From 2683b5b15de6b1fce7e1e6d84694c7cc441f0e49 Mon Sep 17 00:00:00 2001
From: Felix Fietkau <nbd@openwrt.org>
Date: Mon, 16 Jan 2012 08:14:27 +0000
Subject: [PATCH] rt2x00: add a patch to avoid excessive memmove() on packet
 contents to improve performance (patch by Helmut Schaa)

SVN-Revision: 29760
---
 .../patches/606-rt2x00_no_realign.patch       | 67 +++++++++++++++++++
 1 file changed, 67 insertions(+)
 create mode 100644 package/mac80211/patches/606-rt2x00_no_realign.patch

diff --git a/package/mac80211/patches/606-rt2x00_no_realign.patch b/package/mac80211/patches/606-rt2x00_no_realign.patch
new file mode 100644
index 0000000000..e0a920a582
--- /dev/null
+++ b/package/mac80211/patches/606-rt2x00_no_realign.patch
@@ -0,0 +1,67 @@
+[RFC] rt2x00: For drivers that only need L2 padding don't realign frames
+
+Signed-off-by: Helmut Schaa <helmut.schaa@...>
+---
+
+Ivo, Gertjan, do you remeber by any chance why this alignment stuff was added
+in the first place? Was it because of DMA restrictions?
+
+While doing some profiling on the rt3052 SoC I noticed that 30-40% time was
+spent in memmove calls. And the culprit is the memmove aligning the payload
+to a 4byte boundary since that has to move a whole bunch of data.
+
+Interesstingly the legacy drivers insert an l2pad between the header and the
+payload but doesn't realign the payload itself to a 4-byte boundary. Hence,
+I came up with this patch and indeed CPU usage improves impressively.
+
+Only tested on rt2800pci!
+
+Thanks,
+Helmut
+
+ drivers/net/wireless/rt2x00/rt2x00queue.c |   30 +++-------------------------
+ 1 files changed, 4 insertions(+), 26 deletions(-)
+
+--- a/drivers/net/wireless/rt2x00/rt2x00queue.c
++++ b/drivers/net/wireless/rt2x00/rt2x00queue.c
+@@ -151,36 +151,14 @@ void rt2x00queue_align_frame(struct sk_b
+ void rt2x00queue_insert_l2pad(struct sk_buff *skb, unsigned int header_length)
+ {
+ 	unsigned int payload_length = skb->len - header_length;
+-	unsigned int header_align = ALIGN_SIZE(skb, 0);
+-	unsigned int payload_align = ALIGN_SIZE(skb, header_length);
+ 	unsigned int l2pad = payload_length ? L2PAD_SIZE(header_length) : 0;
+ 
+-	/*
+-	 * Adjust the header alignment if the payload needs to be moved more
+-	 * than the header.
+-	 */
+-	if (payload_align > header_align)
+-		header_align += 4;
+-
+-	/* There is nothing to do if no alignment is needed */
+-	if (!header_align)
++	if (!l2pad)
+ 		return;
+ 
+-	/* Reserve the amount of space needed in front of the frame */
+-	skb_push(skb, header_align);
+-
+-	/*
+-	 * Move the header.
+-	 */
+-	memmove(skb->data, skb->data + header_align, header_length);
+-
+-	/* Move the payload, if present and if required */
+-	if (payload_length && payload_align)
+-		memmove(skb->data + header_length + l2pad,
+-			skb->data + header_length + l2pad + payload_align,
+-			payload_length);
+-
+-	/* Trim the skb to the correct size */
++	/* insert l2pad -> Move header */
++	skb_push(skb, l2pad);
++	memmove(skb->data, skb->data + l2pad, header_length);
+ 	skb_trim(skb, header_length + l2pad + payload_length);
+ }
+ 
-- 
2.30.2