Re: [PATCH iwl-net] igc: Fix XSK queue NAPI ID mapping

[florian@bezdeka.de]

Hi Joe,

On 2025-03-05 19:09, Joe Damato wrote:
> In commit b65969856d4f ("igc: Link queues to NAPI instances"), the XSK
> queues were incorrectly unmapped from their NAPI instances. After
> discussion on the mailing list and the introduction of a test to codify
> the expected behavior, we can see that the unmapping causes the
> check_xsk test to fail:
> 
> NETIF=enp86s0 ./tools/testing/selftests/drivers/net/queues.py
> 
> [...]
>   # Check|     ksft_eq(q.get('xsk', None), {},
>   # Check failed None != {} xsk attr on queue we configured
>   not ok 4 queues.check_xsk
> 
> After this commit, the test passes:
> 
>   ok 4 queues.check_xsk
> 
> Note that the test itself is only in net-next, so I tested this change
> by applying it to my local net-next tree, booting, and running the 
> test.
> 
> Cc: stable@vger.kernel.org
> Fixes: b65969856d4f ("igc: Link queues to NAPI instances")
> Signed-off-by: Joe Damato <jdamato@fastly.com>
> ---
>  drivers/net/ethernet/intel/igc/igc_xdp.c | 2 --
>  1 file changed, 2 deletions(-)
> 
> diff --git a/drivers/net/ethernet/intel/igc/igc_xdp.c
> b/drivers/net/ethernet/intel/igc/igc_xdp.c
> index 13bbd3346e01..869815f48ac1 100644
> --- a/drivers/net/ethernet/intel/igc/igc_xdp.c
> +++ b/drivers/net/ethernet/intel/igc/igc_xdp.c
> @@ -86,7 +86,6 @@ static int igc_xdp_enable_pool(struct igc_adapter 
> *adapter,
>  		napi_disable(napi);
>  	}
> 
> -	igc_set_queue_napi(adapter, queue_id, NULL);
>  	set_bit(IGC_RING_FLAG_AF_XDP_ZC, &rx_ring->flags);
>  	set_bit(IGC_RING_FLAG_AF_XDP_ZC, &tx_ring->flags);
> 
> @@ -136,7 +135,6 @@ static int igc_xdp_disable_pool(struct igc_adapter
> *adapter, u16 queue_id)
>  	xsk_pool_dma_unmap(pool, IGC_RX_DMA_ATTR);
>  	clear_bit(IGC_RING_FLAG_AF_XDP_ZC, &rx_ring->flags);
>  	clear_bit(IGC_RING_FLAG_AF_XDP_ZC, &tx_ring->flags);
> -	igc_set_queue_napi(adapter, queue_id, napi);
> 
>  	if (needs_reset) {
>  		napi_enable(napi);

That doesn't look correct to me. You removed both invocations of
igc_set_queue_napi() from igc_xdp.c. Where is the napi mapping now
done (in case XDP is enabled)?

To me it seems flipped. igc_xdp_enable_pool() should do the mapping
(previously did the unmapping) and igc_xdp_disable_pool() should do
the unmapping (previously did the mapping). No?

Btw: I got this patch via stable. It doesn't make sense to send it
to stable where this patch does not apply.

> 
> base-commit: 3c9231ea6497dfc50ac0ef69fff484da27d0df66