* [Patch bpf-next v2 0/4] sockmap: some performance optimizations
@ 2022-05-02 18:23 Cong Wang
2022-05-02 18:23 ` [Patch bpf-next v2 1/4] tcp: introduce tcp_read_skb() Cong Wang
` (3 more replies)
0 siblings, 4 replies; 7+ messages in thread
From: Cong Wang @ 2022-05-02 18:23 UTC (permalink / raw)
To: netdev; +Cc: bpf, Cong Wang
From: Cong Wang <cong.wang@bytedance.com>
This patchset contains two optimizations for sockmap. The first one
eliminates a skb_clone() and the second one eliminates a memset(). After
this patchset, the throughput of UDP transmission via sockmap gets
improved by 61%.
v2: clean up coding style for tcp_read_skb()
get rid of some redundant variables
add a comment for ->read_skb()
Cong Wang (4):
tcp: introduce tcp_read_skb()
net: introduce a new proto_ops ->read_skb()
skmsg: get rid of skb_clone()
skmsg: get rid of unncessary memset()
include/linux/net.h | 4 +++
include/net/tcp.h | 1 +
include/net/udp.h | 3 +--
net/core/skmsg.c | 48 ++++++++++++++----------------------
net/ipv4/af_inet.c | 3 ++-
net/ipv4/tcp.c | 60 +++++++++++++++++++++++++++++++++++++++------
net/ipv4/udp.c | 11 ++++-----
net/ipv6/af_inet6.c | 3 ++-
net/unix/af_unix.c | 23 +++++++----------
9 files changed, 94 insertions(+), 62 deletions(-)
--
2.32.0
^ permalink raw reply [flat|nested] 7+ messages in thread* [Patch bpf-next v2 1/4] tcp: introduce tcp_read_skb() 2022-05-02 18:23 [Patch bpf-next v2 0/4] sockmap: some performance optimizations Cong Wang @ 2022-05-02 18:23 ` Cong Wang 2022-05-03 0:02 ` Eric Dumazet 2022-05-02 18:23 ` [Patch bpf-next v2 2/4] net: introduce a new proto_ops ->read_skb() Cong Wang ` (2 subsequent siblings) 3 siblings, 1 reply; 7+ messages in thread From: Cong Wang @ 2022-05-02 18:23 UTC (permalink / raw) To: netdev Cc: bpf, Cong Wang, Eric Dumazet, John Fastabend, Daniel Borkmann, Jakub Sitnicki From: Cong Wang <cong.wang@bytedance.com> This patch inroduces tcp_read_skb() based on tcp_read_sock(), a preparation for the next patch which actually introduces a new sock ops. TCP is special here, because it has tcp_read_sock() which is mainly used by splice(). tcp_read_sock() supports partial read and arbitrary offset, neither of them is needed for sockmap. Cc: Eric Dumazet <edumazet@google.com> Cc: John Fastabend <john.fastabend@gmail.com> Cc: Daniel Borkmann <daniel@iogearbox.net> Cc: Jakub Sitnicki <jakub@cloudflare.com> Signed-off-by: Cong Wang <cong.wang@bytedance.com> --- include/net/tcp.h | 2 ++ net/ipv4/tcp.c | 63 +++++++++++++++++++++++++++++++++++++++++------ 2 files changed, 57 insertions(+), 8 deletions(-) diff --git a/include/net/tcp.h b/include/net/tcp.h index 94a52ad1101c..ab7516e5cc56 100644 --- a/include/net/tcp.h +++ b/include/net/tcp.h @@ -667,6 +667,8 @@ void tcp_get_info(struct sock *, struct tcp_info *); /* Read 'sendfile()'-style from a TCP socket */ int tcp_read_sock(struct sock *sk, read_descriptor_t *desc, sk_read_actor_t recv_actor); +int tcp_read_skb(struct sock *sk, read_descriptor_t *desc, + sk_read_actor_t recv_actor); void tcp_initialize_rcv_mss(struct sock *sk); diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c index db55af9eb37b..8d48126e3694 100644 --- a/net/ipv4/tcp.c +++ b/net/ipv4/tcp.c @@ -1600,7 +1600,7 @@ static void tcp_eat_recv_skb(struct sock *sk, struct sk_buff *skb) __kfree_skb(skb); } -static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off) +static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off, bool unlink) { struct sk_buff *skb; u32 offset; @@ -1613,6 +1613,8 @@ static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off) } if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) { *off = offset; + if (unlink) + __skb_unlink(skb, &sk->sk_receive_queue); return skb; } /* This looks weird, but this can happen if TCP collapsing @@ -1646,7 +1648,7 @@ int tcp_read_sock(struct sock *sk, read_descriptor_t *desc, if (sk->sk_state == TCP_LISTEN) return -ENOTCONN; - while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) { + while ((skb = tcp_recv_skb(sk, seq, &offset, false)) != NULL) { if (offset < skb->len) { int used; size_t len; @@ -1677,7 +1679,7 @@ int tcp_read_sock(struct sock *sk, read_descriptor_t *desc, * getting here: tcp_collapse might have deleted it * while aggregating skbs from the socket queue. */ - skb = tcp_recv_skb(sk, seq - 1, &offset); + skb = tcp_recv_skb(sk, seq - 1, &offset, false); if (!skb) break; /* TCP coalescing might have appended data to the skb. @@ -1702,13 +1704,58 @@ int tcp_read_sock(struct