On 04/10, David Marchand wrote:
>On Wed, Apr 10, 2019 at 12:58 PM Xiaolong Ye <xiaolong...@intel.com> wrote:
>
>> Since tx callback only picks mbufs that can be copied to the xdp desc, it
>> should call xsk_ring_prod__submit with valid count other than nb_pkts.
>>
>> Fixes: f1debd77efaf ("net/af_xdp: introduce AF_XDP PMD")
>>
>> Reported-by: David Marchand <david.march...@redhat.com>
>> Signed-off-by: Xiaolong Ye <xiaolong...@intel.com>
>> ---
>> drivers/net/af_xdp/rte_eth_af_xdp.c | 2 +-
>> 1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/drivers/net/af_xdp/rte_eth_af_xdp.c
>> b/drivers/net/af_xdp/rte_eth_af_xdp.c
>> index 5cc643ce2..8c2ba5740 100644
>> --- a/drivers/net/af_xdp/rte_eth_af_xdp.c
>> +++ b/drivers/net/af_xdp/rte_eth_af_xdp.c
>> @@ -300,7 +300,7 @@ eth_af_xdp_tx(void *queue, struct rte_mbuf **bufs,
>> uint16_t nb_pkts)
>> rte_pktmbuf_free(mbuf);
>> }
>>
>> - xsk_ring_prod__submit(&txq->tx, nb_pkts);
>> + xsk_ring_prod__submit(&txq->tx, valid);
>>
>
>Err, well, I think we will have an issue here.
>
>Taking from the 5.1.0-rc4 sources I have:
>
>static inline __u32 xsk_prod_nb_free(struct xsk_ring_prod *r, __u32 nb)
>{
> __u32 free_entries = r->cached_cons - r->cached_prod;
>
> if (free_entries >= nb)
> return free_entries;
>
> /* Refresh the local tail pointer.
> * cached_cons is r->size bigger than the real consumer pointer so
> * that this addition can be avoided in the more frequently
> * executed code that computs free_entries in the beginning of
> * this function. Without this optimization it whould have been
> * free_entries = r->cached_prod - r->cached_cons + r->size.
> */
> r->cached_cons = *r->consumer + r->size;
>
> return r->cached_cons - r->cached_prod;
>}
>
>static inline size_t xsk_ring_prod__reserve(struct xsk_ring_prod *prod,
> size_t nb, __u32 *idx)
>{
> if (unlikely(xsk_prod_nb_free(prod, nb) < nb))
> return 0;
>
> *idx = prod->cached_prod;
> prod->cached_prod += nb;
>
> return nb;
>}
>
>static inline void xsk_ring_prod__submit(struct xsk_ring_prod *prod, size_t
>nb)
>{
> /* Make sure everything has been written to the ring before
>signalling
> * this to the kernel.
> */
> smp_wmb();
>
> *prod->producer += nb;
>}
>
>
>If we reserve N slots, but only submit n slots, we end up with an incorrect
>opinion of the number of available slots later.
>Either xsk_ring_prod__submit should also update cached_prod (but I am not
>sure it was the intent of this api), or we must ensure that both reserve
>and submit are consistent.
I think you are right, current design does have the flaw, I haven't thought of
it before :( So in order to make sure both reserve and submit are consistent,
what
about we check the valid count of mbuf at the beginning, then reserve the valid
count slots?
Thanks,
Xiaolong
>
>Did I miss some subtility ?
>
>
>--
>David Marchand
