net/rds: reduce memory footprint during rds_sendmsg with 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 removes this dependency and uses multiple sge in the IB work requests
to support large fragment size.

Orabug: 27339270

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]>
(cherry picked from commit 0a5d76db3b352d06cf677c668d02e0326b20d8b1)
(cherry picked from commit c5cc5d4c5fb98f5d4fed920cd615b10f9b91ff00)
Signed-off-by: Somasundaram Krishnasamy <[email protected]>

Author: Wei Lin Guay

net/rds/ib_send.c

@@ -237,7 +237,9 @@ static struct rds_message *rds_ib_send_unmap_op(struct rds_ib_connection *ic,
 void rds_ib_send_init_ring(struct rds_ib_connection *ic)
 {
 	struct rds_ib_send_work *send;
+	u32 num_send_sge = ic->i_frag_pages;
 	u32 i;
+	u32 j;
 
 	for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
 		struct ib_sge *sge;
@@ -253,7 +255,8 @@ void rds_ib_send_init_ring(struct rds_ib_connection *ic)
 		sge->length = sizeof(struct rds_header);
 		sge->lkey = ic->i_mr->lkey;
 
-		send->s_sge[1].lkey = ic->i_mr->lkey;
+		for (j = 1; j <= num_send_sge; j++)
+			send->s_sge[j].lkey = ic->i_mr->lkey;
 	}
 }
 
@@ -561,6 +564,7 @@ int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
 	struct rds_ib_send_work *prev;
 	struct ib_send_wr *failed_wr;
 	struct scatterlist *scat;
+	int remaining_sge = 0;
 	u32 pos;
 	u32 i;
 	u32 work_alloc;
@@ -573,7 +577,6 @@ int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
 	int flow_controlled = 0;
 	int nr_sig = 0;
 
-	BUG_ON(off % ic->i_frag_sz);
 	BUG_ON(hdr_off != 0 && hdr_off != sizeof(struct rds_header));
 
 	/* Do not send cong updates to IB loopback */
@@ -715,6 +718,7 @@ int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
 	prev = NULL;
 	scat = &ic->i_data_op->op_sg[rm->data.op_dmasg];
 	i = 0;
+	remaining_sge = rm->data.op_count - sg;
 	do {
 		unsigned int len = 0;
 
@@ -735,20 +739,27 @@ int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
 		/* Set up the data, if present */
 		if (i < work_alloc
 		    && scat != &rm->data.op_sg[rm->data.op_count]) {
-			len = min((unsigned int)ic->i_frag_sz,
-				  ib_sg_dma_len(dev, scat) - rm->data.op_dmaoff);
-			send->s_wr.num_sge = 2;
-
-			send->s_sge[1].addr = ib_sg_dma_address(dev, scat);
-			send->s_sge[1].addr += rm->data.op_dmaoff;
-			send->s_sge[1].length = len;
-
-			bytes_sent += len;
-			rm->data.op_dmaoff += len;
-			if (rm->data.op_dmaoff == ib_sg_dma_len(dev, scat)) {
-				scat++;
-				rm->data.op_dmasg++;
-				rm->data.op_dmaoff = 0;
+			unsigned int num_sge = min_t(unsigned long, remaining_sge,
+						     ic->i_frag_pages);
+			unsigned int j = 1;
+
+			send->s_wr.num_sge += num_sge;
+			while (j <= num_sge) {
+				len = min((unsigned int)PAGE_SIZE,
+					  ib_sg_dma_len(dev, scat) - rm->data.op_dmaoff);
+				send->s_sge[j].addr = ib_sg_dma_address(dev, scat);
+				send->s_sge[j].addr += rm->data.op_dmaoff;
+				send->s_sge[j].length = len;
+
+				bytes_sent += len;
+				rm->data.op_dmaoff += len;
+				if (rm->data.op_dmaoff == ib_sg_dma_len(dev, scat)) {
+					scat++;
+					rm->data.op_dmasg++;
+					rm->data.op_dmaoff = 0;
+				}
+				j++;
+				remaining_sge--;
 			}
 		}
 

net/rds/send.c

@@ -1277,7 +1277,7 @@ int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len)
 	if (payload_len) {
 		rm->data.op_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE));
 		ret = rds_message_copy_from_user(rm, &msg->msg_iter, GFP_KERNEL,
-						 large_page);
+						 false);
 		if (ret)
 			goto out;
 	}