sock *sk, read_descriptor_t *desc, /* Clean up data we have read: This will do ACK frames. */ if (copied > 0) { - tcp_recv_skb(sk, seq, &offset); + tcp_recv_skb(sk, seq, &offset, false); tcp_cleanup_rbuf(sk, copied); } return copied; } EXPORT_SYMBOL(tcp_read_sock); +int tcp_read_skb(struct sock *sk, read_descriptor_t *desc, + sk_read_actor_t recv_actor) +{ + struct tcp_sock *tp = tcp_sk(sk); + u32 seq = tp->copied_seq; + struct sk_buff *skb; + int copied = 0; + u32 offset; + + if (sk->sk_state == TCP_LISTEN) + return -ENOTCONN; + + while ((skb = tcp_recv_skb(sk, seq, &offset, true)) != NULL) { + int used = recv_actor(desc, skb, 0, skb->len); + + if (used <= 0) { + if (!copied) + copied = used; + break; + } + seq += used; + copied += used; + + if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) { + kfree_skb(skb); + ++seq; + break; + } + kfree_skb(skb); + if (!desc->count) + break; + WRITE_ONCE(tp->copied_seq, seq); + } + WRITE_ONCE(tp->copied_seq, seq); + + tcp_rcv_space_adjust(sk); + + /* Clean up data we have read: This will do ACK frames. */ + if (copied > 0) + tcp_cleanup_rbuf(sk, copied); + + return copied; +} +EXPORT_SYMBOL(tcp_read_skb); + int tcp_peek_len(struct socket *sock) { return tcp_inq(sock->sk); @@ -1890,7 +1937,7 @@ static int receive_fallback_to_copy(struct sock *sk, struct sk_buff *skb; u32 offset; - skb = tcp_recv_skb(sk, tcp_sk(sk)->copied_seq, &offset); + skb = tcp_recv_skb(sk, tcp_sk(sk)->copied_seq, &offset, false); if (skb) tcp_zerocopy_set_hint_for_skb(sk, zc, skb, offset); } @@ -1937,7 +1984,7 @@ static int tcp_zc_handle_leftover(struct tcp_zerocopy_receive *zc, if (skb) { offset = *seq - TCP_SKB_CB(skb)->seq; } else { - skb = tcp_recv_skb(sk, *seq, &offset); + skb = tcp_recv_skb(sk, *seq, &offset, false); if (TCP_SKB_CB(skb)->has_rxtstamp) { tcp_update_recv_tstamps(skb, tss); zc->msg_flags |= TCP_CMSG_TS; @@ -2130,7 +2177,7 @@ static int tcp_zerocopy_receive(struct sock *sk, skb = skb->next; offset = seq - TCP_SKB_CB(skb)->seq; } else { - skb = tcp_recv_skb(sk, seq, &offset); + skb = tcp_recv_skb(sk, seq, &offset, false); } if (TCP_SKB_CB(skb)->has_rxtstamp) { @@ -2186,7 +2233,7 @@ static int tcp_zerocopy_receive(struct sock *sk, tcp_rcv_space_adjust(sk); /* Clean up data we have read: This will do ACK frames. */ - tcp_recv_skb(sk, seq, &offset); + tcp_recv_skb(sk, seq, &offset, false); tcp_cleanup_rbuf(sk, length + copylen); ret = 0; if (length == zc->length) -- 2.32.0 ^ permalink raw reply related [flat|nested] 7+ messages in thread
* Re: [Patch bpf-next v2 1/4] tcp: introduce tcp_read_skb() 2022-05-02 18:23 ` [Patch bpf-next v2 1/4] tcp: introduce tcp_read_skb() Cong Wang @ 2022-05-03 0:02 ` Eric Dumazet 2022-05-10 18:27 ` Cong Wang 0 siblings, 1 reply; 7+ messages in thread From: Eric Dumazet @ 2022-05-03 0:02 UTC (permalink / raw) To: Cong Wang Cc: netdev, bpf, Cong Wang, John Fastabend, Daniel Borkmann, Jakub Sitnicki On Mon, May 2, 2022 at 11:24 AM Cong Wang <xiyou.wangcong@gmail.com> wrote: > > From: Cong Wang <cong.wang@bytedance.com> > > This patch inroduces tcp_read_skb() based on tcp_read_sock(), > a preparation for the next patch which actually introduces > a new sock ops. > > TCP is special here, because it has tcp_read_sock() which is > mainly used by splice(). tcp_read_sock() supports partial read > and arbitrary offset, neither of them is needed for sockmap. > > Cc: Eric Dumazet <edumazet@google.com> > Cc: John Fastabend <john.fastabend@gmail.com> > Cc: Daniel Borkmann <daniel@iogearbox.net> > Cc: Jakub Sitnicki <jakub@cloudflare.com> > Signed-off-by: Cong Wang <cong.wang@bytedance.com> > --- > include/net/tcp.h | 2 ++ > net/ipv4/tcp.c | 63 +++++++++++++++++++++++++++++++++++++++++------ > 2 files changed, 57 insertions(+), 8 deletions(-) > > diff --git a/include/net/tcp.h b/include/net/tcp.h > index 94a52ad1101c..ab7516e5cc56 100644 > --- a/include/net/tcp.h > +++ b/include/net/tcp.h > @@ -667,6 +667,8 @@ void tcp_get_info(struct sock *, struct tcp_info *); > /* Read 'sendfile()'-style from a TCP socket */ > int tcp_read_sock(struct sock *sk, read_descriptor_t *desc, > sk_read_actor_t recv_actor); > +int tcp_read_skb(struct sock *sk, read_descriptor_t *desc, > + sk_read_actor_t recv_actor); > > void tcp_initialize_rcv_mss(struct sock *sk); > > diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c > index db55af9eb37b..8d48126e3694 100644 > --- a/net/ipv4/tcp.c > +++ b/net/ipv4/tcp.c > @@ -1600,7 +1600,7 @@ static void tcp_eat_recv_skb(struct