aq(4): Distribute RSS over min(rings, 8) with a modulo

aq_if_attach_post() built the RSS indirection table with
	rss_table[i] = i & (rx_rings_count - 1);
which assumes rx_rings_count is a power of two.  iflib does not
guarantee that (queue counts such as 6, 12 or 24 are possible), so the
mask left some queues with no entries and skewed the hash distribution.
When rx_rings_count was 0 the (count - 1) underflow also turned the mask
into a no-op rather than failing safe.

Use a modulo over min(rx_rings_count, HW_ATL_RSS_QUEUES_MAX) instead.
The redirection-table field is 3 bits wide, so the hardware can only
hash across 8 queues regardless of ring count; capping here keeps the
table values in range and complements the bounds fix in aq_hw_rss_set().

Signed-off-by: Nick Price <nick@spun.io>

Author: DominoTree

sys/dev/aq/aq_hw.h

@@ -197,6 +197,8 @@ struct aq_hw {
 
 #define HW_ATL_RSS_INDIRECTION_TABLE_MAX  64U
 #define HW_ATL_RSS_HASHKEY_SIZE           40U
+/* Redirection-table entries are 3 bits wide, so RSS spans <= 8 queues. */
+#define HW_ATL_RSS_QUEUES_MAX             8U
 
 /* PCI core control register */
 #define AQ_HW_PCI_REG_CONTROL_6_ADR 0x1014U

sys/dev/aq/aq_main.c

@@ -464,10 +464,16 @@ aq_if_attach_post(if_ctx_t ctx)
 	}
 
 	aq_add_stats_sysctls(softc);
-	/* RSS */
+	/*
+	 * RSS: modulo over min(rings, 8) queues.  Ring count isn't always a
+	 * power of two, so an AND mask would starve queues; HW caps RSS at 8.
+	 */
 	arc4rand(softc->rss_key, HW_ATL_RSS_HASHKEY_SIZE, 0);
+	uint32_t rss_qs = MIN(softc->rx_rings_count, HW_ATL_RSS_QUEUES_MAX);
+	if (rss_qs == 0)
+		rss_qs = 1;
 	for (int i = ARRAY_SIZE(softc->rss_table); i--;){
-		softc->rss_table[i] = i & (softc->rx_rings_count - 1);
+		softc->rss_table[i] = i % rss_qs;
 	}
 exit:
 	AQ_DBG_EXIT(rc);