rt2x00: add a patch to avoid excessive memmove() on packet contents to improve perfor...

SVN-Revision: 29760

diff --git a/package/mac80211/patches/606-rt2x00_no_realign.patch b/package/mac80211/patches/606-rt2x00_no_realign.patch
new file mode 100644 (file)
index 0000000..e0a920a
--- /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);
+ }
+