sock *sk, struct sk_buff *skb) > __kfree_skb(skb); > } > > -static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off) > +static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off, bool unlink) > { > struct sk_buff *skb; > u32 offset; > @@ -1613,6 +1613,8 @@ static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off) > } > if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) { > *off = offset; > + if (unlink) > + __skb_unlink(skb, &sk->sk_receive_queue); Why adding this @unlink parameter ? This makes your patch more invasive than needed. Can not this unlink happen from your new helper instead ? See [3] later. > return skb; > } > /* This looks weird, but this can happen if TCP collapsing > @@ -1646,7 +1648,7 @@ int tcp_read_sock(struct sock *sk, read_descriptor_t *desc, > > if (sk->sk_state == TCP_LISTEN) > return -ENOTCONN; > - while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) { > + while ((skb = tcp_recv_skb(sk, seq, &offset, false)) != NULL) { > if (offset < skb->len) { > int used; > size_t len; > @@ -1677,7 +1679,7 @@ int tcp_read_sock(struct sock *sk, read_descriptor_t *desc, > * getting here: tcp_collapse might have deleted it > * while aggregating skbs from the socket queue. > */ > - skb = tcp_recv_skb(sk, seq - 1, &offset); > + skb = tcp_recv_skb(sk, seq - 1, &offset, false); > if (!skb) > break; > /* TCP coalescing might have appended data to the skb. > @@ -1702,13 +1704,58 @@ int tcp_read_sock(struct sock *sk, read_descriptor_t *desc, > > /* Clean up data we have read: This will do ACK frames. */ > if (copied > 0) { > - tcp_recv_skb(sk, seq, &offset); > + tcp_recv_skb(sk, seq, &offset, false); > tcp_cleanup_rbuf(sk, copied); > } > return copied; > } > EXPORT_SYMBOL(tcp_read_sock); > > +int tcp_read_skb(struct sock *sk, read_descriptor_t *desc, > + sk_read_actor_t recv_actor) > +{ > + struct tcp_sock *tp = tcp_sk(sk); > + u32 seq = tp->copied_seq; > + struct sk_buff *skb; > + int copied = 0; > + u32 offset; > + > + if (sk->sk_state == TCP_LISTEN) > + return -ENOTCONN; > + > + while ((skb = tcp_recv_skb(sk, seq, &offset, true)) != NULL) { [3] The unlink from sk->sk_receive_queue could happen here. > + int used = recv_actor(desc, skb, 0, skb->len); > + > + if (used <= 0) { > + if (!copied) > + copied = used; > + break; > + } > + seq += used; > + copied += used; > + > + if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) { > + kfree_skb(skb); [1] The two kfree_skb() ([1] & [2]) should be a consume_skb() ? > + ++seq; > + break; > + } > + kfree_skb(skb); [2] > + if (!desc->count) > + break; > + WRITE_ONCE(tp->copied_seq, seq); > + } > + WRITE_ONCE(tp->copied_seq, seq); > + > + tcp_rcv_space_adjust(sk); > + > + /* Clean up data we have read: This will do ACK frames. */ > + if (copied > 0) > + tcp_cleanup_rbuf(sk, copied); > + > + return copied; > +} > +EXPORT_SYMBOL(tcp_read_skb); > + > int tcp_peek_len(struct socket *sock) > { > return tcp_inq(sock->sk); > @@ -1890,7 +1937,7 @@ static int receive_fallback_to_copy(struct sock *sk, > struct sk_buff *skb; > u32 offset; > > - skb = tcp_recv_skb(sk, tcp_sk(sk)->copied_seq, &offset); > + skb = tcp_recv_skb(sk, tcp_sk(sk)->copied_seq, &offset, false); > if (skb) > tcp_zerocopy_set_hint_for_skb(sk, zc, skb, offset); > } > @@ -1937,7 +1984,7 @@ static int tcp_zc_handle_leftover(struct tcp_zerocopy_receive *zc, > if (skb) { > offset = *seq - TCP_SKB_CB(skb)->seq; > } else { > - skb = tcp_recv_skb(sk, *seq, &offset); > + skb = tcp_recv_skb(sk, *seq, &offset, false); > if (TCP_SKB_CB(skb)->has_rxtstamp) { > tcp_update_recv_tstamps(skb, tss); > zc->msg_flags |= TCP_CMSG_TS; > @@ -2130,7 +2177,7 @@ static int tcp_zerocopy_receive(struct sock *sk, > skb = skb->next; > offset = seq - TCP_SKB_CB(skb)->seq; > } else { > - skb = tcp_recv_skb(sk, seq, &offset); > + skb = tcp_recv_skb(sk, seq, &offset, false); > } > > if (TCP_SKB_CB(skb)->has_rxtstamp) { > @@ -2186,7 +2233,7 @@ static int tcp_zerocopy_receive(struct sock *sk, > tcp_rcv_space_adjust(sk); > > /* Clean up data we have read: This will do ACK frames. */ > - tcp_recv_skb(sk, seq, &offset); > + tcp_recv_skb(sk, seq, &offset, false); > tcp_cleanup_rbuf(sk, length + copylen); > ret = 0; > if (length == zc->length) > -- > 2.32.0 > ^ permalink raw reply [flat|nested] 7+ messages in thread
