On Wed, Sep 24, 2025 at 07:41:28AM +0200, Simon Schippers wrote: > Hi, > first of all thank you very much for your detailed replies! :) > > On 23.09.25 16:47, Michael S. Tsirkin wrote: > > On Tue, Sep 23, 2025 at 12:15:48AM +0200, Simon Schippers wrote: > >> 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 <[email protected]> > >> Signed-off-by: Tim Gebauer <[email protected]> > >> Signed-off-by: Simon Schippers <[email protected]> > >> --- > >> 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. > > > > I agree. > > > > > But it is racy isn't it? So first of all I suspect you > > are missing an mb before netif_tx_stop_queue. > > > > I don’t really get this point right now.
ring full next is a read. stop queue is a write. if you are relying on ordering them in some way you need a full mb generally. > > 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. > > > > Will be covered in the next mail. > > > > > > > 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. > > > > > > You convinced me about changing my implementation anyway but here my (old) > idea: > I did this in V1-V4. The problem is that vhost_net is only called on > EPOLLIN triggered by tun_net_xmit. Then, after consuming a batch from the > ptr_ring, it must be able to see if the netdev queue stopped or not. If > this is not the case the ptr_ring might get empty and vhost_net is not > able to wake the queue again (because it is not stopped from its POV), > which happened in my testing in my V4. > > This is the reason why, now in the V5, in tun_net_xmit I stop the netdev > queue before producing. With that I exploit the smp_wmb() in > __ptr_ring_produce which is paired with the READ_ONCE in __ptr_ring_peek > to ensure that the consumer in vhost_net sees that the netdev queue > stopped after consuming a batch. yea you said it somewhere in code, too, and I am not sure I understand it all, but wmb isn't paired with READ_ONCE generally. barrier pairing is described in memory-barriers.txt, READ_ONCE is not a barrier at all. > > > > > >> - > >> - /* 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 > >
