Re: [PATCH net] xsk: correct tx_ring_empty_descs count statistics

[Wang Liang <wangliang74@huawei.com>]

在 2025/4/1 6:03, Stanislav Fomichev 写道:
> On 03/31, Stanislav Fomichev wrote:
>> On 03/29, Wang Liang wrote:
>>> The tx_ring_empty_descs count may be incorrect, when set the XDP_TX_RING
>>> option but do not reserve tx ring. Because xsk_poll() try to wakeup the
>>> driver by calling xsk_generic_xmit() for non-zero-copy mode. So the
>>> tx_ring_empty_descs count increases once the xsk_poll()is called:
>>>
>>>    xsk_poll
>>>      xsk_generic_xmit
>>>        __xsk_generic_xmit
>>>          xskq_cons_peek_desc
>>>            xskq_cons_read_desc
>>>              q->queue_empty_descs++;
>>>
>>> To avoid this count error, add check for tx descs before send msg in poll.
>>>
>>> Fixes: df551058f7a3 ("xsk: Fix crash in poll when device does not support ndo_xsk_wakeup")
>>> Signed-off-by: Wang Liang <wangliang74@huawei.com>
>> Acked-by: Stanislav Fomichev <sdf@fomichev.me>
> Hmm, wait, I stumbled upon xskq_has_descs again and it looks only at
> cached prod/cons. How is it supposed to work when the actual tx
> descriptor is posted? Is there anything besides xskq_cons_peek_desc from
> __xsk_generic_xmit that refreshes cached_prod?


Yes, you are right!

How about using xskq_cons_nb_entries() to check free descriptors?

Like this:


diff --git a/net/xdp/xsk.c b/net/xdp/xsk.c
index e5d104ce7b82..babb7928d335 100644
--- a/net/xdp/xsk.c
+++ b/net/xdp/xsk.c
@@ -993,7 +993,7 @@ static __poll_t xsk_poll(struct file *file, struct 
socket *sock,
         if (pool->cached_need_wakeup) {
                 if (xs->zc)
                         xsk_wakeup(xs, pool->cached_need_wakeup);
-               else if (xs->tx)
+               else if (xs->tx && xskq_cons_nb_entries(xs->tx, 1))
                         /* Poll needs to drive Tx also in copy mode */
                         xsk_generic_xmit(sk);
         }