Skip to contentSkip to conversation

tcp: Reduce chance of collisions in inet6_hashfn().

PATCH Applied Kuniyuki Iwashima 5 participants 7 messages Jul 21, 2023 – Jul 25, 2023

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 tcp: Reduce chance of collisions in inet6_hashfn().

From: Stewart Smith <[email protected]>

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 the full length of IPv6 address in __ipv6_addr_jhash() so that all users can benefit from a more secure version.

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 ~16 cycles on Skylake (from a base of ~155), and an extra ~9 on Nehalem (base of ~173).

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.

Fixes 08dcdbf6a7b9 ("ipv6: use a stronger hash for tcp")

---
v2:
* Hash all IPv6 bytes once in __ipv6_addr_jhash() instead of reusing some bytes twice.

v1: https://lore.kernel.org/netdev/[email protected]/

---
 include/net/ipv6.h | 8 ++------
 1 file changed, 2 insertions(+), 6 deletions(-)
include/net/ipv6.h +2 -6
index 7332296eca44..2acc4c808d45 100644
@@ -752,12 +752,8 @@ 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,
-			    (__force u32)a->s6_addr32[2],
-			    (__force u32)a->s6_addr32[3],
-			    initval);
+	return jhash2((__force const u32 *)a->s6_addr32,
+		      ARRAY_SIZE(a->s6_addr32), initval);
 }
 
 static inline bool ipv6_addr_loopback(const struct in6_addr *a)
Signature · 1 line
2.30.2
#1 lore raw

Resending because some recipients require plaintext email. Sorry. Original message:

This is certainly better than the original code.

Two remarks:

1. In general, using SipHash is more secure, even if only for its longer key (128 bits, cf. jhash's 32 bits), which renders simple enumeration attacks infeasible. I understand that in a different context, switching from jhash to siphash incurred some overhead, but maybe here it won't.

2. Taking a more holistic approach to __ipv6_addr_jhash(), I surveyed where and how it's used. In most cases, it is used for hashing together the IPv6 local address, foreign address and optionally some more data (e.g. layer 4 protocol number, layer 4 ports). Security-wise, it makes more sense to hash all data together once, and not piecewise as it's done today (i.e. today it's jhash(....,jhash(faddr),...), which cases the faddr into 32 bits, whereas the recommended way is to hash all items in their entirety, i.e. jhash(...,faddr,...)). This requires scrutinizing all 6 invocations of __ipv6_addr_jhash() one by one and modifying the calling code accordingly.

Thanks,
-Amit

Quoted 71 lines · Kuniyuki Iwashima

From: Stewart Smith <[email protected]>

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 the full length of IPv6 address in __ipv6_addr_jhash() so that all users can benefit from a more secure version.

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 ~16 cycles on Skylake (from a base of ~155), and an extra ~9 on Nehalem (base of ~173).

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.

Fixes 08dcdbf6a7b9 ("ipv6: use a stronger hash for tcp")

---
v2:
* Hash all IPv6 bytes once in __ipv6_addr_jhash() instead of reusing some bytes twice.

v1: https://lore.kernel.org/netdev/[email protected]/

---
 include/net/ipv6.h | 8 ++------
 1 file changed, 2 insertions(+), 6 deletions(-)
include/net/ipv6.h +2 -7
index 7332296eca44..2acc4c808d45 100644
@@ -752,12 +752,8 @@ 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,
-                           (__force u32)a->s6_addr32[2],
-                           (__force u32)a->s6_addr32[3],
-                           initval);
+       return jhash2((__force const u32 *)a->s6_addr32,
+                     ARRAY_SIZE(a->s6_addr32), initval);
 }
 
 static inline bool ipv6_addr_loopback(const struct in6_addr *a)
--

2.30.2

#2 lore raw

From: Amit Klein <[email protected]>
Date: Sat, 22 Jul 2023 03:07:49 +0300

Quoted 4 lines

Resending because some recipients require plaintext email. Sorry. Original message:

This is certainly better than the original code.

Thanks for reviewing!

Quoted 8 lines

Two remarks:

1. In general, using SipHash is more secure, even if only for its longer key (128 bits, cf. jhash's 32 bits), which renders simple enumeration attacks infeasible. I understand that in a different context, switching from jhash to siphash incurred some overhead, but maybe here it won't.

