Message-ID: <182D446E-8871-4811-9275-98FF067B1BA9@amazon.com>
Date:   Sat, 1 Jul 2023 00:26:05 +0000
From:   "Smith, Stewart" <trawets@...zon.com>
To:     Eric Dumazet <edumazet@...gle.com>
CC:     "Mendoza-Jonas, Samuel" <samjonas@...zon.com>,
        "netdev@...r.kernel.org" <netdev@...r.kernel.org>,
        "David S . Miller" <davem@...emloft.net>,
        "David Ahern" <dsahern@...nel.org>,
        Jakub Kicinski <kuba@...nel.org>,
        Paolo Abeni <pabeni@...hat.com>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
        "Herrenschmidt, Benjamin" <benh@...zon.com>,
        "stable@...r.kernel.org" <stable@...r.kernel.org>
Subject: Re: [PATCH net] net/ipv6: Reduce chance of collisions in
 inet6_hashfn()



> On Jun 29, 2023, at 3:05 AM, Eric Dumazet <edumazet@...gle.com> wrote:
> 
> 
> On Thu, Jun 29, 2023 at 3:59 AM Samuel Mendoza-Jonas
> <samjonas@...zon.com> wrote:
>> 
>> From: Stewart Smith <trawets@...zon.com>
>> 
>> For both IPv4 and IPv6 incoming TCP connections are tracked in a hash
>> table with a hash over the source & destination addresses and ports.
>> However, the IPv6 hash is insufficient and can lead to a high rate of
>> collisions.
>> 
>> The IPv6 hash used an XOR to fit everything into the 96 bits for the
>> fast jenkins hash, meaning it is possible for an external entity to
>> ensure the hash collides, thus falling back to a linear search in the
>> bucket, which is slow.
>> 
>> We take the approach of hash half the data; hash the other half; and
>> then hash them together. We do this with 3x jenkins hashes rather than
>> 2x to calculate the hashing value for the connection covering the full
>> length of the addresses and ports.
>> 
> 
> ...
> 
>> While this may look like it adds overhead, the reality of modern CPUs
>> means that this is unmeasurable in real world scenarios.
>> 
>> In simulating with llvm-mca, the increase in cycles for the hashing code
>> was ~5 cycles on Skylake (from a base of ~50), and an extra ~9 on
>> Nehalem (base of ~62).
>> 
>> In commit dd6d2910c5e0 ("netfilter: conntrack: switch to siphash")
>> netfilter switched from a jenkins hash to a siphash, but even the faster
>> hsiphash is a more significant overhead (~20-30%) in some preliminary
>> testing. So, in this patch, we keep to the more conservative approach to
>> ensure we don't add much overhead per SYN.
>> 
>> In testing, this results in a consistently even spread across the
>> connection buckets. In both testing and real-world scenarios, we have
>> not found any measurable performance impact.
>> 
>> Cc: stable@...r.kernel.org
>> Fixes: 08dcdbf6a7b9 ("ipv6: use a stronger hash for tcp")
>> Fixes: b3da2cf37c5c ("[INET]: Use jhash + random secret for ehash.")
>> Signed-off-by: Stewart Smith <trawets@...zon.com>
>> Signed-off-by: Samuel Mendoza-Jonas <samjonas@...zon.com>
>> ---
>> include/net/ipv6.h          | 4 +---
>> net/ipv6/inet6_hashtables.c | 5 ++++-
>> 2 files changed, 5 insertions(+), 4 deletions(-)
>> 
>> diff --git a/include/net/ipv6.h b/include/net/ipv6.h
>> index 7332296eca44..f9bb54869d82 100644
>> --- a/include/net/ipv6.h
>> +++ b/include/net/ipv6.h
>> @@ -752,9 +752,7 @@ static inline u32 ipv6_addr_hash(const struct in6_addr *a)
>> /* more secured version of ipv6_addr_hash() */
>> static inline u32 __ipv6_addr_jhash(const struct in6_addr *a, const u32 initval)
>> {
>> -       u32 v = (__force u32)a->s6_addr32[0] ^ (__force u32)a->s6_addr32[1];
>> -
>> -       return jhash_3words(v,
>> +       return jhash_3words((__force u32)a->s6_addr32[1],
>>                            (__force u32)a->s6_addr32[2],
>>                            (__force u32)a->s6_addr32[3],
>>                            initval);
> 
> Hmmm... see my following comment.
> 
>> diff --git a/net/ipv6/inet6_hashtables.c b/net/ipv6/inet6_hashtables.c
>> index b64b49012655..bb7198081974 100644
>> --- a/net/ipv6/inet6_hashtables.c
>> +++ b/net/ipv6/inet6_hashtables.c
>> @@ -33,7 +33,10 @@ u32 inet6_ehashfn(const struct net *net,
>>        net_get_random_once(&inet6_ehash_secret, sizeof(inet6_ehash_secret));
>>        net_get_random_once(&ipv6_hash_secret, sizeof(ipv6_hash_secret));
>> 
>> -       lhash = (__force u32)laddr->s6_addr32[3];
>> +       lhash = jhash_3words((__force u32)laddr->s6_addr32[3],
>> +                           (((u32)lport) << 16) | (__force u32)fport,
>> +                           (__force u32)faddr->s6_addr32[0],
>> +                           ipv6_hash_secret);
> 
> This seems wrong to me.
> 
> Reusing ipv6_hash_secret and other keys twice is not good, I am sure
> some security researchers
> would love this...

My personal math here is nowhere near what’s needed to work out if it’s a problem or not, it passed the “Am I a complete idiot here?” question of someone much smarter than me in the area, but that’s not sustained scrutiny of course. It’s quite possible there’s something there given enough time to noodle on it.

> Please just change __ipv6_addr_jhash(), so that all users can benefit
> from a more secure version ?
> It also leaves lhash / fhash names relevant here.
> 
> We will probably have to switch to sip (or other stronger hash than
> jhash)  at some point, it is a tradeoff.

Probably to a hsiphash?

When using the same kind of sim with llvm-mca, hsiphash appears to be about the same number of cycles as jhash2 you suggest, so maybe we should just go there and be done with it?

I put my tests and output up at https://github.com/stewartsmith/inet6_hashfn-sim 

I’ll throw some traffic at the hsiphash and see if we can observe a difference.

If nobody is madly complaining about netfilter switching to it as of dd6d2910c5e071a8683827df1a89e527aa5145ab, then it may be fine, but will throw some more benchmarks at it.

Powered by blists - more mailing lists