net/intel: ensure correct Rx path is selected

The common rx path selection logic iterates through an array of
candidate paths and selects the best fit for the requested features.
Currently, in the event that two potential candidates are identified,
the one with the fewer offloads (and thus less complex path) is
selected. However this is not correct, because if the path with more
offloads has a greater SIMD width, that should be chosen. This commit
reworks the logic so that the number of offloads is only taken into
consideration when choosing between two paths with the same SIMD width.

Since the paths arrays are ordered from lowest SIMD width to highest,
and vector paths tend to have fewer offloads enabled than scalar paths,
"new" candidate paths with greater SIMDs widths tended to have fewer or
equal offloads than the "current" candidate paths and thus were
correctly accepted as the best candidate. For this reason the incorrect
logic did not cause any incorrect path selections in practise.

Fixes: 9d99641 ("net/intel: introduce infrastructure for Rx path selection")

Signed-off-by: Ciara Loftus <[email protected]>
Acked-by: Bruce Richardson <[email protected]>

Author: cloftus

drivers/net/intel/common/rx.h

@@ -300,8 +300,11 @@ ci_rx_path_select(struct ci_rx_path_features req_features,
 			/* Do not select paths with lower SIMD width than the current path. */
 			if (path_features->simd_width < current_features->simd_width)
 				continue;
-			/* Do not select paths with more offloads enabled than the current path. */
-			if (rte_popcount32(path_features->rx_offloads) >
+			/* Do not select paths with more offloads enabled than the current path if
+			 * the SIMD widths are the same.
+			 */
+			if (path_features->simd_width == current_features->simd_width &&
+					rte_popcount32(path_features->rx_offloads) >
 					rte_popcount32(current_features->rx_offloads))
 				continue;
 			/* Do not select paths without bulk alloc support if requested and the