I see. Stewart tested hsiphash and observed more overhead as noted in the changelog, but let me give another shot to SipHash and HSiphash.

I'll report back here next week.

Quoted 10 lines

2. Taking a more holistic approach to __ipv6_addr_jhash(), I surveyed where and how it's used. In most cases, it is used for hashing together the IPv6 local address, foreign address and optionally some more data (e.g. layer 4 protocol number, layer 4 ports). Security-wise, it makes more sense to hash all data together once, and not piecewise as it's done today (i.e. today it's jhash(....,jhash(faddr),...), which cases the faddr into 32 bits, whereas the recommended way is to hash all items in their entirety, i.e. jhash(...,faddr,...)).

Agree.

Quoted 3 lines

This requires scrutinizing all 6
invocations of __ipv6_addr_jhash() one by one and modifying the calling code accordingly.

At a glance, only rds_conn_bucket() seems a little bit tricky as it uses v4 hash function later.

But I'll take a deeper look.

Thanks!

Quoted 75 lines

Thanks,
-Amit

Quoted 70 lines · Kuniyuki Iwashima

From: Stewart Smith <[email protected]>

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 the full length of IPv6 address in __ipv6_addr_jhash() so that all users can benefit from a more secure version.

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 ~16 cycles on Skylake (from a base of ~155), and an extra ~9 on Nehalem (base of ~173).

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.

Fixes 08dcdbf6a7b9 ("ipv6: use a stronger hash for tcp")

---
v2:
* Hash all IPv6 bytes once in __ipv6_addr_jhash() instead of reusing some bytes twice.

v1: https://lore.kernel.org/netdev/[email protected]/

---
 include/net/ipv6.h | 8 ++------
 1 file changed, 2 insertions(+), 6 deletions(-)
include/net/ipv6.h +2 -7
index 7332296eca44..2acc4c808d45 100644
@@ -752,12 +752,8 @@ 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,
-                           (__force u32)a->s6_addr32[2],
-                           (__force u32)a->s6_addr32[3],
-                           initval);
+       return jhash2((__force const u32 *)a->s6_addr32,
+                     ARRAY_SIZE(a->s6_addr32), initval);
 }
 
 static inline bool ipv6_addr_loopback(const struct in6_addr *a)
--

2.30.2

#3 lore raw
Quoted 20 lines · Kuniyuki

From: Amit Klein <[email protected]>
Date: Sat, 22 Jul 2023 03:07:49 +0300

Quoted 4 lines

Resending because some recipients require plaintext email. Sorry. Original message:

This is certainly better than the original code.

Thanks for reviewing!

Quoted 8 lines

Two remarks:

1. In general, using SipHash is more secure, even if only for its longer key (128 bits, cf. jhash's 32 bits), which renders simple enumeration attacks infeasible. I understand that in a different context, switching from jhash to siphash incurred some overhead, but maybe here it won't.

Quoted 3 lines

I see. Stewart tested hsiphash and observed more overhead as noted in the changelog, but let me give another shot to SipHash and HSiphash.

When originally looking at what to do for the collisions, it did seem like siphash would be the better hash, but I did have some concern around what the real-world performance impact could be, and wanted to have something that would alleviate the issue at hand that nobody could even *possibly* remotely contemplate complaining was a problem to apply the patch to their systems because of “performance”.. which was why keeping jhash but with modifications was where we started.

Two things about siphash/hsiphash that I had open questions about:
- is the extra overhead of the hash calculation going to be noticeable at all in any regular workload
- all the jhash stuff gets inlined nicely and the compiler does wonderful things, but it’s possible that siphash will end up always with a jump, which would add more to the extra overhead (and that llvm-mca doesn’t really model well, so it was harder to prove in sim). Again, not quite sure the real world impact to real world workloads.

Quoted 15 lines

I'll report back here next week.

