On 30.05.2014 14:11, Wei Liu wrote:

On Fri, May 30, 2014 at 10:06:48AM +0200, Stefan Bader wrote:
[...]

quoted

I had been idly wondering about this onwards. And trying to understand the whole
skb handling environment, I tried to come up with some idea as well. It may be
totally stupid and using the wrong assumptions. It seems to work in the sense
that things did not blow up into my face immediately and somehow I did not see
dropped packages due to the number of slots either.
But again, I am not sure I am doing the right thing. The idea was to just try to
get rid of so many compound pages (which I believe are the only ones that can
have an offset big enough to allow some alignment savings)...

-Stefan

Thanks. I think the general idea is OK, but it still involves
unnecessary page allocation. We don't actually need to get rid of
compound page by replacing it with a new page, we just need to make sure
the data inside is aligned.

If you look at xennet_make_frags, it only grants the 4K page which
contains data. I presume a simple memove would be better than alloc_page
+ memcpy. What do you think?

Like:
   memmove(page_address(fpage), page_address(fpage)+offset, size);
   frag->page_offset = 0;

I was hesitating to do that as I could not really tell whether I can make any
assumptions about those memory areas. So my cautious guess was to leave the
original pages alone altogether (maybe whoever owns those has something
important in the starting area). If I did it right, my new page potentially
could be a smaller allocation unit than the original, since I only ask for
something big enough to hold the frag size (ignoring a potential offset over
full 4K areas).

-Stefan

Wei.

quoted

From 8571b106643b32296e58526e2fbe97c330877ac8 Mon Sep 17 00:00:00 2001
From: Stefan Bader <redacted>
Date: Thu, 29 May 2014 12:18:01 +0200
Subject: [PATCH] xen-netfront: Align frags to fit max slots

In cases where the frags in a skb require more than MAX_SKB_FRAGS + 1
(= 18) 4K pages of grant pages, try to reduce the footprint by moving
the data to new pages and have it aligned to the beginning.
Then replace the page in the frag and release the old one. This sure is
more expensive in compute but should happen not too often and sounds
better than to just drop the packet in that case.

Signed-off-by: Stefan Bader <redacted>
---
 drivers/net/xen-netfront.c | 65 +++++++++++++++++++++++++++++++++++++++++++---
 1 file changed, 62 insertions(+), 3 deletions(-)

diff --git a/drivers/net/xen-netfront.c b/drivers/net/xen-netfront.c
index 158b5e6..ad71e5c 100644
--- a/drivers/net/xen-netfront.c
+++ b/drivers/net/xen-netfront.c

@@ -544,6 +544,61 @@ static int xennet_count_skb_frag_slots(struct sk_buff *skb)
 	return pages;
 }

+/*
+ * Align data to new pages in order to save slots required to
+ * transmit this buffer.
+ * @skb - socket buffer
+ * @target - number of pages to save
+ * returns the number of pages the fragments have been reduced of
+ */
+static int xennet_align_frags(struct sk_buff *skb, int target)
+{
+	int i, frags = skb_shinfo(skb)->nr_frags;
+	int reduced = 0;
+
+	for (i = 0; i < frags; i++) {
+		skb_frag_t *frag = skb_shinfo(skb)->frags + i;
+		struct page *fpage = skb_frag_page(frag);
+		struct page *npage;
+		unsigned long size;
+		unsigned long offset;
+		gfp_t gfp;
+		int order;
+
+		if (!PageCompound(fpage))
+			continue;
+
+		size = skb_frag_size(frag);
+		offset = frag->page_offset & ~PAGE_MASK;
+
+		/*
+		 * If the length of data in the last subpage of a compound
+		 * page is smaller than the offset into the first data sub-
+		 * page, we can save a subpage by copying data around.
+		 */
+		if ( ((offset + size) & ~PAGE_MASK) > offset )
+			continue;
+
+		gfp = GFP_ATOMIC | __GFP_COLD;
+		order = PFN_UP(size);
+		if (order)
+			gfp |= __GFP_COMP | __GFP_NOWARN;
+
+		npage = alloc_pages(gfp, order);
+		if (!npage)
+			break;
+		memcpy(page_address(npage), skb_frag_address(frag), size);
+		frag->page.p = npage;
+		frag->page_offset = 0;
+		put_page(fpage);
+
+		if (++reduced >= target)
+			break;
+	}
+
+	return reduced;
+}
+
 static int xennet_start_xmit(struct sk_buff *skb, struct net_device *dev)
 {
 	unsigned short id;

@@ -573,9 +628,13 @@ static int xennet_start_xmit(struct sk_buff *skb, struct

net_device *dev)
 	slots = DIV_ROUND_UP(offset + len, PAGE_SIZE) +
 		xennet_count_skb_frag_slots(skb);
 	if (unlikely(slots > MAX_SKB_FRAGS + 1)) {
-		net_alert_ratelimited(
-			"xennet: skb rides the rocket: %d slots\n", slots);
-		goto drop;
+		slots -= xennet_align_frags(skb, slots - (MAX_SKB_FRAGS + 1));
+		if (slots > MAX_SKB_FRAGS + 1) {
+			net_alert_ratelimited(
+				"xennet: skb rides the rocket: %d slots\n",
+				slots);
+			goto drop;
+		}
 	}

 	spin_lock_irqsave(&np->tx_lock, flags);
-- 
1.9.1