Merge branch 'ionic-better-tx-sg=handling'

Shannon Nelson says:

====================
ionic: better Tx SG handling

The primary patch here is to be sure we're not hitting linearize on a Tx
skb when we don't really need to.  The other two are related details.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/ethernet/pensando/ionic/ionic_dev.h

@@ -7,6 +7,7 @@
 #include <linux/atomic.h>
 #include <linux/mutex.h>
 #include <linux/workqueue.h>
+#include <linux/skbuff.h>
 
 #include "ionic_if.h"
 #include "ionic_regs.h"
@@ -216,7 +217,7 @@ struct ionic_desc_info {
 	};
 	unsigned int bytes;
 	unsigned int nbufs;
-	struct ionic_buf_info bufs[IONIC_MAX_FRAGS];
+	struct ionic_buf_info bufs[MAX_SKB_FRAGS + 1];
 	ionic_desc_cb cb;
 	void *cb_arg;
 };

drivers/net/ethernet/pensando/ionic/ionic_lif.c

@@ -3831,6 +3831,18 @@ static void ionic_lif_queue_identify(struct ionic_lif *lif)
 			qtype, qti->max_sg_elems);
 		dev_dbg(ionic->dev, " qtype[%d].sg_desc_stride = %d\n",
 			qtype, qti->sg_desc_stride);
+
+		if (qti->max_sg_elems >= IONIC_MAX_FRAGS) {
+			qti->max_sg_elems = IONIC_MAX_FRAGS - 1;
+			dev_dbg(ionic->dev, "limiting qtype %d max_sg_elems to IONIC_MAX_FRAGS-1 %d\n",
+				qtype, qti->max_sg_elems);
+		}
+
+		if (qti->max_sg_elems > MAX_SKB_FRAGS) {
+			qti->max_sg_elems = MAX_SKB_FRAGS;
+			dev_dbg(ionic->dev, "limiting qtype %d max_sg_elems to MAX_SKB_FRAGS %d\n",
+				qtype, qti->max_sg_elems);
+		}
 	}
 }
 

drivers/net/ethernet/pensando/ionic/ionic_txrx.c

@@ -1239,25 +1239,84 @@ static int ionic_tx(struct ionic_queue *q, struct sk_buff *skb)
 static int ionic_tx_descs_needed(struct ionic_queue *q, struct sk_buff *skb)
 {
 	struct ionic_tx_stats *stats = q_to_tx_stats(q);
+	bool too_many_frags = false;
+	skb_frag_t *frag;
+	int desc_bufs;
+	int chunk_len;
+	int frag_rem;
+	int tso_rem;
+	int seg_rem;
+	bool encap;
+	int hdrlen;
 	int ndescs;
 	int err;
 
 	/* Each desc is mss long max, so a descriptor for each gso_seg */
-	if (skb_is_gso(skb))
+	if (skb_is_gso(skb)) {
 		ndescs = skb_shinfo(skb)->gso_segs;
-	else
+	} else {
 		ndescs = 1;
+		if (skb_shinfo(skb)->nr_frags > q->max_sg_elems) {
+			too_many_frags = true;
+			goto linearize;
+		}
+	}
 
-	/* If non-TSO, just need 1 desc and nr_frags sg elems */
-	if (skb_shinfo(skb)->nr_frags <= q->max_sg_elems)
+	/* If non-TSO, or no frags to check, we're done */
+	if (!skb_is_gso(skb) || !skb_shinfo(skb)->nr_frags)
 		return ndescs;
 
-	/* Too many frags, so linearize */
-	err = skb_linearize(skb);
-	if (err)
-		return err;
+	/* We need to scan the skb to be sure that none of the MTU sized
+	 * packets in the TSO will require more sgs per descriptor than we
+	 * can support.  We loop through the frags, add up the lengths for
+	 * a packet, and count the number of sgs used per packet.
+	 */
+	tso_rem = skb->len;
+	frag = skb_shinfo(skb)->frags;
+	encap = skb->encapsulation;
+
+	/* start with just hdr in first part of first descriptor */
+	if (encap)
+		hdrlen = skb_inner_tcp_all_headers(skb);
+	else
+		hdrlen = skb_tcp_all_headers(skb);
+	seg_rem = min_t(int, tso_rem, hdrlen + skb_shinfo(skb)->gso_size);
+	frag_rem = hdrlen;
+
+	while (tso_rem > 0) {
+		desc_bufs = 0;
+		while (seg_rem > 0) {
+			desc_bufs++;
+
+			/* We add the +1 because we can take buffers for one
+			 * more than we have SGs: one for the initial desc data
+			 * in addition to the SG segments that might follow.
+			 */
+			if (desc_bufs > q->max_sg_elems + 1) {
+				too_many_frags = true;
+				goto linearize;
+			}
+
+			if (frag_rem == 0) {
+				frag_rem = skb_frag_size(frag);
+				frag++;
+			}
+			chunk_len = min(frag_rem, seg_rem);
+			frag_rem -= chunk_len;
+			tso_rem -= chunk_len;
+			seg_rem -= chunk_len;
+		}
+
+		seg_rem = min_t(int, tso_rem, skb_shinfo(skb)->gso_size);
+	}
 
-	stats->linearize++;
+linearize:
+	if (too_many_frags) {
+		err = skb_linearize(skb);
+		if (err)
+			return err;
+		stats->linearize++;
+	}
 
 	return ndescs;
 }