Quoted 10 lines

2. Taking a more holistic approach to __ipv6_addr_jhash(), I surveyed where and how it's used. In most cases, it is used for hashing together the IPv6 local address, foreign address and optionally some more data (e.g. layer 4 protocol number, layer 4 ports). Security-wise, it makes more sense to hash all data together once, and not piecewise as it's done today (i.e. today it's jhash(....,jhash(faddr),...), which cases the faddr into 32 bits, whereas the recommended way is to hash all items in their entirety, i.e. jhash(...,faddr,...)).

Quoted 1 line

Agree.

When looking at this originally, I was trying to work out what was intentional for some reason I wasn’t smart enough to understand and what had just grown over the past (eep.. nearly 3 decades I think) and could do with a bit of a rethink.

Quoted 8 lines
Quoted 3 lines

This requires scrutinizing all 6
invocations of __ipv6_addr_jhash() one by one and modifying the calling code accordingly.

At a glance, only rds_conn_bucket() seems a little bit tricky as it uses v4 hash function later.

But I'll take a deeper look.

#4 lore raw

From: "Smith, Stewart" <[email protected]>
Date: Sat, 22 Jul 2023 03:17:53 +0000

Quoted 33 lines · Kuniyuki
Quoted 20 lines

From: Amit Klein <[email protected]>
Date: Sat, 22 Jul 2023 03:07:49 +0300

Quoted 4 lines

Resending because some recipients require plaintext email. Sorry. Original message:

This is certainly better than the original code.

Thanks for reviewing!

Quoted 8 lines

Two remarks:

1. In general, using SipHash is more secure, even if only for its longer key (128 bits, cf. jhash's 32 bits), which renders simple enumeration attacks infeasible. I understand that in a different context, switching from jhash to siphash incurred some overhead, but maybe here it won't.

Quoted 3 lines

I see. Stewart tested hsiphash and observed more overhead as noted in the changelog, but let me give another shot to SipHash and HSiphash.

When originally looking at what to do for the collisions, it did seem like siphash would be the better hash, but I did have some concern around what the real-world performance impact could be, and wanted to have something that would alleviate the issue at hand that nobody could even *possibly* remotely contemplate complaining was a problem to apply the patch to their systems because of “performance”.. which was why keeping jhash but with modifications was where we started.

Quoted 8 lines

Two things about siphash/hsiphash that I had open questions about:
- is the extra overhead of the hash calculation going to be noticeable at all in any regular workload
- all the jhash stuff gets inlined nicely and the compiler does wonderful things, but it’s possible that siphash will end up always with a jump, which would add more to the extra overhead (and that llvm-mca doesn’t really model well, so it was harder to prove in sim). Again, not quite sure the real world impact to real world workloads.

Here is the llvm-mca result. Base commit is 1671bcfd76fd. In Total Cycles, siphash is much slower than others.

On Skylake: jhash2: +10%, hsiphash +8%, siphash: +23% On Nehalem: jahsh2: +5%, hsiphash +9%, siphash: +26%

  Skylake             	base		jhash2                  hsiphash                siphash                 
  Iterations          	  1.00		  1.00 (100.00)		  1.00 (100.00)		  1.00 (100.00)		
  Instructions        	104.00		137.00 (131.73)		166.00 (159.62)		222.00 (213.46)		
  Total Cycles        	155.00		171.00 (110.32)		168.00 (108.39)		191.00 (123.23)		
  Total uOps          	132.00		166.00 (125.76)		192.00 (145.45)		248.00 (187.88)		
  Dispatch Width      	  6.00		  6.00 (100.00)		  6.00 (100.00)		  6.00 (100.00)		
  uOps Per Cycle      	  0.85		  0.97 (114.12)		  1.14 (134.12)		  1.30 (152.94)		
  IPC                 	  0.67		  0.80 (119.40)		  0.99 (147.76)		  1.16 (173.13)		
  Nehalem             	base		jhash2                  hsiphash                siphash                 
  Iterations          	  1.00		  1.00 (100.00)		  1.00 (100.00)		  1.00 (100.00)		
  Instructions        	104.00		137.00 (131.73)		166.00 (159.62)		222.00 (213.46)		
  Total Cycles        	173.00		182.00 (105.20)		190.00 (109.83)		218.00 (126.01)		
  Total uOps          	130.00		170.00 (130.77)		215.00 (165.38)		295.00 (226.92)		
  Dispatch Width      	  4.00		  4.00 (100.00)		  4.00 (100.00)		  4.00 (100.00)		
  uOps Per Cycle      	  0.75		  0.93 (124.00)		  1.13 (150.67)		  1.35 (180.00)		
  IPC                 	  0.60		  0.75 (125.00)		  0.87 (145.00)		  1.02 (170.00)	