* Re: [Patch bpf-next v2 1/4] tcp: introduce tcp_read_skb() 2022-05-03 0:02 ` Eric Dumazet @ 2022-05-10 18:27 ` Cong Wang 0 siblings, 0 replies; 7+ messages in thread From: Cong Wang @ 2022-05-10 18:27 UTC (permalink / raw) To: Eric Dumazet Cc: netdev, bpf, Cong Wang, John Fastabend, Daniel Borkmann, Jakub Sitnicki On Mon, May 2, 2022 at 5:02 PM Eric Dumazet <edumazet@google.com> wrote: > > On Mon, May 2, 2022 at 11:24 AM Cong Wang <xiyou.wangcong@gmail.com> wrote: > > > > From: Cong Wang <cong.wang@bytedance.com> > > > > This patch inroduces tcp_read_skb() based on tcp_read_sock(), > > a preparation for the next patch which actually introduces > > a new sock ops. > > > > TCP is special here, because it has tcp_read_sock() which is > > mainly used by splice(). tcp_read_sock() supports partial read > > and arbitrary offset, neither of them is needed for sockmap. > > > > Cc: Eric Dumazet <edumazet@google.com> > > Cc: John Fastabend <john.fastabend@gmail.com> > > Cc: Daniel Borkmann <daniel@iogearbox.net> > > Cc: Jakub Sitnicki <jakub@cloudflare.com> > > Signed-off-by: Cong Wang <cong.wang@bytedance.com> > > --- > > include/net/tcp.h | 2 ++ > > net/ipv4/tcp.c | 63 +++++++++++++++++++++++++++++++++++++++++------ > > 2 files changed, 57 insertions(+), 8 deletions(-) > > > > diff --git a/include/net/tcp.h b/include/net/tcp.h > > index 94a52ad1101c..ab7516e5cc56 100644 > > --- a/include/net/tcp.h > > +++ b/include/net/tcp.h > > @@ -667,6 +667,8 @@ void tcp_get_info(struct sock *, struct tcp_info *); > > /* Read 'sendfile()'-style from a TCP socket */ > > int tcp_read_sock(struct sock *sk, read_descriptor_t *desc, > > sk_read_actor_t recv_actor); > > +int tcp_read_skb(struct sock *sk, read_descriptor_t *desc, > > + sk_read_actor_t recv_actor); > > > > void tcp_initialize_rcv_mss(struct sock *sk); > > > > diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c > > index db55af9eb37b..8d48126e3694 100644 > > --- a/net/ipv4/tcp.c > > +++ b/net/ipv4/tcp.c > > @@ -1600,7 +1600,7 @@ static void tcp_eat_recv_skb(struct sock *sk, struct sk_buff *skb) > > __kfree_skb(skb); > > } > > > > -static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off) > > +static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off, bool unlink) > > { > > struct sk_buff *skb; > > u32 offset; > > @@ -1613,6 +1613,8 @@ static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off) > > } > > if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) { > > *off = offset; > > + if (unlink) > > + __skb_unlink(skb, &sk->sk_receive_queue); > > Why adding this @unlink parameter ? > This makes your patch more invasive than needed. > Can not this unlink happen from your new helper instead ? See [3] later. Good point, I was trying to reuse the code there, but it is just one __skb_unlink(). > > > return skb; > > } > > /* This looks weird, but this can happen if TCP collapsing > > @@ -1646,7 +1648,7 @@ int tcp_read_sock(struct sock *sk, read_descriptor_t *desc, > > > > if (sk->sk_state == TCP_LISTEN) > > return -ENOTCONN; > > - while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) { > > + while ((skb = tcp_recv_skb(sk, seq, &offset, false)) != NULL) { > > if (offset < skb->len) { > > int used; > > size_t len; > > @@ -1677,7 +1679,7 @@ int tcp_read_sock(struct sock *sk, read_descriptor_t *desc, > > * getting here: tcp_collapse might have deleted it > > * while aggregating skbs from the socket queue. > > */ > > - skb = tcp_recv_skb(sk, seq - 1, &offset); > > + skb = tcp_recv_skb(sk, seq - 1, &offset, false); > > if (!skb) > > break; > > /* TCP coalescing might have appended data to the skb. > > @@ -1702,13 +1704,58 @@ int tcp_read_sock(struct sock *sk, read_descriptor_t *desc, > > > > /* Clean up data we have read: This will do ACK frames. */ > > if (copied > 0) { > > - tcp_recv_skb(sk, seq, &offset); > > + tcp_recv_skb(sk, seq, &offset, false); > > tcp_cleanup_rbuf(sk, copied); > > } > > return copied; > > } > > EXPORT_SYMBOL(tcp_read_sock); > > > > +int tcp_read_skb(struct sock *sk, read_descriptor_t *desc, > > + sk_read_actor_t recv_actor) > > +{ > > + struct tcp_sock *tp = tcp_sk(sk); > > + u32 seq = tp->copied_seq; > > + struct sk_buff *skb; > > + int copied = 0; > > + u32 offset; > > + > > + if (sk->sk_state == TCP_LISTEN) > > + return -ENOTCONN; > > + > > + while ((skb = tcp_recv_skb(sk, seq, &offset, true)) != NULL) { > > [3] > The unlink from sk->sk_receive_queue could happen here. Right. > > > + int used = recv_actor(desc, skb, 0, skb->len); > > + > > + if (used <= 0) { > > + if (!copied) > > + copied = used; > > + break; > > + } > > + seq += used; > > + copied += used; > > + > > + if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) { > > + kfree_skb(skb); > > [1] > > The two kfree_skb() ([1] & [2]) should be a consume_skb() ? Hm, it is tricky here, we use the skb refcount after this patchset, so it could be a real drop from another kfree_skb() in net/core/skmsg.c which initiates the drop. Thanks. ^ permalink raw reply [flat|nested] 7+ messages in thread
