| Message ID | dafe09ca2e14c4ab45f3d9db56b768e06750e382.1666173045.git.pabeni@redhat.com (mailing list archive) |
|---|---|
| State | Superseded |
| Delegated to: | Netdev Maintainers |
| Headers | show |
| Series | udp: avoid false sharing on receive | expand |
From: Paolo Abeni <pabeni@redhat.com> Date: Wed, 19 Oct 2022 12:02:01 +0200 > When the receiver process and the BH runs on different cores, > udp_rmem_release() experience a cache miss while accessing sk_rcvbuf, > as the latter shares the same cacheline with sk_forward_alloc, written > by the BH. > > With this patch, UDP tracks the rcvbuf value and its update via custom > SOL_SOCKET socket options, and copies the forward memory threshold value > used by udp_rmem_release() in a different cacheline, already accessed by > the above function and uncontended. > > Overall the above give a 10% peek throughput increase under UDP flood. > > Signed-off-by: Paolo Abeni <pabeni@redhat.com> > --- > include/linux/udp.h | 3 +++ > net/ipv4/udp.c | 22 +++++++++++++++++++--- > net/ipv6/udp.c | 8 ++++++-- > 3 files changed, 28 insertions(+), 5 deletions(-) > > diff --git a/include/linux/udp.h b/include/linux/udp.h > index e96da4157d04..5cdba00a904a 100644 > --- a/include/linux/udp.h > +++ b/include/linux/udp.h > @@ -87,6 +87,9 @@ struct udp_sock { > > /* This field is dirtied by udp_recvmsg() */ > int forward_deficit; > + > + /* This fields follows rcvbuf value, and is touched by udp_recvmsg */ > + int forward_threshold; > }; > > #define UDP_MAX_SEGMENTS (1 << 6UL) > diff --git a/net/ipv4/udp.c b/net/ipv4/udp.c > index 8126f67d18b3..915f573587fa 100644 > --- a/net/ipv4/udp.c > +++ b/net/ipv4/udp.c > @@ -1448,7 +1448,7 @@ static void udp_rmem_release(struct sock *sk, int size, int partial, > if (likely(partial)) { > up->forward_deficit += size; > size = up->forward_deficit; > - if (size < (sk->sk_rcvbuf >> 2) && > + if (size < READ_ONCE(up->forward_threshold) && > !skb_queue_empty(&up->reader_queue)) > return; > } else { > @@ -1622,8 +1622,12 @@ static void udp_destruct_sock(struct sock *sk) > > int udp_init_sock(struct sock *sk) > { > - skb_queue_head_init(&udp_sk(sk)->reader_queue); > + struct udp_sock *up = udp_sk(sk); > + > + skb_queue_head_init(&up->reader_queue); > + up->forward_threshold = sk->sk_rcvbuf >> 2; > sk->sk_destruct = udp_destruct_sock; > + set_bit(SOCK_CUSTOM_SOCKOPT, &sk->sk_socket->flags); > return 0; > } > > @@ -2671,6 +2675,18 @@ int udp_lib_setsockopt(struct sock *sk, int level, int optname, > int err = 0; > int is_udplite = IS_UDPLITE(sk); > > + if (level == SOL_SOCKET) { > + err = sk_setsockopt(sk, level, optname, optval, optlen); > + > + if (optname == SO_RCVBUF || optname == SO_RCVBUFFORCE) { > + sockopt_lock_sock(sk); Can we drop this lock by adding READ_ONCE() to sk->sk_rcvbuf below ? > + /* paired with READ_ONCE in udp_rmem_release() */ > + WRITE_ONCE(up->forward_threshold, sk->sk_rcvbuf >> 2); > + sockopt_release_sock(sk); > + } > + return err; > + } > + > if (optlen < sizeof(int)) > return -EINVAL; > > @@ -2784,7 +2800,7 @@ EXPORT_SYMBOL(udp_lib_setsockopt); > int udp_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval, > unsigned int optlen) > { > - if (level == SOL_UDP || level == SOL_UDPLITE) > + if (level == SOL_UDP || level == SOL_UDPLITE || level == SOL_SOCKET) > return udp_lib_setsockopt(sk, level, optname, > optval, optlen, > udp_push_pending_frames); > diff --git a/net/ipv6/udp.c b/net/ipv6/udp.c > index 8d09f0ea5b8c..1ed20bcfd7a0 100644 > --- a/net/ipv6/udp.c > +++ b/net/ipv6/udp.c > @@ -64,8 +64,12 @@ static void udpv6_destruct_sock(struct sock *sk) > > int udpv6_init_sock(struct sock *sk) > { > - skb_queue_head_init(&udp_sk(sk)->reader_queue); > + struct udp_sock *up = udp_sk(sk); > + > + skb_queue_head_init(&up->reader_queue); > + up->forward_threshold = sk->sk_rcvbuf >> 2; > sk->sk_destruct = udpv6_destruct_sock; > + set_bit(SOCK_CUSTOM_SOCKOPT, &sk->sk_socket->flags); > return 0; > } It's time to factorise this part like udp_destruct_common() ? > > @@ -1671,7 +1675,7 @@ void udpv6_destroy_sock(struct sock *sk) > int udpv6_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval, > unsigned int optlen) > { > - if (level == SOL_UDP || level == SOL_UDPLITE) > + if (level == SOL_UDP || level == SOL_UDPLITE || level == SOL_SOCKET) > return udp_lib_setsockopt(sk, level, optname, > optval, optlen, > udp_v6_push_pending_frames); > -- > 2.37.3
On Wed, 2022-10-19 at 09:33 -0700, Kuniyuki Iwashima wrote: > From: Paolo Abeni <pabeni@redhat.com> > Date: Wed, 19 Oct 2022 12:02:01 +0200 > > When the receiver process and the BH runs on different cores, > > udp_rmem_release() experience a cache miss while accessing sk_rcvbuf, > > as the latter shares the same cacheline with sk_forward_alloc, written > > by the BH. > > > > With this patch, UDP tracks the rcvbuf value and its update via custom > > SOL_SOCKET socket options, and copies the forward memory threshold value > > used by udp_rmem_release() in a different cacheline, already accessed by > > the above function and uncontended. > > > > Overall the above give a 10% peek throughput increase under UDP flood. > > > > Signed-off-by: Paolo Abeni <pabeni@redhat.com> > > --- > > include/linux/udp.h | 3 +++ > > net/ipv4/udp.c | 22 +++++++++++++++++++--- > > net/ipv6/udp.c | 8 ++++++-- > > 3 files changed, 28 insertions(+), 5 deletions(-) > > > > diff --git a/include/linux/udp.h b/include/linux/udp.h > > index e96da4157d04..5cdba00a904a 100644 > > --- a/include/linux/udp.h > > +++ b/include/linux/udp.h > > @@ -87,6 +87,9 @@ struct udp_sock { > > > > /* This field is dirtied by udp_recvmsg() */ > > int forward_deficit; > > + > > + /* This fields follows rcvbuf value, and is touched by udp_recvmsg */ > > + int forward_threshold; > > }; > > > > #define UDP_MAX_SEGMENTS (1 << 6UL) > > diff --git a/net/ipv4/udp.c b/net/ipv4/udp.c > > index 8126f67d18b3..915f573587fa 100644 > > --- a/net/ipv4/udp.c > > +++ b/net/ipv4/udp.c > > @@ -1448,7 +1448,7 @@ static void udp_rmem_release(struct sock *sk, int size, int partial, > > if (likely(partial)) { > > up->forward_deficit += size; > > size = up->forward_deficit; > > - if (size < (sk->sk_rcvbuf >> 2) && > > + if (size < READ_ONCE(up->forward_threshold) && > > !skb_queue_empty(&up->reader_queue)) > > return; > > } else { > > @@ -1622,8 +1622,12 @@ static void udp_destruct_sock(struct sock *sk) > > > > int udp_init_sock(struct sock *sk) > > { > > - skb_queue_head_init(&udp_sk(sk)->reader_queue); > > + struct udp_sock *up = udp_sk(sk); > > + > > + skb_queue_head_init(&up->reader_queue); > > + up->forward_threshold = sk->sk_rcvbuf >> 2; > > sk->sk_destruct = udp_destruct_sock; > > + set_bit(SOCK_CUSTOM_SOCKOPT, &sk->sk_socket->flags); > > return 0; > > } > > > > @@ -2671,6 +2675,18 @@ int udp_lib_setsockopt(struct sock *sk, int level, int optname, > > int err = 0; > > int is_udplite = IS_UDPLITE(sk); > > > > + if (level == SOL_SOCKET) { > > + err = sk_setsockopt(sk, level, optname, optval, optlen); > > + > > + if (optname == SO_RCVBUF || optname == SO_RCVBUFFORCE) { > > + sockopt_lock_sock(sk); > > Can we drop this lock by adding READ_ONCE() to sk->sk_rcvbuf below ? I think we can't. If there are racing thread updating rcvbuf, we could end-up with mismatching value in forward_threshold. Not a likely scenario, but still... This is control path, acquiring the lock once more should not be a problem. > > + /* paired with READ_ONCE in udp_rmem_release() */ > > + WRITE_ONCE(up->forward_threshold, sk->sk_rcvbuf >> 2); > > + sockopt_release_sock(sk); > > + } > > + return err; > > + } > > + > > if (optlen < sizeof(int)) > > return -EINVAL; > > > > @@ -2784,7 +2800,7 @@ EXPORT_SYMBOL(udp_lib_setsockopt); > > int udp_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval, > > unsigned int optlen) > > { > > - if (level == SOL_UDP || level == SOL_UDPLITE) > > + if (level == SOL_UDP || level == SOL_UDPLITE || level == SOL_SOCKET) > > return udp_lib_setsockopt(sk, level, optname, > > optval, optlen, > > udp_push_pending_frames); > > diff --git a/net/ipv6/udp.c b/net/ipv6/udp.c > > index 8d09f0ea5b8c..1ed20bcfd7a0 100644 > > --- a/net/ipv6/udp.c > > +++ b/net/ipv6/udp.c > > @@ -64,8 +64,12 @@ static void udpv6_destruct_sock(struct sock *sk) > > > > int udpv6_init_sock(struct sock *sk) > > { > > - skb_queue_head_init(&udp_sk(sk)->reader_queue); > > + struct udp_sock *up = udp_sk(sk); > > + > > + skb_queue_head_init(&up->reader_queue); > > + up->forward_threshold = sk->sk_rcvbuf >> 2; > > sk->sk_destruct = udpv6_destruct_sock; > > + set_bit(SOCK_CUSTOM_SOCKOPT, &sk->sk_socket->flags); > > return 0; > > } > > It's time to factorise this part like udp_destruct_common() ? I guess it makes sense. Possibly 'udp_lib_destruct()' just to follow others helper style? > Thanks, Paolo
From: Paolo Abeni <pabeni@redhat.com> Date: Wed, 19 Oct 2022 18:58:33 +0200 > On Wed, 2022-10-19 at 09:33 -0700, Kuniyuki Iwashima wrote: > > From: Paolo Abeni <pabeni@redhat.com> > > Date: Wed, 19 Oct 2022 12:02:01 +0200 > > > When the receiver process and the BH runs on different cores, > > > udp_rmem_release() experience a cache miss while accessing sk_rcvbuf, > > > as the latter shares the same cacheline with sk_forward_alloc, written > > > by the BH. > > > > > > With this patch, UDP tracks the rcvbuf value and its update via custom > > > SOL_SOCKET socket options, and copies the forward memory threshold value > > > used by udp_rmem_release() in a different cacheline, already accessed by > > > the above function and uncontended. > > > > > > Overall the above give a 10% peek throughput increase under UDP flood. > > > > > > Signed-off-by: Paolo Abeni <pabeni@redhat.com> > > > --- > > > include/linux/udp.h | 3 +++ > > > net/ipv4/udp.c | 22 +++++++++++++++++++--- > > > net/ipv6/udp.c | 8 ++++++-- > > > 3 files changed, 28 insertions(+), 5 deletions(-) > > > > > > diff --git a/include/linux/udp.h b/include/linux/udp.h > > > index e96da4157d04..5cdba00a904a 100644 > > > --- a/include/linux/udp.h > > > +++ b/include/linux/udp.h > > > @@ -87,6 +87,9 @@ struct udp_sock { > > > > > > /* This field is dirtied by udp_recvmsg() */ > > > int forward_deficit; > > > + > > > + /* This fields follows rcvbuf value, and is touched by udp_recvmsg */ > > > + int forward_threshold; > > > }; > > > > > > #define UDP_MAX_SEGMENTS (1 << 6UL) > > > diff --git a/net/ipv4/udp.c b/net/ipv4/udp.c > > > index 8126f67d18b3..915f573587fa 100644 > > > --- a/net/ipv4/udp.c > > > +++ b/net/ipv4/udp.c > > > @@ -1448,7 +1448,7 @@ static void udp_rmem_release(struct sock *sk, int size, int partial, > > > if (likely(partial)) { > > > up->forward_deficit += size; > > > size = up->forward_deficit; > > > - if (size < (sk->sk_rcvbuf >> 2) && > > > + if (size < READ_ONCE(up->forward_threshold) && > > > !skb_queue_empty(&up->reader_queue)) > > > return; > > > } else { > > > @@ -1622,8 +1622,12 @@ static void udp_destruct_sock(struct sock *sk) > > > > > > int udp_init_sock(struct sock *sk) > > > { > > > - skb_queue_head_init(&udp_sk(sk)->reader_queue); > > > + struct udp_sock *up = udp_sk(sk); > > > + > > > + skb_queue_head_init(&up->reader_queue); > > > + up->forward_threshold = sk->sk_rcvbuf >> 2; > > > sk->sk_destruct = udp_destruct_sock; > > > + set_bit(SOCK_CUSTOM_SOCKOPT, &sk->sk_socket->flags); > > > return 0; > > > } > > > > > > @@ -2671,6 +2675,18 @@ int udp_lib_setsockopt(struct sock *sk, int level, int optname, > > > int err = 0; > > > int is_udplite = IS_UDPLITE(sk); > > > > > > + if (level == SOL_SOCKET) { > > > + err = sk_setsockopt(sk, level, optname, optval, optlen); > > > + > > > + if (optname == SO_RCVBUF || optname == SO_RCVBUFFORCE) { > > > + sockopt_lock_sock(sk); > > > > Can we drop this lock by adding READ_ONCE() to sk->sk_rcvbuf below ? > > I think we can't. If there are racing thread updating rcvbuf, we could > end-up with mismatching value in forward_threshold. Not a likely > scenario, but still... This is control path, acquiring the lock once > more should not be a problem. I see. Thank you! > > > + /* paired with READ_ONCE in udp_rmem_release() */ > > > + WRITE_ONCE(up->forward_threshold, sk->sk_rcvbuf >> 2); > > > + sockopt_release_sock(sk); > > > + } > > > + return err; > > > + } > > > + > > > if (optlen < sizeof(int)) > > > return -EINVAL; > > > > > > @@ -2784,7 +2800,7 @@ EXPORT_SYMBOL(udp_lib_setsockopt); > > > int udp_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval, > > > unsigned int optlen) > > > { > > > - if (level == SOL_UDP || level == SOL_UDPLITE) > > > + if (level == SOL_UDP || level == SOL_UDPLITE || level == SOL_SOCKET) > > > return udp_lib_setsockopt(sk, level, optname, > > > optval, optlen, > > > udp_push_pending_frames); > > > diff --git a/net/ipv6/udp.c b/net/ipv6/udp.c > > > index 8d09f0ea5b8c..1ed20bcfd7a0 100644 > > > --- a/net/ipv6/udp.c > > > +++ b/net/ipv6/udp.c > > > @@ -64,8 +64,12 @@ static void udpv6_destruct_sock(struct sock *sk) > > > > > > int udpv6_init_sock(struct sock *sk) > > > { > > > - skb_queue_head_init(&udp_sk(sk)->reader_queue); > > > + struct udp_sock *up = udp_sk(sk); > > > + > > > + skb_queue_head_init(&up->reader_queue); > > > + up->forward_threshold = sk->sk_rcvbuf >> 2; > > > sk->sk_destruct = udpv6_destruct_sock; > > > + set_bit(SOCK_CUSTOM_SOCKOPT, &sk->sk_socket->flags); > > > return 0; > > > } > > > > It's time to factorise this part like udp_destruct_common() ? > > I guess it makes sense. Possibly 'udp_lib_destruct()' just to follow > others helper style? Ah, I should have named it so :) > > > > Thanks, > > Paolo
