Thread (37 messages) 37 messages, 4 authors, 2025-10-11

Re: [PATCH net-next v5 3/8] TUN, TAP & vhost_net: Stop netdev queue before reaching a full ptr_ring

From: "Michael S. Tsirkin" <mst@redhat.com>
Date: 2025-09-23 14:48:00
Also in: kvm, lkml, virtualization

On Tue, Sep 23, 2025 at 12:15:48AM +0200, Simon Schippers wrote:
quoted hunk ↗ jump to hunk
Stop the netdev queue ahead of __ptr_ring_produce when
__ptr_ring_full_next signals the ring is about to fill. Due to the
smp_wmb() of __ptr_ring_produce the consumer is guaranteed to be able to
notice the stopped netdev queue after seeing the new ptr_ring entry. As
both __ptr_ring_full_next and __ptr_ring_produce need the producer_lock,
the lock is held during the execution of both methods.

dev->lltx is disabled to ensure that tun_net_xmit is not called even
though the netdev queue is stopped (which happened in my testing,
resulting in rare packet drops). Consequently, the update of trans_start
in tun_net_xmit is also removed.

Co-developed-by: Tim Gebauer <redacted>
Signed-off-by: Tim Gebauer <redacted>
Signed-off-by: Simon Schippers <redacted>
---
 drivers/net/tun.c | 16 ++++++++++------
 1 file changed, 10 insertions(+), 6 deletions(-)
diff --git a/drivers/net/tun.c b/drivers/net/tun.c
index 86a9e927d0ff..c6b22af9bae8 100644
--- a/drivers/net/tun.c
+++ b/drivers/net/tun.c
@@ -931,7 +931,7 @@ static int tun_net_init(struct net_device *dev)
 	dev->vlan_features = dev->features &
 			     ~(NETIF_F_HW_VLAN_CTAG_TX |
 			       NETIF_F_HW_VLAN_STAG_TX);
-	dev->lltx = true;
+	dev->lltx = false;
 
 	tun->flags = (tun->flags & ~TUN_FEATURES) |
 		      (ifr->ifr_flags & TUN_FEATURES);
@@ -1060,14 +1060,18 @@ static netdev_tx_t tun_net_xmit(struct sk_buff *skb, struct net_device *dev)
 
 	nf_reset_ct(skb);
 
-	if (ptr_ring_produce(&tfile->tx_ring, skb)) {
+	queue = netdev_get_tx_queue(dev, txq);
+
+	spin_lock(&tfile->tx_ring.producer_lock);
+	if (__ptr_ring_full_next(&tfile->tx_ring))
+		netif_tx_stop_queue(queue);
+
+	if (unlikely(__ptr_ring_produce(&tfile->tx_ring, skb))) {
+		spin_unlock(&tfile->tx_ring.producer_lock);
 		drop_reason = SKB_DROP_REASON_FULL_RING;
 		goto drop;
 	}
The comment makes it sound like you always keep one slot free
in the queue but that is not the case - you just
check before calling __ptr_ring_produce.


But it is racy isn't it? So first of all I suspect you
are missing an mb before netif_tx_stop_queue.

Second it's racy because more entries can get freed
afterwards. Which maybe is ok in this instance?
But it really should be explained in more detail, if so.



Now - why not just check ring full *after* __ptr_ring_produce?
Why do we need all these new APIs, and we can
use existing ones which at least are not so hard to understand.



-
-	/* dev->lltx requires to do our own update of trans_start */
-	queue = netdev_get_tx_queue(dev, txq);
-	txq_trans_cond_update(queue);
+	spin_unlock(&tfile->tx_ring.producer_lock);
 
 	/* Notify and wake up reader process */
 	if (tfile->flags & TUN_FASYNC)
-- 
2.43.0
  
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help