[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <54526D3C.2040604@mellanox.com>
Date: Thu, 30 Oct 2014 18:54:20 +0200
From: Or Gerlitz <ogerlitz@...lanox.com>
To: David Laight <David.Laight@...LAB.COM>
CC: "David S. Miller" <davem@...emloft.net>,
"netdev@...r.kernel.org" <netdev@...r.kernel.org>,
Matan Barak <matanb@...lanox.com>,
Amir Vadai <amirv@...lanox.com>,
Saeed Mahameed <saeedm@...lanox.com>,
Shani Michaeli <shanim@...lanox.com>
Subject: Re: [PATCH net-next 7/8] net: Add calaulation of non folded IPV6
pseudo header checksum
On 10/30/2014 6:39 PM, David Laight wrote:
> From: Or Gerlitz [mailto:ogerlitz@...lanox.com]
>> On 10/30/2014 6:25 PM, David Laight wrote:
>>>>> +static inline __wsum csum_ipv6_magic_nofold(const struct in6_addr *saddr,
>>>>> + const struct in6_addr *daddr,
>>>>> + __u32 len, unsigned short proto,
>>>>> + __wsum sum)
>>>>> +{
>>>>> + __wsum res = sum;
>>>>> +
>>>>> + res = csum_add(res, (__force __wsum)saddr->in6_u.u6_addr32[0]);
>>>>> + res = csum_add(res, (__force __wsum)saddr->in6_u.u6_addr32[1]);
>>>>> + res = csum_add(res, (__force __wsum)saddr->in6_u.u6_addr32[2]);
>>>>> + res = csum_add(res, (__force __wsum)saddr->in6_u.u6_addr32[3]);
>>>>> + res = csum_add(res, (__force __wsum)daddr->in6_u.u6_addr32[0]);
>>>>> + res = csum_add(res, (__force __wsum)daddr->in6_u.u6_addr32[1]);
>>>>> + res = csum_add(res, (__force __wsum)daddr->in6_u.u6_addr32[2]);
>>>>> + res = csum_add(res, (__force __wsum)daddr->in6_u.u6_addr32[3]);
>>> That probably generates a very long dependency chain.
>>>
>> Could you clarify this comment a bit?
> csum_add() probably generates a 32bit 'add with carry' instruction.
> So the above generates 8 instructions that have to be executed in series
> (dependencies on the register and the carry flag).
>
> On a 64bit cpu there are other options, eg adding 32bit values into
> several 64bit registers, then adding those together and finally
> collapsing the value to 32 then 16 bits.
>
> Maybe __wsum does end up being 64bit (not looked), but gcc won't
> generate a 'tree' of additions, it will still generate a dependency chain.
>
> Hopefully the software 'checksum a buffer' function is written to
> avoid these problems.
OK, talking to Matan, hecame up with this super-quick (compile tested
only) alternative, is
this what you were advocating for?
Or.
diff --git a/include/net/ip6_checksum.h b/include/net/ip6_checksum.h
index c45d690..cadffdc 100644
--- a/include/net/ip6_checksum.h
+++ b/include/net/ip6_checksum.h
@@ -46,6 +46,19 @@ static inline __wsum csum_ipv6_magic_nofold(const
struct in6_addr *saddr,
__u32 len, unsigned short
proto,
__wsum sum)
{
+ sum = csum_partial(saddr, sizeof(saddr->in6_u.u6_addr32), sum);
+ sum = csum_partial(saddr, sizeof(daddr->in6_u.u6_addr32), sum);
+ sum = csum_add(sum, (__force __wsum)htonl(len));
+ sum = csum_add(sum, (__force __wsum)htons(proto));
+
+ return sum;
+}
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Powered by blists - more mailing lists