diff --git a/include/linux/udp.h b/include/linux/udp.h index e96da4157d04..5cdba00a904a 100644 --- a/include/linux/udp.h +++ b/include/linux/udp.h @@ -87,6 +87,9 @@ struct udp_sock { /* This field is dirtied by udp_recvmsg() */ int forward_deficit; + + /* This fields follows rcvbuf value, and is touched by udp_recvmsg */ + int forward_threshold; }; #define UDP_MAX_SEGMENTS (1 << 6UL) diff --git a/net/ipv4/udp.c b/net/ipv4/udp.c index 8126f67d18b3..915f573587fa 100644 --- a/net/ipv4/udp.c +++ b/net/ipv4/udp.c @@ -1448,7 +1448,7 @@ static void udp_rmem_release(struct sock *sk, int size, int partial, if (likely(partial)) { up->forward_deficit += size; size = up->forward_deficit; - if (size < (sk->sk_rcvbuf >> 2) && + if (size < READ_ONCE(up->forward_threshold) && !skb_queue_empty(&up->reader_queue)) return; } else { @@ -1622,8 +1622,12 @@ static void udp_destruct_sock(struct sock *sk) int udp_init_sock(struct sock *sk) { - skb_queue_head_init(&udp_sk(sk)->reader_queue); + struct udp_sock *up = udp_sk(sk); + + skb_queue_head_init(&up->reader_queue); + up->forward_threshold = sk->sk_rcvbuf >> 2; sk->sk_destruct = udp_destruct_sock; + set_bit(SOCK_CUSTOM_SOCKOPT, &sk->sk_socket->flags); return 0; } @@ -2671,6 +2675,18 @@ int udp_lib_setsockopt(struct sock *sk, int level, int optname, int err = 0; int is_udplite = IS_UDPLITE(sk); + if (level == SOL_SOCKET) { + err = sk_setsockopt(sk, level, optname, optval, optlen); + + if (optname == SO_RCVBUF || optname == SO_RCVBUFFORCE) { + sockopt_lock_sock(sk); + /* paired with READ_ONCE in udp_rmem_release() */ + WRITE_ONCE(up->forward_threshold, sk->sk_rcvbuf >> 2); + sockopt_release_sock(sk); + } + return err; + } + if (optlen < sizeof(int)) return -EINVAL; @@ -2784,7 +2800,7 @@ EXPORT_SYMBOL(udp_lib_setsockopt); int udp_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval, unsigned int optlen) { - if (level == SOL_UDP || level == SOL_UDPLITE) + if (level == SOL_UDP || level == SOL_UDPLITE || level == SOL_SOCKET) return udp_lib_setsockopt(sk, level, optname, optval, optlen, udp_push_pending_frames); diff --git a/net/ipv6/udp.c b/net/ipv6/udp.c index 8d09f0ea5b8c..1ed20bcfd7a0 100644 --- a/net/ipv6/udp.c +++ b/net/ipv6/udp.c @@ -64,8 +64,12 @@ static void udpv6_destruct_sock(struct sock *sk) int udpv6_init_sock(struct sock *sk) { - skb_queue_head_init(&udp_sk(sk)->reader_queue); + struct udp_sock *up = udp_sk(sk); + + skb_queue_head_init(&up->reader_queue); + up->forward_threshold = sk->sk_rcvbuf >> 2; sk->sk_destruct = udpv6_destruct_sock; + set_bit(SOCK_CUSTOM_SOCKOPT, &sk->sk_socket->flags); return 0; } @@ -1671,7 +1675,7 @@ void udpv6_destroy_sock(struct sock *sk) int udpv6_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval, unsigned int optlen) { - if (level == SOL_UDP || level == SOL_UDPLITE) + if (level == SOL_UDP || level == SOL_UDPLITE || level == SOL_SOCKET) return udp_lib_setsockopt(sk, level, optname, optval, optlen, udp_v6_push_pending_frames);
When the receiver process and the BH runs on different cores, udp_rmem_release() experience a cache miss while accessing sk_rcvbuf, as the latter shares the same cacheline with sk_forward_alloc, written by the BH. With this patch, UDP tracks the rcvbuf value and its update via custom SOL_SOCKET socket options, and copies the forward memory threshold value used by udp_rmem_release() in a different cacheline, already accessed by the above function and uncontended. Overall the above give a 10% peek throughput increase under UDP flood. Signed-off-by: Paolo Abeni <pabeni@redhat.com> --- include/linux/udp.h | 3 +++ net/ipv4/udp.c | 22 +++++++++++++++++++--- net/ipv6/udp.c | 8 ++++++-- 3 files changed, 28 insertions(+), 5 deletions(-)