On 4/15/2024 3:40 PM, Alan Elder wrote:
> The previous code allowed the number of Tx queues to be set higher than the
> number of Rx queues. If a packet was sent on a Tx queue with index
>> = number Rx queues there was a segfault due to accessing beyond the end of
>> the dev->data->rx_queues[] array.
>
> #0 rte_spinlock_trylock (sl = invalid address) at /include/rte_spinlock.h L63
> #1 hn_process_events at /drivers/net/netvsc/hn_rxtx.c L 1129
> #2 hn_xmit_pkts at /drivers/net/netvsc/hn_rxtx.c L1553
>
> This commit fixes the issue by creating an Rx queue for every Tx queue
> meaning that an event buffer is allocated to handle receiving Tx completion
> messages.
>
> mbuf pool and Rx ring are not allocated for these additional Rx queues and
> RSS configuration ensures that no packets are received on them.
>
> Fixes: 4e9c73e96e83 ("net/netvsc: add Hyper-V network device")
> Cc: [email protected]
> Cc: [email protected]
>
> Signed-off-by: Alan Elder <[email protected]>
<...>
> @@ -552,10 +595,12 @@ static void hn_rxpkt(struct hn_rx_queue *rxq, struct
> hn_rx_bufinfo *rxb,
> const struct hn_rxinfo *info)
> {
> struct hn_data *hv = rxq->hv;
> - struct rte_mbuf *m;
> + struct rte_mbuf *m = NULL;
> bool use_extbuf = false;
>
> - m = rte_pktmbuf_alloc(rxq->mb_pool);
> + if (likely(rxq->mb_pool != NULL))
> + m = rte_pktmbuf_alloc(rxq->mb_pool);
> +
>
This introduced additional check in Rx path, not sure what is the
performance impact.
I can see Long already acked the v3, I just want to double check.
If Tx queue number > Rx queue number is not a common usecase, perhaps it
can be an option to forbid it instead of getting performance hit.
Or it can be possible to have a dedicated Rx queue, like queue_id 0, for
Tx completion events for Tx queue_id > Rx queue number, etc..
But Long if you prefer to continue with this patch, please ack it and I
can continue with it.