* [Patch bpf-next v2 2/4] net: introduce a new proto_ops ->read_skb() 2022-05-02 18:23 [Patch bpf-next v2 0/4] sockmap: some performance optimizations Cong Wang 2022-05-02 18:23 ` [Patch bpf-next v2 1/4] tcp: introduce tcp_read_skb() Cong Wang @ 2022-05-02 18:23 ` Cong Wang 2022-05-02 18:23 ` [Patch bpf-next v2 3/4] skmsg: get rid of skb_clone() Cong Wang 2022-05-02 18:23 ` [Patch bpf-next v2 4/4] skmsg: get rid of unncessary memset() Cong Wang 3 siblings, 0 replies; 7+ messages in thread From: Cong Wang @ 2022-05-02 18:23 UTC (permalink / raw) To: netdev Cc: bpf, Cong Wang, Eric Dumazet, John Fastabend, Daniel Borkmann, Jakub Sitnicki From: Cong Wang <cong.wang@bytedance.com> Currently both splice() and sockmap use ->read_sock() to read skb from receive queue, but for sockmap we only read one entire skb at a time, so ->read_sock() is too conservative to use. Introduce a new proto_ops ->read_skb() which supports this sematic, with this we can finally pass the ownership of skb to recv actors. For non-TCP protocols, all ->read_sock() can be simply converted to ->read_skb(). Cc: Eric Dumazet <edumazet@google.com> Cc: John Fastabend <john.fastabend@gmail.com> Cc: Daniel Borkmann <daniel@iogearbox.net> Cc: Jakub Sitnicki <jakub@cloudflare.com> Signed-off-by: Cong Wang <cong.wang@bytedance.com> --- include/linux/net.h | 4 ++++ include/net/tcp.h | 3 +-- include/net/udp.h | 3 +-- net/core/skmsg.c | 20 +++++--------------- net/ipv4/af_inet.c | 3 ++- net/ipv4/tcp.c | 9 +++------ net/ipv4/udp.c | 10 ++++------ net/ipv6/af_inet6.c | 3 ++- net/unix/af_unix.c | 23 +++++++++-------------- 9 files changed, 31 insertions(+), 47 deletions(-) diff --git a/include/linux/net.h b/include/linux/net.h index 12093f4db50c..a03485e8cbb2 100644 --- a/include/linux/net.h +++ b/include/linux/net.h @@ -152,6 +152,8 @@ struct module; struct sk_buff; typedef int (*sk_read_actor_t)(read_descriptor_t *, struct sk_buff *, unsigned int, size_t); +typedef int (*skb_read_actor_t)(struct sock *, struct sk_buff *); + struct proto_ops { int family; @@ -214,6 +216,8 @@ struct proto_ops { */ int (*read_sock)(struct sock *sk, read_descriptor_t *desc, sk_read_actor_t recv_actor); + /* This is different from read_sock(), it reads an entire skb at a time. */ + int (*read_skb)(struct sock *sk, skb_read_actor_t recv_actor); int (*sendpage_locked)(struct sock *sk, struct page *page, int offset, size_t size, int flags); int (*sendmsg_locked)(struct sock *sk, struct msghdr *msg, diff --git a/include/net/tcp.h b/include/net/tcp.h index ab7516e5cc56..9f4fe3b80e30 100644 --- a/include/net/tcp.h +++ b/include/net/tcp.h @@ -667,8 +667,7 @@ void tcp_get_info(struct sock *, struct tcp_info *); /* Read 'sendfile()'-style from a TCP socket */ int tcp_read_sock(struct sock *sk, read_descriptor_t *desc, sk_read_actor_t recv_actor); -int tcp_read_skb(struct sock *sk, read_descriptor_t *desc, - sk_read_actor_t recv_actor); +int tcp_read_skb(struct sock *sk, skb_read_actor_t recv_actor); void tcp_initialize_rcv_mss(struct sock *sk); diff --git a/include/net/udp.h b/include/net/udp.h index b83a00330566..47a0e3359771 100644 --- a/include/net/udp.h +++ b/include/net/udp.h @@ -305,8 +305,7 @@ struct sock *__udp6_lib_lookup(struct net *net, struct sk_buff *skb); struct sock *udp6_lib_lookup_skb(const struct sk_buff *skb, __be16 sport, __be16 dport); -int udp_read_sock(struct sock *sk, read_descriptor_t *desc, - sk_read_actor_t recv_actor); +int udp_read_skb(struct sock *sk, skb_read_actor_t recv_actor); /* UDP uses skb->dev_scratch to cache as much information as possible and avoid * possibly multiple cache miss on dequeue() diff --git a/net/core/skmsg.c b/net/core/skmsg.c index 22b983ade0e7..50405e3eda88 100644 --- a/net/core/skmsg.c +++ b/net/core/skmsg.c @@ -1159,21 +1159,17 @@ static void sk_psock_done_strp(struct sk_psock *psock) } #endif /* CONFIG_BPF_STREAM_PARSER */ -static int sk_psock_verdict_recv(read_descriptor_t *desc, struct sk_buff *skb, - unsigned int offset, size_t orig_len) +static int sk_psock_verdict_recv(struct sock *sk, struct sk_buff *skb) { - struct sock *sk = (struct sock *)desc->arg.data; struct