Ice driver tries to always create XDP rings array to be
num_possible_cpus() sized, regardless of user's queue count setting that
can be changed via ethtool -L for example.
Currently, ice_tx_xsk_pool() calculates the qid by decrementing the
ring->q_index by the count of XDP queues, but ring->q_index is set to 'i
+ vsi->alloc_txq'.
When user did ethtool -L $IFACE combined 1, alloc_txq is 1, but
vsi->num_xdp_txq is still num_possible_cpus(). Then, ice_tx_xsk_pool()
will do OOB access and in the final result ring would not get xsk_pool
pointer assigned. Then, each ice_xsk_wakeup() call will fail with error
and it will not be possible to get into NAPI and do the processing from
driver side.
Fix this by decrementing vsi->alloc_txq instead of vsi->num_xdp_txq from
ring-q_index in ice_tx_xsk_pool() so the calculation is reflected to the
setting of ring->q_index.
CC: Ciara Loftus <ciara.loftus@xxxxxxxxx>
Fixes: 22bf877e528f ("ice: introduce XDP_TX fallback path")
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@xxxxxxxxx>
---
drivers/net/ethernet/intel/ice/ice.h | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/drivers/net/ethernet/intel/ice/ice.h b/drivers/net/ethernet/intel/ice/ice.h
index b0b27bfcd7a2..d4f1874df7d0 100644
--- a/drivers/net/ethernet/intel/ice/ice.h
+++ b/drivers/net/ethernet/intel/ice/ice.h
@@ -710,7 +710,7 @@ static inline struct xsk_buff_pool *ice_tx_xsk_pool(struct ice_tx_ring *ring)
struct ice_vsi *vsi = ring->vsi;
u16 qid;
- qid = ring->q_index - vsi->num_xdp_txq;
+ qid = ring->q_index - vsi->alloc_txq;
if (!ice_is_xdp_ena_vsi(vsi) || !test_bit(qid, vsi->af_xdp_zc_qps))
return NULL;
--
2.27.0
