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

[Wang Liang <wangliang74@huawei.com>]

在 2025/4/1 14:57, Magnus Karlsson 写道:
> On Tue, 1 Apr 2025 at 04:36, Wang Liang <wangliang74@huawei.com> wrote:
>>
>> 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++;
> Sorry, but I do not understand how to reproduce this error. So you
> first issue a setsockopt with the XDP_TX_RING option and then you do
> not "reserve tx ring". What does that last "not reserve tx ring" mean?
> No mmap() of that ring, or something else? I guess you have bound the
> socket with a bind()? Some pseudo code on how to reproduce this would
> be helpful. Just want to understand so I can help. Thank you.
>
Ok. Some pseudo code like below: fd = socket(AF_XDP, SOCK_RAW, 0); 
setsockopt(fd, SOL_XDP, XDP_UMEM_REG, &mr, sizeof(mr)); setsockopt(fd, 
SOL_XDP, XDP_UMEM_FILL_RING, &fill_size, sizeof(fill_size)); 
setsockopt(fd, SOL_XDP, XDP_UMEM_COMPLETION_RING, &comp_size, 
sizeof(comp_size)); mmap(NULL, off.fr.desc + fill_size * sizeof(__u64), 
..., XDP_UMEM_PGOFF_FILL_RING); mmap(NULL, off.cr.desc + comp_size * 
sizeof(__u64), ..., XDP_UMEM_PGOFF_COMPLETION_RING); setsockopt(fd, 
SOL_XDP, XDP_RX_RING, &rx_size, sizeof(rx_size)); setsockopt(fd, 
SOL_XDP, XDP_TX_RING, &tx_size, sizeof(tx_size)); mmap(NULL, off.rx.desc 
+ rx_size * sizeof(struct xdp_desc), ..., XDP_PGOFF_RX_RING); mmap(NULL, 
off.tx.desc + tx_size * sizeof(struct xdp_desc), ..., 
XDP_PGOFF_TX_RING); bind(fd, (struct sockaddr *)&sxdp, sizeof(sxdp)); 
bpf_map_update_elem(xsk_map_fd, &queue_id, &fd, 0); while(!global_exit) 
{ poll(fds, 1, -1); handle_receive_packets(...); } The xsk is created 
success, and xs->tx is initialized. The "not reserve tx ring" means user 
app do not update tx ring producer. Like: xsk_ring_prod__reserve(tx, 1, 
&tx_idx); xsk_ring_prod__tx_desc(tx, tx_idx)->addr = frame; 
xsk_ring_prod__tx_desc(tx, tx_idx)->len = pkg_length; 
xsk_ring_prod__submit(tx, 1); These functions (xsk_ring_prod__reserve, 
etc.) is provided by libxdp. The tx->producer is not updated, so the 
xs->tx->cached_cons and xs->tx->cached_prod are always zero. When 
receive packets and user app call poll(), xsk_generic_xmit() will be 
triggered by xsk_poll(), leading to this issue.