sk_psock *psock; struct bpf_prog *prog; int ret = __SK_DROP; - int len = orig_len; + int len = skb->len; /* clone here so sk_eat_skb() in tcp_read_sock does not drop our data */ skb = skb_clone(skb, GFP_ATOMIC); - if (!skb) { - desc->error = -ENOMEM; + if (!skb) return 0; - } rcu_read_lock(); psock = sk_psock(sk); @@ -1203,16 +1199,10 @@ static int sk_psock_verdict_recv(read_descriptor_t *desc, struct sk_buff *skb, static void sk_psock_verdict_data_ready(struct sock *sk) { struct socket *sock = sk->sk_socket; - read_descriptor_t desc; - if (unlikely(!sock || !sock->ops || !sock->ops->read_sock)) + if (unlikely(!sock || !sock->ops || !sock->ops->read_skb)) return; - - desc.arg.data = sk; - desc.error = 0; - desc.count = 1; - - sock->ops->read_sock(sk, &desc, sk_psock_verdict_recv); + sock->ops->read_skb(sk, sk_psock_verdict_recv); } void sk_psock_start_verdict(struct sock *sk, struct sk_psock *psock) diff --git a/net/ipv4/af_inet.c b/net/ipv4/af_inet.c index 93da9f783bec..f615263855d0 100644 --- a/net/ipv4/af_inet.c +++ b/net/ipv4/af_inet.c @@ -1040,6 +1040,7 @@ const struct proto_ops inet_stream_ops = { .sendpage = inet_sendpage, .splice_read = tcp_splice_read, .read_sock = tcp_read_sock, + .read_skb = tcp_read_skb, .sendmsg_locked = tcp_sendmsg_locked, .sendpage_locked = tcp_sendpage_locked, .peek_len = tcp_peek_len, @@ -1067,7 +1068,7 @@ const struct proto_ops inet_dgram_ops = { .setsockopt = sock_common_setsockopt, .getsockopt = sock_common_getsockopt, .sendmsg = inet_sendmsg, - .read_sock = udp_read_sock, + .read_skb = udp_read_skb, .recvmsg = inet_recvmsg, .mmap = sock_no_mmap, .sendpage = inet_sendpage, diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c index 8d48126e3694..d62490d10fd8 100644 --- a/net/ipv4/tcp.c +++ b/net/ipv4/tcp.c @@ -1711,8 +1711,7 @@ int tcp_read_sock(struct sock *sk, read_descriptor_t *desc, } EXPORT_SYMBOL(tcp_read_sock); -int tcp_read_skb(struct sock *sk, read_descriptor_t *desc, - sk_read_actor_t recv_actor) +int tcp_read_skb(struct sock *sk, skb_read_actor_t recv_actor) { struct tcp_sock *tp = tcp_sk(sk); u32 seq = tp->copied_seq; @@ -1724,7 +1723,7 @@ int tcp_read_skb(struct sock *sk, read_descriptor_t *desc, return -ENOTCONN; while ((skb = tcp_recv_skb(sk, seq, &offset, true)) != NULL) { - int used = recv_actor(desc, skb, 0, skb->len); + int used = recv_actor(sk, skb); if (used <= 0) { if (!copied) @@ -1740,9 +1739,7 @@ int tcp_read_skb(struct sock *sk, read_descriptor_t *desc, break; } kfree_skb(skb); - if (!desc->count) - break; - WRITE_ONCE(tp->copied_seq, seq); + break; } WRITE_ONCE(tp->copied_seq, seq); diff --git a/net/ipv4/udp.c b/net/ipv4/udp.c index aa8545ca6964..b8cfa0c3de59 100644 --- a/net/ipv4/udp.c +++ b/net/ipv4/udp.c @@ -1795,8 +1795,7 @@ struct sk_buff *__skb_recv_udp(struct sock *sk, unsigned int flags, } EXPORT_SYMBOL(__skb_recv_udp); -int udp_read_sock(struct sock *sk, read_descriptor_t *desc, - sk_read_actor_t recv_actor) +int udp_read_skb(struct sock *sk, skb_read_actor_t recv_actor) { int copied = 0; @@ -1818,7 +1817,7 @@ int udp_read_sock(struct sock *sk, read_descriptor_t *desc, continue; } - used = recv_actor(desc, skb, 0, skb->len); + used = recv_actor(sk, skb); if (used <= 0) { if (!copied) copied = used; @@ -1829,13 +1828,12 @@ int udp_read_sock(struct sock *sk, read_descriptor_t *desc, } kfree_skb(skb); - if (!desc->count) - break; + break; } return copied; } -EXPORT_SYMBOL(udp_read_sock); +EXPORT_SYMBOL(udp_read_skb); /* * This should be easy, if there is something there we diff --git a/net/ipv6/af_inet6.c b/net/ipv6/af_inet6.c index 70564ddccc46..1aea5ef9bdea 100644 --- a/net/ipv6/af_inet6.c +++ b/net/ipv6/af_inet6.c @@ -701,6 +701,7 @@ const struct proto_ops inet6_stream_ops = { .sendpage_locked = tcp_sendpage_locked, .splice_read = tcp_splice_read, .read_sock = tcp_read_sock, + .read_skb = tcp_read_skb, .peek_len = tcp_peek_len, #ifdef CONFIG_COMPAT .compat_ioctl = inet6_compat_ioctl, @@ -726,7 +727,7 @@ const struct proto_ops inet6_dgram_ops = { .getsockopt = sock_common_getsockopt, /* ok */ .sendmsg = inet6_sendmsg, /* retpoline's sake */ .recvmsg = inet6_recvmsg, /* retpoline's sake */ - .read_sock = udp_read_sock, + .read_skb = udp_read_skb, .mmap = sock_no_mmap, .sendpage = sock_no_sendpage, .set_peek_off = sk_set_peek_off, diff --git a/net/unix/af_unix.c b/net/unix/af_unix.c index e1dd9e9c8452..71deefaaf373 100644 --- a/net/unix/af_unix.c +++ b/net/unix/af_unix.c @@ -741,10 +741,8 @@ static ssize_t