From: Willy Tarreau <w@1wt.eu>
Date: Tue, 11 Dec 2018 18:50:43 +0000 (+0100)
Subject: MEDIUM: mux-h1: implement true zero-copy of DATA blocks
X-Git-Tag: v1.9-dev11~92
X-Git-Url: http://git.ipfire.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=3815b227f;p=thirdparty%2Fhaproxy.git

MEDIUM: mux-h1: implement true zero-copy of DATA blocks

When the mux's buffer is empty and the htx area contains exactly one
data block of the same size as the requested count, then it's possible
to simply swap the caller's buffer with the mux's output buffer and
adjust offsets and length to match the entire DATA HTX block in the
middle. In this case we perform a true zero-copy operation from
end-to-end. This is the situation that happens all the time with large
files. With this change, the HTX bit rate performance catches up again
with the legacy mode (measured at 97%).
---

diff --git a/src/mux_h1.c b/src/mux_h1.c
index a602b21142..baaf0dfedc 100644
--- a/src/mux_h1.c
+++ b/src/mux_h1.c
@@ -1356,20 +1356,47 @@ static size_t h1_process_output(struct h1c *h1c, struct buffer *buf, size_t coun
 		errflag = H1S_F_REQ_ERROR;
 	}
 
+	blk = htx_get_head_blk(chn_htx);
 
 	tmp = get_trash_chunk();
 
-	/* pre-align the output buffer like the HTX in case it's empty. In this
-	 * case since it's aligned we don't need to use the temporary trash.
+	/* Perform some optimizations to reduce the number of buffer copies.
+	 * First, if the mux's buffer is empty and the htx area contains
+	 * exactly one data block of the same size as the requested count,
+	 * then it's possible to simply swap the caller's buffer with the
+	 * mux's output buffer and adjust offsets and length to match the
+	 * entire DATA HTX block in the middle. In this case we perform a
+	 * true zero-copy operation from end-to-end. This is the situation
+	 * that happens all the time with large files. Second, if this is not
+	 * possible, but the mux's output buffer is empty, we still have an
+	 * opportunity to avoid the copy to the intermediary buffer, by making
+	 * the intermediary buffer's area point to the output buffer's area.
+	 * In this case we want to skip the HTX header to make sure that copies
+	 * remain aligned and that this operation remains possible all the
+	 * time. This goes for headers, data blocks and any data extracted from
+	 * the HTX blocks.
 	 */
 	if (!b_data(&h1c->obuf)) {
 		h1c->obuf.head = sizeof(struct htx);
+
+		if (chn_htx->used == 1 &&
+		    blk && htx_get_blk_type(blk) == HTX_BLK_DATA &&
+		    htx_get_blk_value(chn_htx, blk).len == count) {
+			void *old_area = h1c->obuf.area;
+
+			h1c->obuf.area = buf->area;
+			h1c->obuf.data = count;
+
+			buf->area = old_area;
+			buf->data = buf->head = 0;
+			total += count;
+			goto out;
+		}
 		tmp->area = h1c->obuf.area + h1c->obuf.head;
 	}
 
 	tmp->size = b_room(&h1c->obuf);
 
-	blk = htx_get_head_blk(chn_htx);
 	while (count && !(h1s->flags & errflag) && blk) {
 		struct htx_sl *sl;
 		struct ist n, v;
@@ -1535,9 +1562,10 @@ static size_t h1_process_output(struct h1c *h1c, struct buffer *buf, size_t coun
 	else
 		b_putblk(&h1c->obuf, tmp->area, tmp->data);
 
+	htx_to_buf(chn_htx, buf);
+ out:
 	if (!buf_room_for_htx_data(&h1c->obuf))
 		h1c->flags |= H1C_F_OUT_FULL;
-	htx_to_buf(chn_htx, buf);
   end:
 	return total;
 }