However, inet6_ehashfn() is not dominant in the fast path, and there seems to be not so big impact during benchmarking with super_netperf.

  for i in $(seq 1 5);
  do
      for i in 64 128 256 512 1024 2048 4096;
      do
  	printf "%4d: " $i;
  	./super_netperf 32 -H ::1 -l 60 -- -m $i -M $i;
      done
  done

Average of 5 run:

$i 10^6 bps
size base jhash2 hsiphash siphash
64 20735.80 20419.96 (98.48) 20335.60 (98.07) 20404.08 (98.40)
128 40450.40 40167.66 (99.30) 39957.32 (98.78) 40222.06 (99.44)
256 73740.84 74219.26 (100.65) 73625.76 (99.84) 74081.52 (100.46)
512 142075.60 142461.80 (100.27) 141237.00 (99.41) 141789.60 (99.80)
1024 254391.00 253553.20 (99.67) 252751.00 (99.36) 253718.00 (99.74)
2048 421221.20 421175.60 (99.99) 420142.80 (99.74) 418189.20 (99.28)
4096 599158.20 597958.20 (99.80) 594472.20 (99.22) 598591.00 (99.91)

Also, in the kernel with hsiphash/siphash, there is call op as hsiphash_4u32() and siphash_2u64() are exposed to modules, but both functions did not have jumps.

hsiphash:

  $ gdb -batch -ex 'file vmlinux' -ex 'disassemble inet6_ehashfn' | grep hsiphash
     0xffffffff81de5bcc <+92>:	call   0xffffffff81f637e0 <hsiphash_4u32>
  $ gdb -batch -ex 'file vmlinux' -ex 'disassemble hsiphash_4u32' | grep j
     0xffffffff81f63970 <+400>:	jmp    0xffffffff81f85ec0 <__x86_return_thunk>

siphash:

  $ gdb -batch -ex 'file vmlinux' -ex 'disassemble inet6_ehashfn' | grep siphash
     0xffffffff81de5bc4 <+84>:	call   0xffffffff81f625e0 <siphash_2u64>
  $ gdb -batch -ex 'file vmlinux' -ex 'disassemble siphash_2u64' | grep j
     0xffffffff81f62837 <+599>:	jmp    0xffffffff81f85ec0 <__x86_return_thunk>

All of the data above are uploaded here.
https://github.com/q2ven/inet6_ehashfn

Thanks!

Reviewed-by on v2 1/1
#5 lore raw
In reply to Kuniyuki Iwashima
Quoted 39 lines · Kuniyuki Iwashima

From: Stewart Smith <[email protected]>

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 the full length of IPv6 address in __ipv6_addr_jhash() so that all users can benefit from a more secure version.

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 ~16 cycles on Skylake (from a base of ~155), and an extra ~9 on Nehalem (base of ~173).

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.

Fixes 08dcdbf6a7b9 ("ipv6: use a stronger hash for tcp")
Quoted 4 lines

---
v2:
* Hash all IPv6 bytes once in __ipv6_addr_jhash() instead of reusing some bytes twice.

Thanks, I think we should merge this patch as is, and think about further improvements later.

#6 lore raw
In reply to Kuniyuki Iwashima
This patch was applied to netdev/net.git (main) by Jakub Kicinski <[email protected]>:

Hello:

This patch was applied to netdev/net.git (main)
by Jakub Kicinski <[email protected]>:

Quoted 8 lines · On Fri, 21 Jul 2023 15:24:10 -0700 you wrote:

From: Stewart Smith <[email protected]>

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.

[...]

Here is the summary with links:
- [v2,net] tcp: Reduce chance of collisions in inet6_hashfn().

    https://git.kernel.org/netdev/net/c/d11b0df7ddf1

You are awesome, thank you!

Signature · 2 lines
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html