On 5/7/26 08:32, Simon Schippers wrote:
> On 5/7/26 00:56, Michael S. Tsirkin wrote:
>> On Wed, May 06, 2026 at 06:28:06PM -0400, Michael S. Tsirkin wrote:
>>> On Wed, May 06, 2026 at 04:10:33PM +0200, Simon Schippers wrote:
>>>> This commit prevents tail-drop when a qdisc is present and the ptr_ring
>>>> becomes full. Once an entry is successfully produced and the ptr_ring
>>>> reaches capacity, the netdev queue is stopped instead of dropping
>>>> subsequent packets. If no qdisc is present, the previous tail-drop
>>>> behavior is preserved.
>>>>
>>>> If producing an entry fails anyways due to a race, tun_net_xmit() drops
>>>> the packet. Such races are expected because LLTX is enabled and the
>>>> transmit path operates without the usual locking.
>>>>
>>>> The __tun_wake_queue() function of the consumer races with the producer
>>>> for waking/stopping the netdev queue, which could result in a stalled
>>>> queue. Therefore, an smp_mb__after_atomic() is introduced that pairs
>>>> with the smp_mb() of the consumer. It follows the principle of store
>>>> buffering described in tools/memory-model/Documentation/recipes.txt:
>>>>
>>>> - The producer in tun_net_xmit() first sets __QUEUE_STATE_DRV_XOFF,
>>>> followed by an smp_mb__after_atomic() (= smp_mb()), and then reads the
>>>> ring with __ptr_ring_produce_peek().
>>>>
>>>> - The consumer in __tun_wake_queue() first writes zero to the ring in
>>>> __ptr_ring_consume(), followed by an smp_mb(), and then reads the queue
>>>> status with netif_tx_queue_stopped().
>>>>
>>>> => Following the aforementioned principle, it is impossible for the
>>>> producer to see a full ring (and therefore not wake the queue on the
>>>> re-check) while the consumer simultaneously fails to see a stopped
>>>> queue (and therefore also does not wake it).
>>>>
>>>> Benchmarks:
>>>> The benchmarks show a slight regression in raw transmission performance
>>>> when using two sending threads. Packet loss also occurs only in the
>>>> two-thread sending case; no packet loss was observed with a single
>>>> sending thread.
>>>>
>>>> Test setup:
>>>> AMD Ryzen 5 5600X at 4.3 GHz, 3200 MHz RAM, isolated QEMU threads;
>>>> Average over 50 runs @ 100,000,000 packets. SRSO and spectre v2
>>>> mitigations disabled.
>>>>
>>>> Note for tap+vhost-net:
>>>> XDP drop program active in VM -> ~2.5x faster; slower for tap due to
>>>> more syscalls (high utilization of entry_SYSRETQ_unsafe_stack in perf)
>>>>
>>>> +--------------------------+--------------+----------------+----------+
>>>> | 1 thread | Stock | Patched with | diff |
>>>> | sending | | fq_codel qdisc | |
>>>> +------------+-------------+--------------+----------------+----------+
>>>> | TAP | Received | 1.132 Mpps | 1.133 Mpps | +0.1% |
>>>> | +-------------+--------------+----------------+----------+
>>>> | | Lost/s | 3.765 Mpps | 0 pps | |
>>>> +------------+-------------+--------------+----------------+----------+
>>>> | TAP | Received | 3.857 Mpps | 3.905 Mpps | +1.2% |
>>>> | +-------------+--------------+----------------+----------+
>>>> | +vhost-net | Lost/s | 0.802 Mpps | 0 pps | |
>>>> +------------+-------------+--------------+----------------+----------+
>>>>
>>>> +--------------------------+--------------+----------------+----------+
>>>> | 2 threads | Stock | Patched with | diff |
>>>> | sending | | fq_codel qdisc | |