unix_stream_splice_read(struct socket *, loff_t *ppos, unsigned int flags); static int unix_dgram_sendmsg(struct socket *, struct msghdr *, size_t); static int unix_dgram_recvmsg(struct socket *, struct msghdr *, size_t, int); -static int unix_read_sock(struct sock *sk, read_descriptor_t *desc, - sk_read_actor_t recv_actor); -static int unix_stream_read_sock(struct sock *sk, read_descriptor_t *desc, - sk_read_actor_t recv_actor); +static int unix_read_skb(struct sock *sk, skb_read_actor_t recv_actor); +static int unix_stream_read_skb(struct sock *sk, skb_read_actor_t recv_actor); static int unix_dgram_connect(struct socket *, struct sockaddr *, int, int); static int unix_seqpacket_sendmsg(struct socket *, struct msghdr *, size_t); @@ -798,7 +796,7 @@ static const struct proto_ops unix_stream_ops = { .shutdown = unix_shutdown, .sendmsg = unix_stream_sendmsg, .recvmsg = unix_stream_recvmsg, - .read_sock = unix_stream_read_sock, + .read_skb = unix_stream_read_skb, .mmap = sock_no_mmap, .sendpage = unix_stream_sendpage, .splice_read = unix_stream_splice_read, @@ -823,7 +821,7 @@ static const struct proto_ops unix_dgram_ops = { .listen = sock_no_listen, .shutdown = unix_shutdown, .sendmsg = unix_dgram_sendmsg, - .read_sock = unix_read_sock, + .read_skb = unix_read_skb, .recvmsg = unix_dgram_recvmsg, .mmap = sock_no_mmap, .sendpage = sock_no_sendpage, @@ -2489,8 +2487,7 @@ static int unix_dgram_recvmsg(struct socket *sock, struct msghdr *msg, size_t si return __unix_dgram_recvmsg(sk, msg, size, flags); } -static int unix_read_sock(struct sock *sk, read_descriptor_t *desc, - sk_read_actor_t recv_actor) +static int unix_read_skb(struct sock *sk, skb_read_actor_t recv_actor) { int copied = 0; @@ -2505,7 +2502,7 @@ static int unix_read_sock(struct sock *sk, read_descriptor_t *desc, if (!skb) return err; - used = recv_actor(desc, skb, 0, skb->len); + used = recv_actor(sk, skb); if (used <= 0) { if (!copied) copied = used; @@ -2516,8 +2513,7 @@ static int unix_read_sock(struct sock *sk, read_descriptor_t *desc, } kfree_skb(skb); - if (!desc->count) - break; + break; } return copied; @@ -2652,13 +2648,12 @@ static struct sk_buff *manage_oob(struct sk_buff *skb, struct sock *sk, } #endif -static int unix_stream_read_sock(struct sock *sk, read_descriptor_t *desc, - sk_read_actor_t recv_actor) +static int unix_stream_read_skb(struct sock *sk, skb_read_actor_t recv_actor) { if (unlikely(sk->sk_state != TCP_ESTABLISHED)) return -ENOTCONN; - return unix_read_sock(sk, desc, recv_actor); + return unix_read_skb(sk, recv_actor); } static int unix_stream_read_generic(struct unix_stream_read_state *state, -- 2.32.0 ^ permalink raw reply related [flat|nested] 7+ messages in thread
* [Patch bpf-next v2 3/4] skmsg: get rid of skb_clone() 2022-05-02 18:23 [Patch bpf-next v2 0/4] sockmap: some performance optimizations Cong Wang 2022-05-02 18:23 ` [Patch bpf-next v2 1/4] tcp: introduce tcp_read_skb() Cong Wang 2022-05-02 18:23 ` [Patch bpf-next v2 2/4] net: introduce a new proto_ops ->read_skb() Cong Wang @ 2022-05-02 18:23 ` Cong Wang 2022-05-02 18:23 ` [Patch bpf-next v2 4/4] skmsg: get rid of unncessary memset() Cong Wang 3 siblings, 0 replies; 7+ messages in thread From: Cong Wang @ 2022-05-02 18:23 UTC (permalink / raw) To: netdev Cc: bpf, Cong Wang, Eric Dumazet, John Fastabend, Daniel Borkmann, Jakub Sitnicki From: Cong Wang <cong.wang@bytedance.com> With ->read_skb() now we have an entire skb dequeued from receive queue, now we just need to grab an addtional refcnt before passing its ownership to recv actors. And we should not touch them any more, particularly for skb->sk. Fortunately, skb->sk is already set for most of the protocols except UDP where skb->sk has been stolen, so we have to fix it up for UDP case. Cc: Eric Dumazet <edumazet@google.com> Cc: John Fastabend <john.fastabend@gmail.com> Cc: Daniel Borkmann <daniel@iogearbox.net> Cc: Jakub Sitnicki <jakub@cloudflare.com> Signed-off-by: Cong Wang <cong.wang@bytedance.com> --- net/core/skmsg.c | 7 +------ net/ipv4/udp.c | 1 + 2 files changed, 2 insertions(+), 6 deletions(-) diff --git a/net/core/skmsg.c b/net/core/skmsg.c index 50405e3eda88..3ff86d73672c 100644 --- a/net/core/skmsg.c +++ b/net/core/skmsg.c @@ -1166,10 +1166,7 @@ static int sk_psock_verdict_recv(struct sock *sk, struct sk_buff *skb) int ret = __SK_DROP; int len = skb->len; - /* clone here so sk_eat_skb() in tcp_read_sock does not drop our data */ - skb = skb_clone(skb, GFP_ATOMIC); - if (!skb) - return 