MFC r311877

sfxge(4): avoid unnecessary mbuf data prefetch

Unnecessary prefetch just loads HW prefetcher and displaces other
cache entries (which could be really useful).

If we parse mbuf for TSO early and use firmware-assisted TSO, we do not
expect mbuf data access when we compose firmware-assisted TSO (v1 or v2)
option descriptors.  If packet header needs to be linearized or finally
FATSO cannot be used because of, for example, too big header, we do not
care about a bit more performance degradation because of prefetch
absence (it is better to optimize more common case).

Sponsored by:   Solarflare Communications, Inc.

1 files changed, 14 insertions, 0 deletions

diff --git a/sys/dev/sfxge/sfxge_tx.c b/sys/dev/sfxge/sfxge_tx.c
index cc6fb17..d8ccbf1 100644
--- a/sys/dev/sfxge/sfxge_tx.c
+++ b/sys/dev/sfxge/sfxge_tx.c
@@ -356,8 +356,22 @@ static int sfxge_tx_queue_mbuf(struct sfxge_txq *txq, struct mbuf *mbuf)
 
 	KASSERT(!txq->blocked, ("txq->blocked"));
 
+#if SFXGE_TX_PARSE_EARLY
+	/*
+	 * If software TSO is used, we still need to copy packet header,
+	 * even if we have already parsed it early before enqueue.
+	 */
+	if ((mbuf->m_pkthdr.csum_flags & CSUM_TSO) &&
+	    (txq->tso_fw_assisted == 0))
+		prefetch_read_many(mbuf->m_data);
+#else
+	/*
+	 * Prefetch packet header since we need to parse it and extract
+	 * IP ID, TCP sequence number and flags.
+	 */
 	if (mbuf->m_pkthdr.csum_flags & CSUM_TSO)
 		prefetch_read_many(mbuf->m_data);
+#endif
 
 	if (__predict_false(txq->init_state != SFXGE_TXQ_STARTED)) {
 		rc = EINTR;