svn commit: r318973 - head/contrib/ofed/libcxgb4/src

[Navdeep Parhar]

Author: np
Date: Sat May 27 02:05:21 2017
New Revision: 318973
URL: https://svnweb.freebsd.org/changeset/base/318973

Log:
  libcxgb4: Use memcpy instead of copying WRs 8B at a time in the userspace
  RDMA library for cxgbe(4).
  
  MFC after:	3 days
  Sponsored by:	Chelsio Communications

Modified:
  head/contrib/ofed/libcxgb4/src/qp.c

Modified: head/contrib/ofed/libcxgb4/src/qp.c
==============================================================================
--- head/contrib/ofed/libcxgb4/src/qp.c	Sat May 27 01:35:59 2017	(r318972)
+++ head/contrib/ofed/libcxgb4/src/qp.c	Sat May 27 02:05:21 2017	(r318973)
@@ -47,39 +47,49 @@ struct c4iw_stats c4iw_stats;
 
 static void copy_wr_to_sq(struct t4_wq *wq, union t4_wr *wqe, u8 len16)
 {
-	u64 *src, *dst;
-
-	src = (u64 *)wqe;
-	dst = (u64 *)((u8 *)wq->sq.queue + wq->sq.wq_pidx * T4_EQ_ENTRY_SIZE);
+	void *src, *dst;
+	uintptr_t end;
+	int total, len;
+
+	src = &wqe->flits[0];
+	dst = &wq->sq.queue->flits[wq->sq.wq_pidx *
+	    (T4_EQ_ENTRY_SIZE / sizeof(__be64))];
 	if (t4_sq_onchip(wq)) {
 		len16 = align(len16, 4);
 		wc_wmb();
 	}
-	while (len16) {
-		*dst++ = *src++;
-		if (dst == (u64 *)&wq->sq.queue[wq->sq.size])
-			dst = (u64 *)wq->sq.queue;
-		*dst++ = *src++;
-		if (dst == (u64 *)&wq->sq.queue[wq->sq.size])
-			dst = (u64 *)wq->sq.queue;
-		len16--;
+
+	total = len16 * 16;
+	end = (uintptr_t)&wq->sq.queue[wq->sq.size];
+	if (__predict_true((uintptr_t)dst + total <= end)) {
+		/* Won't wrap around. */
+		memcpy(dst, src, total);
+	} else {
+		len = end - (uintptr_t)dst;
+		memcpy(dst, src, len);
+		memcpy(wq->sq.queue, src + len, total - len);
 	}
 }
 
 static void copy_wr_to_rq(struct t4_wq *wq, union t4_recv_wr *wqe, u8 len16)
 {
-	u64 *src, *dst;
-
-	src = (u64 *)wqe;
-	dst = (u64 *)((u8 *)wq->rq.queue + wq->rq.wq_pidx * T4_EQ_ENTRY_SIZE);
-	while (len16) {
-		*dst++ = *src++;
-		if (dst >= (u64 *)&wq->rq.queue[wq->rq.size])
-			dst = (u64 *)wq->rq.queue;
-		*dst++ = *src++;
-		if (dst >= (u64 *)&wq->rq.queue[wq->rq.size])
-			dst = (u64 *)wq->rq.queue;
-		len16--;
+	void *src, *dst;
+	uintptr_t end;
+	int total, len;
+
+	src = &wqe->flits[0];
+	dst = &wq->rq.queue->flits[wq->rq.wq_pidx *
+	    (T4_EQ_ENTRY_SIZE / sizeof(__be64))];
+
+	total = len16 * 16;
+	end = (uintptr_t)&wq->rq.queue[wq->rq.size];
+	if (__predict_true((uintptr_t)dst + total <= end)) {
+		/* Won't wrap around. */
+		memcpy(dst, src, total);
+	} else {
+		len = end - (uintptr_t)dst;
+		memcpy(dst, src, len);
+		memcpy(wq->rq.queue, src + len, total - len);
 	}
 }