Re: IP header alignment
Thus, I believe, for IP stack it is wise to have IP header aligned by drivers. I warn once, so it do not flood system log with messages. It is wise only on systems that need it, and most can (albeit expensively) handle the unaligned access, so it is absolutely not a requirement in those cases. Look, the point still stands, this thing is going to crap into people's kernel logs and that is not acceptable. Furthermore, there are drivers that do alignment conditionally based upon whether the platform can handle it transparently or not. They do this because fixing the alignment might result in expensive copies. So these drivers will always generate warnings, even though it is entirely intentional. This is just a bad idea, sorry. And anyone who is truly interested in this kind of work can add this as a temporary debugging patch to their tree during development. There are several aspects. 1: I agree that there is platforms that can handle unaligned access. But, there are others that can't. And, on platforms that can't unaligned IP header will lead to incorrect data processing. I want to deal with this issue. So, in this case, I can check alignment and, if platform can't handle unaligned loads, issue error message once and drop unaligned frames always. 2: Even for platforms that can do unaligned loads, overhead is not zero. If it is easy to modify driver to align data without overhead for frame copy, it is worth to do so. In this case, warn once don't seems to be that bad. Drivers that intentionally deliver unaligned data will generate one line in syslog, is it critical? This warning may be put under CONFIG_DEBUG_KERNEL. It will help to identify problem in driver for those who have no platform w/o unaligned loads. Comments? - To unsubscribe from this list: send the line unsubscribe linux-net in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: IP header alignment
On Mon, 14 Jan 2008 10:33:52 +0200
Vladimir Kondratiev [EMAIL PROTECTED] wrote:
Following patch made against 2.6.24-rc7
It warns if IP stack gets frame with IP header not aligned on dword.
Idea is this will help to identify mis-behaving drivers that do not
enforce this alignment.
-
commit d9ff4a47ca7bd902947eb22ed3b01682dce2752f
Author: Vladimir Kondratiev [EMAIL PROTECTED]
Date: Mon Jan 14 10:08:09 2008 +0200
[NET] warn on non-aligned IP header
IP stack silently assumes that IP header is aligned on dword.
While IP stack may be modified to overcome this requirement, it will
impose performance penalty. Better off to force drivers to properly align
data in received frame.
This patch warns once, hoping that frame will be handled properly any way.
Drop frame in this case would be unappropriate - this will break some
wireless drivers.
diff --git a/net/ipv4/ip_input.c b/net/ipv4/ip_input.c
index 168c871..88c8369 100644
--- a/net/ipv4/ip_input.c
+++ b/net/ipv4/ip_input.c
@@ -400,7 +400,23 @@ int ip_rcv(struct sk_buff *skb, struct net_device
*dev, struct packet_type *pt,
goto inhdr_error;
iph = ip_hdr(skb);
-
+/**
+ * IP stack silently assumes that IP header aligned on dword.
+ * If it is not:
+ * 1. extra time for non-aligned access will be paid for every
+ * architecture;
+ * 2. various things will mis-behave on hardware that do not
+ * support non-aligned access.
+ * Driver code should be fixed.
+ */
+if (unlikely((u32)iph % 4)) {
+static int __warned;
+if (unlikely(!__warned)) {
+__warned = 1;
+dev_warn(orig_dev-dev,Non aligned IP header from [%s]\n,
+ orig_dev-name);
+}
+}
/*
*RFC1122: 3.1.2.2 MUST silently discard any IP frame that
fails the checksum.
*
No. There is hardware like Yukon-EC used in mac mini where the cost
of the CPU unaligned access is trivial and the cost of copying every
frame to shut up the stupid warning would be high.
--
Stephen Hemminger [EMAIL PROTECTED]
-
To unsubscribe from this list: send the line unsubscribe linux-net in
the body of a message to [EMAIL PROTECTED]
More majordomo info at http://vger.kernel.org/majordomo-info.html