>>>> +------------+-------------+--------------+----------------+----------+
>>>> | TAP | Received | 1.115 Mpps | 1.092 Mpps | -2.1% |
>>>> | +-------------+--------------+----------------+----------+
>>>> | | Lost/s | 8.490 Mpps | 359 pps | |
>>>> +------------+-------------+--------------+----------------+----------+
>>>> | TAP | Received | 3.664 Mpps | 3.549 Mpps | -3.1% |
>>>> | +-------------+--------------+----------------+----------+
>>>> | +vhost-net | Lost/s | 5.330 Mpps | 832 pps | |
>>>> +------------+-------------+--------------+----------------+----------+
>>>>
>>>> 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 | 25 +++++++++++++++++++++++--
>>>> 1 file changed, 23 insertions(+), 2 deletions(-)
>>>>
>>>> diff --git a/drivers/net/tun.c b/drivers/net/tun.c
>>>> index fc358c4c355b..d9ffbf88cfd8 100644
>>>> --- a/drivers/net/tun.c
>>>> +++ b/drivers/net/tun.c
>>>> @@ -1018,6 +1018,7 @@ static netdev_tx_t tun_net_xmit(struct sk_buff *skb,
>>>> struct net_device *dev)
>>>> struct netdev_queue *queue;
>>>> struct tun_file *tfile;
>>>> int len = skb->len;
>>>> + int ret;
>>>>
>>>> rcu_read_lock();
>>>> tfile = rcu_dereference(tun->tfiles[txq]);
>>>> @@ -1072,13 +1073,33 @@ 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);
>>>> + ret = __ptr_ring_produce(&tfile->tx_ring, skb);
>>>> + if (!qdisc_txq_has_no_queue(queue) &&
>>>> + (__ptr_ring_produce_peek(&tfile->tx_ring) || ret)) {
>>>> + netif_tx_stop_queue(queue);
>>>> + /* Paired with smp_mb() in __tun_wake_queue() */
>>>> + smp_mb__after_atomic();
>>>> + if (!__ptr_ring_produce_peek(&tfile->tx_ring))
>>>> + netif_tx_wake_queue(queue);
>>>> + }
>>>> + spin_unlock(&tfile->tx_ring.producer_lock);
>>>> +
>>>
>>> There's a weird corner case here when tx_queue_len is 0
>>> but a qdisc has been configured - it looks like that
>>> currently it just drops all packets, with this change,
>>> the qdisc will get stuck permanently.
>>>
>>> I suspect just checking tx_ring.size should fix it.
>>> Or if you feel adventurous, change return code for __ptr_ring_produce
>>> to distinguish between "no ring" and "no space".
>>
>>
>> __ptr_ring_produce_peek really.
>>
>
> Yes, I like the approach of returning this from
> __ptr_ring_produce_peek(). Then I will do a switch on the return value
> in tun_net_xmit().
Sashiko reports the same :)
So for the v11 I will:
- Change __ptr_ring_produce_peek() to return -ENOSPC / -EINVAL? (for 0
sized ring) / 0
- Lock the ring.consumer_lock in __tun_detach() to avoid a race with
consumer (Sashiko).
>
> Additionally, I should wake up in tun_queue_resize() after calling
> ptr_ring_resize_multiple_bh(). For a new dev->tx_queue_len > 0, it
> should be fine without waking, but for 0 it is not.
>
>>
>>>
>>>> + if (ret) {
>>>> + /* This should be a rare case if a qdisc is present, but
>>>> + * can happen due to lltx.
>>>> + * Since skb_tx_timestamp(), skb_orphan(),
>>>> + * run_ebpf_filter() and pskb_trim() could have tinkered
>>>> + * with the SKB, returning NETDEV_TX_BUSY is unsafe and
>>>> + * we must drop instead.
>>>> + */
>>>> drop_reason = SKB_DROP_REASON_FULL_RING;
>>>> goto drop;
>>>> }
>>>>
>>>> /* dev->lltx requires to do our own update of trans_start */
>>>> - queue = netdev_get_tx_queue(dev, txq);
>>>> txq_trans_cond_update(queue);
>>>>
>>>> /* Notify and wake up reader process */
>>>> --
>>>> 2.43.0
>>
