net/rds: reduce memory footprint during ib_post_recv in IB transport

The RDS IB large fragment size feature requires order 2 memory allocations
and it introduces memory pressure in the allocation system. Thus, this
patch implements large fragment size support in ib_post_recv with N sge. As
of today, RDS has an assumption that each IB received work request has only
two SGEs. This patch removes this assumption and uses various SGE to
support large fragment size.

Orabug: 26770234

Signed-off-by: Wei Lin Guay <[email protected]>
Reviewed-by: Håkon Bugge <[email protected]>
Tested-by: Shih-Yu Huang <[email protected]>
Acked-by: Santosh Shilimkar <[email protected]>
Signed-off-by: Dhaval Giani <[email protected]>

Author: Wei Lin Guay

net/rds/ib.c

@@ -2836,32 +2836,37 @@ int rds_ib_inc_to_skb(struct rds_incoming *inc, struct sk_buff *skb)
 	int i;
 	struct rds_ib_incoming *ibinc;
 	struct rds_page_frag *ibfrag;
+	struct scatterlist *sg;
 
 	/* pull out initial pointers */
 	ibinc  = container_of(inc, struct rds_ib_incoming, ii_inc);
 	ibfrag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item);
 	len    = be32_to_cpu(inc->i_hdr.h_len);
+	sg     = ibfrag->f_sg;
 	slen   = len;
 	i      = 0;
 
 	/* run through the entire ib fragment list and save off the buffers */
 	while (NULL != ibfrag && slen > 0) {
 		/* one to one mapping of frags to sg structures */
 		frag = &skb_shinfo(skb)->frags[i];
-
 		/* save off all the sg pieces to the skb frags we are creating */
-		frag->size        = ibfrag->f_sg.length;
-		frag->page_offset = ibfrag->f_sg.offset;
-		frag->page.p      = sg_page(&ibfrag->f_sg);
+		frag->size        = sg->length;
+		frag->page_offset = sg->offset;
+		frag->page.p      = sg_page(sg);
 
 		/* AA:  do we need to bump up the page reference */
 		/* get_page(frag->page); */
 
 		/* dec the amount of data we are consuming */
 		slen -= frag->size;
 
-		/* bump to the next entry */
-		ibfrag = list_entry(ibfrag->f_item.next, struct rds_page_frag, f_item);
+		sg  = sg_next(sg);
+		if (!sg) {
+			/* bump to the next entry */
+			ibfrag = list_entry(ibfrag->f_item.next, struct rds_page_frag, f_item);
+			sg = ibfrag->f_sg;
+		}
 		i++;
 
 		/* for now we will only have a single chain of fragments in the skb */

net/rds/ib.h

@@ -50,6 +50,8 @@
 
 #define RDS_WC_MAX 32
 
+#define NUM_RDS_RECV_SG	(PAGE_ALIGN(RDS_MAX_FRAG_SIZE) / PAGE_SIZE)
+
 #define	RDS_IB_CLEAN_CACHE	1
 
 #define RDS_IB_DEFAULT_FREG_PORT_NUM	1
@@ -65,7 +67,7 @@ extern struct list_head rds_ib_devices;
 struct rds_page_frag {
 	struct list_head	f_item;
 	struct list_head	f_cache_entry;
-	struct scatterlist	f_sg;
+	struct scatterlist	f_sg[NUM_RDS_RECV_SG];
 };
 
 struct rds_ib_incoming {
@@ -110,7 +112,7 @@ struct rds_ib_recv_work {
 	struct rds_ib_incoming	*r_ibinc;
 	struct rds_page_frag	*r_frag;
 	struct ib_recv_wr	r_wr;
-	struct ib_sge		r_sge[2];
+	struct ib_sge		r_sge[RDS_IB_MAX_SGE];
 	struct rds_ib_connection	*r_ic;
 	int				r_posted;
 };

net/rds/ib_cm.c

@@ -754,7 +754,7 @@ static int rds_ib_setup_qp(struct rds_connection *conn)
 	attr.cap.max_send_wr = ic->i_send_ring.w_nr + 1 + mr_reg;
 	attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1;
 	attr.cap.max_send_sge = rds_ibdev->max_sge;
-	attr.cap.max_recv_sge = RDS_IB_RECV_SGE;
+	attr.cap.max_recv_sge = rds_ibdev->max_sge;
 	attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 	attr.qp_type = IB_QPT_RC;
 	attr.send_cq = ic->i_scq;