ice: xsk: do not use xdp_return_frame() on tx_buf->raw_buf

Previously ice XDP xmit routine was changed in a way that it avoids
xdp_buff->xdp_frame conversion as it is simply not needed for handling
XDP_TX action and what is more it saves us CPU cycles. This routine is
re-used on ZC driver to handle XDP_TX action.

Although for XDP_TX on Rx ZC xdp_buff that comes from xsk_buff_pool is
converted to xdp_frame, xdp_frame itself is not stored inside
ice_tx_buf, we only store raw data pointer. Casting this pointer to
xdp_frame and calling against it xdp_return_frame in
ice_clean_xdp_tx_buf() results in undefined behavior.

To fix this, simply call page_frag_free() on tx_buf->raw_buf.
Later intention is to remove the buff->frame conversion in order to
simplify the codebase and improve XDP_TX performance on ZC.

Fixes: 126cdfe ("ice: xsk: Improve AF_XDP ZC Tx and use batching API")
Reported-and-tested-by: Robin Cowley <[email protected]>
Signed-off-by: Maciej Fijalkowski <[email protected]>
Tested-by: Chandan Kumar Rout <[email protected]> (A Contingent Worker at Intel)
Signed-off-by: Tony Nguyen <[email protected]>
Reviewed-by: Piotr Raczynski <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Jakub Kicinski <[email protected]>

drivers/net/ethernet/intel/ice/ice_xsk.c

@@ -783,7 +783,7 @@ int ice_clean_rx_irq_zc(struct ice_rx_ring *rx_ring, int budget)
 static void
 ice_clean_xdp_tx_buf(struct ice_tx_ring *xdp_ring, struct ice_tx_buf *tx_buf)
 {
-	xdp_return_frame((struct xdp_frame *)tx_buf->raw_buf);
+	page_frag_free(tx_buf->raw_buf);
 	xdp_ring->xdp_tx_active--;
 	dma_unmap_single(xdp_ring->dev, dma_unmap_addr(tx_buf, dma),
 			 dma_unmap_len(tx_buf, len), DMA_TO_DEVICE);