (req->tp_frame_nr - 1);
 		rb->head = 0;
 		rb->frame_size = req->tp_frame_size;
@@ -4422,6 +4442,7 @@
+	bitmap_free(rx_owner_map);
 	if (pg_vec)
 		free_pg_vec(pg_vec, order, req->tp_block_nr);
diff --git a/net/packet/internal.h b/net/packet/internal.h
index 82fb2b1..907f4cd 100644
--- a/net/packet/internal.h
+++ b/net/packet/internal.h
@@ -70,7 +70,10 @@
 	unsigned int __percpu	*pending_refcnt;
-	struct tpacket_kbdq_core	prb_bdqc;
+	union {
+		unsigned long			*rx_owner_map;
+		struct tpacket_kbdq_core	prb_bdqc;
+	};
 extern struct mutex fanout_mutex;