0; + skb_get(skb); rcu_read_lock(); psock = sk_psock(sk); @@ -1182,12 +1179,10 @@ static int sk_psock_verdict_recv(struct sock *sk, struct sk_buff *skb) if (!prog) prog = READ_ONCE(psock->progs.skb_verdict); if (likely(prog)) { - skb->sk = sk; skb_dst_drop(skb); skb_bpf_redirect_clear(skb); ret = bpf_prog_run_pin_on_cpu(prog, skb); ret = sk_psock_map_verd(ret, skb_bpf_redirect_fetch(skb)); - skb->sk = NULL; } if (sk_psock_verdict_apply(psock, skb, ret) < 0) len = 0; diff --git a/net/ipv4/udp.c b/net/ipv4/udp.c index b8cfa0c3de59..71c2c147f2d0 100644 --- a/net/ipv4/udp.c +++ b/net/ipv4/udp.c @@ -1817,6 +1817,7 @@ int udp_read_skb(struct sock *sk, skb_read_actor_t recv_actor) continue; } + WARN_ON(!skb_set_owner_sk_safe(skb, sk)); used = recv_actor(sk, skb); if (used <= 0) { if (!copied) -- 2.32.0 ^ permalink raw reply related [flat|nested] 7+ messages in thread
* [Patch bpf-next v2 4/4] skmsg: get rid of unncessary memset() 2022-05-02 18:23 [Patch bpf-next v2 0/4] sockmap: some performance optimizations Cong Wang ` (2 preceding siblings ...) 2022-05-02 18:23 ` [Patch bpf-next v2 3/4] skmsg: get rid of skb_clone() Cong Wang @ 2022-05-02 18:23 ` Cong Wang 3 siblings, 0 replies; 7+ messages in thread From: Cong Wang @ 2022-05-02 18:23 UTC (permalink / raw) To: netdev; +Cc: bpf, Cong Wang, John Fastabend, Daniel Borkmann, Jakub Sitnicki From: Cong Wang <cong.wang@bytedance.com> We always allocate skmsg with kzalloc(), so there is no need to call memset(0) on it, the only thing we need from sk_msg_init() is sg_init_marker(). So introduce a new helper which is just kzalloc()+sg_init_marker(), this saves an unncessary memset(0) for skmsg on fast path. Cc: John Fastabend <john.fastabend@gmail.com> Cc: Daniel Borkmann <daniel@iogearbox.net> Cc: Jakub Sitnicki <jakub@cloudflare.com> Signed-off-by: Cong Wang <cong.wang@bytedance.com> --- net/core/skmsg.c | 23 +++++++++++++---------- 1 file changed, 13 insertions(+), 10 deletions(-) diff --git a/net/core/skmsg.c b/net/core/skmsg.c index 3ff86d73672c..6dbb735ec94d 100644 --- a/net/core/skmsg.c +++ b/net/core/skmsg.c @@ -497,23 +497,27 @@ bool sk_msg_is_readable(struct sock *sk) } EXPORT_SYMBOL_GPL(sk_msg_is_readable); -static struct sk_msg *sk_psock_create_ingress_msg(struct sock *sk, - struct sk_buff *skb) +static struct sk_msg *alloc_sk_msg(void) { struct sk_msg *msg; - if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) + msg = kzalloc(sizeof(*msg), __GFP_NOWARN | GFP_KERNEL); + if (unlikely(!msg)) return NULL; + sg_init_marker(msg->sg.data, NR_MSG_FRAG_IDS); + return msg; +} - if (!sk_rmem_schedule(sk, skb, skb->truesize)) +static struct sk_msg *sk_psock_create_ingress_msg(struct sock *sk, + struct sk_buff *skb) +{ + if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) return NULL; - msg = kzalloc(sizeof(*msg), __GFP_NOWARN | GFP_KERNEL); - if (unlikely(!msg)) + if (!sk_rmem_schedule(sk, skb, skb->truesize)) return NULL; - sk_msg_init(msg); - return msg; + return alloc_sk_msg(); } static int sk_psock_skb_ingress_enqueue(struct sk_buff *skb, @@ -590,13 +594,12 @@ static int sk_psock_skb_ingress(struct sk_psock *psock, struct sk_buff *skb, static int sk_psock_skb_ingress_self(struct sk_psock *psock, struct sk_buff *skb, u32 off, u32 len) { - struct sk_msg *msg = kzalloc(sizeof(*msg), __GFP_NOWARN | GFP_ATOMIC); + struct sk_msg *msg = alloc_sk_msg(); struct sock *sk = psock->sk; int err; if (unlikely(!msg)) return -EAGAIN; - sk_msg_init(msg); skb_set_owner_r(skb, sk); err = sk_psock_skb_ingress_enqueue(skb, off, len, psock, sk, msg); if (err < 0) -- 2.32.0 ^ permalink raw reply related [flat|nested] 7+ messages in thread
end of thread, other threads:[~2022-05-10 18:27 UTC | newest] Thread overview: 7+ messages (download: mbox.gz / follow: Atom feed) -- links below jump to the message on this page -- 2022-05-02 18:23 [Patch bpf-next v2 0/4] sockmap: some performance optimizations Cong Wang 2022-05-02 18:23 ` [Patch bpf-next v2 1/4] tcp: introduce tcp_read_skb() Cong Wang 2022-05-03 0:02 ` Eric Dumazet 2022-05-10 18:27 ` Cong Wang 2022-05-02 18:23 ` [Patch bpf-next v2 2/4] net: introduce a new proto_ops ->read_skb() Cong Wang 2022-05-02 18:23 ` [Patch bpf-next v2 3/4] skmsg: get rid of skb_clone() Cong Wang 2022-05-02 18:23 ` [Patch bpf-next v2 4/4] skmsg: get rid of unncessary memset() Cong Wang
This is a public inbox, see mirroring instructions for how to clone and mirror all data and code used for this inbox; as well as URLs for NNTP newsgroup(s).