From: Gabriel Krisman Bertazi krisman@suse.de
stable inclusion from stable-v6.6.33 commit 71d865be7c2f58bdfb95528e8e5416f71f25ad4a category: bugfix bugzilla: https://gitee.com/openeuler/kernel/issues/IAD6H2
Reference: https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/commit/?id=...
--------------------------------
[ Upstream commit 50aee97d15113b95a68848db1f0cb2a6c09f753a ]
We've observed a 7-12% performance regression in iperf3 UDP ipv4 and ipv6 tests with multiple sockets on Zen3 cpus, which we traced back to commit f0ea27e7bfe1 ("udp: re-score reuseport groups when connected sockets are present"). The failing tests were those that would spawn UDP sockets per-cpu on systems that have a high number of cpus.
Unsurprisingly, it is not caused by the extra re-scoring of the reused socket, but due to the compiler no longer inlining compute_score, once it has the extra call site in udp4_lib_lookup2. This is augmented by the "Safe RET" mitigation for SRSO, needed in our Zen3 cpus.
We could just explicitly inline it, but compute_score() is quite a large function, around 300b. Inlining in two sites would almost double udp4_lib_lookup2, which is a silly thing to do just to workaround a mitigation. Instead, this patch shuffles the code a bit to avoid the multiple calls to compute_score. Since it is a static function used in one spot, the compiler can safely fold it in, as it did before, without increasing the text size.
With this patch applied I ran my original iperf3 testcases. The failing cases all looked like this (ipv4): iperf3 -c 127.0.0.1 --udp -4 -f K -b $R -l 8920 -t 30 -i 5 -P 64 -O 2
where $R is either 1G/10G/0 (max, unlimited). I ran 3 times each. baseline is v6.9-rc3. harmean == harmonic mean; CV == coefficient of variation.
ipv4: 1G 10G MAX HARMEAN (CV) HARMEAN (CV) HARMEAN (CV) baseline 1743852.66(0.0208) 1725933.02(0.0167) 1705203.78(0.0386) patched 1968727.61(0.0035) 1962283.22(0.0195) 1923853.50(0.0256)
ipv6: 1G 10G MAX HARMEAN (CV) HARMEAN (CV) HARMEAN (CV) baseline 1729020.03(0.0028) 1691704.49(0.0243) 1692251.34(0.0083) patched 1900422.19(0.0067) 1900968.01(0.0067) 1568532.72(0.1519)
This restores the performance we had before the change above with this benchmark. We obviously don't expect any real impact when mitigations are disabled, but just to be sure it also doesn't regresses:
mitigations=off ipv4: 1G 10G MAX HARMEAN (CV) HARMEAN (CV) HARMEAN (CV) baseline 3230279.97(0.0066) 3229320.91(0.0060) 2605693.19(0.0697) patched 3242802.36(0.0073) 3239310.71(0.0035) 2502427.19(0.0882)
Cc: Lorenz Bauer lmb@isovalent.com Fixes: f0ea27e7bfe1 ("udp: re-score reuseport groups when connected sockets are present") Signed-off-by: Gabriel Krisman Bertazi krisman@suse.de Reviewed-by: Kuniyuki Iwashima kuniyu@amazon.com Reviewed-by: Willem de Bruijn willemb@google.com Signed-off-by: David S. Miller davem@davemloft.net Signed-off-by: Sasha Levin sashal@kernel.org Signed-off-by: Wang Hai wanghai38@huawei.com --- net/ipv4/udp.c | 21 ++++++++++++++++----- net/ipv6/udp.c | 20 ++++++++++++++++---- 2 files changed, 32 insertions(+), 9 deletions(-)
diff --git a/net/ipv4/udp.c b/net/ipv4/udp.c index ec07d9429ead..faeb95d92f02 100644 --- a/net/ipv4/udp.c +++ b/net/ipv4/udp.c @@ -429,15 +429,21 @@ static struct sock *udp4_lib_lookup2(struct net *net, { struct sock *sk, *result; int score, badness; + bool need_rescore;
result = NULL; badness = 0; udp_portaddr_for_each_entry_rcu(sk, &hslot2->head) { - score = compute_score(sk, net, saddr, sport, - daddr, hnum, dif, sdif); + need_rescore = false; +rescore: + score = compute_score(need_rescore ? result : sk, net, saddr, + sport, daddr, hnum, dif, sdif); if (score > badness) { badness = score;
+ if (need_rescore) + continue; + if (sk->sk_state == TCP_ESTABLISHED) { result = sk; continue; @@ -458,9 +464,14 @@ static struct sock *udp4_lib_lookup2(struct net *net, if (IS_ERR(result)) continue;
- badness = compute_score(result, net, saddr, sport, - daddr, hnum, dif, sdif); - + /* compute_score is too long of a function to be + * inlined, and calling it again here yields + * measureable overhead for some + * workloads. Work around it by jumping + * backwards to rescore 'result'. + */ + need_rescore = true; + goto rescore; } } return result; diff --git a/net/ipv6/udp.c b/net/ipv6/udp.c index 124cf2bb2a6d..c77ee9a3cde2 100644 --- a/net/ipv6/udp.c +++ b/net/ipv6/udp.c @@ -171,15 +171,21 @@ static struct sock *udp6_lib_lookup2(struct net *net, { struct sock *sk, *result; int score, badness; + bool need_rescore;
result = NULL; badness = -1; udp_portaddr_for_each_entry_rcu(sk, &hslot2->head) { - score = compute_score(sk, net, saddr, sport, - daddr, hnum, dif, sdif); + need_rescore = false; +rescore: + score = compute_score(need_rescore ? result : sk, net, saddr, + sport, daddr, hnum, dif, sdif); if (score > badness) { badness = score;
+ if (need_rescore) + continue; + if (sk->sk_state == TCP_ESTABLISHED) { result = sk; continue; @@ -200,8 +206,14 @@ static struct sock *udp6_lib_lookup2(struct net *net, if (IS_ERR(result)) continue;
- badness = compute_score(sk, net, saddr, sport, - daddr, hnum, dif, sdif); + /* compute_score is too long of a function to be + * inlined, and calling it again here yields + * measureable overhead for some + * workloads. Work around it by jumping + * backwards to rescore 'result'. + */ + need_rescore = true; + goto rescore; } } return result;