Skip to contentSkip to conversation

net: zerocopy: combine pages in zerocopy_sg_from_iter()

PATCH Applied Eric Dumazet 2 participants 2 messages Aug 20, 2020

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 net: zerocopy: combine pages in zerocopy_sg_from_iter()

Currently, tcp sendmsg(MSG_ZEROCOPY) is building skbs with order-0 fragments. Compared to standard sendmsg(), these skbs usually contain up to 16 fragments on arches with 4KB page sizes, instead of two.

This adds considerable costs on various ndo_start_xmit() handlers, especially when IOMMU is in the picture.

As high performance applications are often using huge pages, we can try to combine adjacent pages belonging to same compound page.

Tested on AMD Rome platform, with IOMMU, nominal single TCP flow speed is roughly doubled (~55Gbit -> ~100Gbit), when user application is using hugepages.

For reference, nominal single TCP flow speed on this platform without MSG_ZEROCOPY is ~65Gbit.

Cc Willem de Bruijn <[email protected]>
---
 net/core/datagram.c | 33 +++++++++++++++++++++++++++++----
 1 file changed, 29 insertions(+), 4 deletions(-)
net/core/datagram.c +29 -4
index 639745d4f3b94a248da9a685f45158410a85bec7..9fcaa544f11a92f1b833d03e9db0863c32905673 100644
@@ -623,10 +623,11 @@ int __zerocopy_sg_from_iter(struct sock *sk, struct sk_buff *skb,
 
 	while (length && iov_iter_count(from)) {
 		struct page *pages[MAX_SKB_FRAGS];
+		struct page *last_head = NULL;
 		size_t start;
 		ssize_t copied;
 		unsigned long truesize;
-		int n = 0;
+		int refs, n = 0;
 
 		if (frag == MAX_SKB_FRAGS)
 			return -EMSGSIZE;
@@ -649,13 +650,37 @@ int __zerocopy_sg_from_iter(struct sock *sk, struct sk_buff *skb,
 		} else {
 			refcount_add(truesize, &skb->sk->sk_wmem_alloc);
 		}
-		while (copied) {
+		for (refs = 0; copied != 0; start = 0) {
 			int size = min_t(int, copied, PAGE_SIZE - start);
-			skb_fill_page_desc(skb, frag++, pages[n], start, size);
-			start = 0;
+			struct page *head = compound_head(pages[n]);
+
+			start += (pages[n] - head) << PAGE_SHIFT;
 			copied -= size;
 			n++;
+			if (frag) {
+				skb_frag_t *last = &skb_shinfo(skb)->frags[frag - 1];
+
+				if (head == skb_frag_page(last) &&
+				    start == skb_frag_off(last) + skb_frag_size(last)) {
+					skb_frag_size_add(last, size);
+					/* We combined this page, we need to release
+					 * a reference. Since compound pages refcount
+					 * is shared among many pages, batch the refcount
+					 * adjustments to limit false sharing.
+					 */
+					last_head = head;
+					refs++;
+					continue;
+				}
+			}
+			if (refs) {
+				page_ref_sub(last_head, refs);
+				refs = 0;
+			}
+			skb_fill_page_desc(skb, frag++, head, start, size);
 		}
+		if (refs)
+			page_ref_sub(last_head, refs);
 	}
 	return 0;
 }
Signature · 1 line
2.28.0.220.ged08abb693-goog
maintainer Applied
#1 lore raw

From: Eric Dumazet <[email protected]>
Date: Thu, 20 Aug 2020 08:43:59 -0700

Quoted 11 lines

Currently, tcp sendmsg(MSG_ZEROCOPY) is building skbs with order-0 fragments. Compared to standard sendmsg(), these skbs usually contain up to 16 fragments on arches with 4KB page sizes, instead of two.

This adds considerable costs on various ndo_start_xmit() handlers, especially when IOMMU is in the picture.

As high performance applications are often using huge pages, we can try to combine adjacent pages belonging to same compound page.

Quoted 8 lines

Tested on AMD Rome platform, with IOMMU, nominal single TCP flow speed is roughly doubled (~55Gbit -> ~100Gbit), when user application is using hugepages.

For reference, nominal single TCP flow speed on this platform without MSG_ZEROCOPY is ~65Gbit.

Applied, the refcounitng in these kinds of patchs is always fun to audit :-)