Change Details

qls_init_rss() built the RSS indirection table with rss_icb->cq_id[i] = (i & (ha->num_rx_rings - 1)); which assumes num_rx_rings is a power of two. When it is not, the AND mask leaves some completion queues with no entries and skews the hash distribution. num_rx_rings is currently hardcoded to MAX_RX_RINGS (1), so this is latent today, but use a modulo so the table stays correct if the ring count is ever raised to a non-power-of-two value. This mirrors the same fix made in aq(4). Signed-off-by: Nick Price <nick@spun.io>

`qls_init_rss()` built the RSS indirection table with `rss_icb->cq_id[i] = (i & (ha->num_rx_rings - 1));` which assumes `num_rx_rings` is a power of two. When it is not, the AND mask leaves some completion queues with no entries and skews the hash distribution. `MAX_RX_RINGS` is currently hardcoded to `1`, so this is latent today, but use a modulo so the table stays correct if the ring count is ever raised to a non-power-of-two value.

`qls_init_rss()` built the RSS indirection table with `rss_icb->cq_id[i] = (i & (ha->num_rx_rings - 1));` which assumes `num_rx_rings` is a power of two. When it is not, the AND mask leaves some completion queues with no entries and skews the hash rss_icb->cq_id[i] = (i & (ha->num_rx_rings - 1)); which assumes num_rx_rings is a power of two. When it is not, the AND mask leaves some completion queues with no entries and skews the hash distribution. num_rx_rings`MAX_RX_RINGS` is currently hardcoded to MAX_RX_RINGS (1)`1`, so this is latent today, but use a modulo so the table stays correct if the ring count is ever raised to a non-power-of-two value. This mirrors the same fix made in aq(4). Signed-off-by: Nick Price <nick@spun.io>