| Message ID | d0b676c036479fe304218a05e1b03a7ba274426c.1460495323.git.marcelo.leitner@gmail.com |
|---|---|
| State | Accepted, archived |
| Delegated to: | David Miller |
| Headers | show |
On Tue, Apr 12, 2016 at 2:11 PM, Marcelo Ricardo Leitner <marcelo.leitner@gmail.com> wrote: > This patch adds what's missing to properly support RPS and RFS on SCTP, > as some of it is already implemented in common calls. > > Having support for RPS and RFS allows better scaling specially because > not all NICs support hashing SCTP headers. > > Save the hash right when we dequeue a skb from inqueue so we do it only > once per skb instead of per chunk. New sockets will then inherit the > hash through sctp_copy_sock(). > > Signed-off-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com> > --- > net/sctp/inqueue.c | 3 +++ > net/sctp/socket.c | 3 +++ > 2 files changed, 6 insertions(+) > > diff --git a/net/sctp/inqueue.c b/net/sctp/inqueue.c > index 7e8a16c77039e1b70ef89f3e862dbb332bcc614f..b335ffcef0b901b0e71bf0843057dbf5877da31b 100644 > --- a/net/sctp/inqueue.c > +++ b/net/sctp/inqueue.c > @@ -163,6 +163,9 @@ struct sctp_chunk *sctp_inq_pop(struct sctp_inq *queue) > chunk->singleton = 1; > ch = (sctp_chunkhdr_t *) chunk->skb->data; > chunk->data_accepted = 0; > + > + if (chunk->asoc) > + sock_rps_save_rxhash(chunk->asoc->base.sk, chunk->skb); > } > > chunk->chunk_hdr = ch; > diff --git a/net/sctp/socket.c b/net/sctp/socket.c > index 878d28eda1a68dc639d7b9f3f663d7518f21bb32..36697f85ce48be39f41ddedd9f53369c7f9e28d8 100644 > --- a/net/sctp/socket.c > +++ b/net/sctp/socket.c > @@ -6430,6 +6430,8 @@ unsigned int sctp_poll(struct file *file, struct socket *sock, poll_table *wait) > > poll_wait(file, sk_sleep(sk), wait); > > + sock_rps_record_flow(sk); > + > /* A TCP-style listening socket becomes readable when the accept queue > * is not empty. > */ > @@ -7186,6 +7188,7 @@ void sctp_copy_sock(struct sock *newsk, struct sock *sk, > newsk->sk_lingertime = sk->sk_lingertime; > newsk->sk_rcvtimeo = sk->sk_rcvtimeo; > newsk->sk_sndtimeo = sk->sk_sndtimeo; > + newsk->sk_rxhash = sk->sk_rxhash; > > newinet = inet_sk(newsk); > HI Marcelo, sock_rps_record_flow should probably be in sctp_poll also (like it is in tcp_poll). Thanks, Tom > -- > 2.5.0 >
On Tue, Apr 12, 2016 at 02:50:45PM -0700, Tom Herbert wrote: > On Tue, Apr 12, 2016 at 2:11 PM, Marcelo Ricardo Leitner > <marcelo.leitner@gmail.com> wrote: > > This patch adds what's missing to properly support RPS and RFS on SCTP, > > as some of it is already implemented in common calls. > > > > Having support for RPS and RFS allows better scaling specially because > > not all NICs support hashing SCTP headers. > > > > Save the hash right when we dequeue a skb from inqueue so we do it only > > once per skb instead of per chunk. New sockets will then inherit the > > hash through sctp_copy_sock(). > > > > Signed-off-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com> > > --- > > net/sctp/inqueue.c | 3 +++ > > net/sctp/socket.c | 3 +++ > > 2 files changed, 6 insertions(+) > > > > diff --git a/net/sctp/inqueue.c b/net/sctp/inqueue.c > > index 7e8a16c77039e1b70ef89f3e862dbb332bcc614f..b335ffcef0b901b0e71bf0843057dbf5877da31b 100644 > > --- a/net/sctp/inqueue.c > > +++ b/net/sctp/inqueue.c > > @@ -163,6 +163,9 @@ struct sctp_chunk *sctp_inq_pop(struct sctp_inq *queue) > > chunk->singleton = 1; > > ch = (sctp_chunkhdr_t *) chunk->skb->data; > > chunk->data_accepted = 0; > > + > > + if (chunk->asoc) > > + sock_rps_save_rxhash(chunk->asoc->base.sk, chunk->skb); > > } > > > > chunk->chunk_hdr = ch; > > diff --git a/net/sctp/socket.c b/net/sctp/socket.c > > index 878d28eda1a68dc639d7b9f3f663d7518f21bb32..36697f85ce48be39f41ddedd9f53369c7f9e28d8 100644 > > --- a/net/sctp/socket.c > > +++ b/net/sctp/socket.c > > @@ -6430,6 +6430,8 @@ unsigned int sctp_poll(struct file *file, struct socket *sock, poll_table *wait) > > > > poll_wait(file, sk_sleep(sk), wait); > > > > + sock_rps_record_flow(sk); > > + > > /* A TCP-style listening socket becomes readable when the accept queue > > * is not empty. > > */ > > @@ -7186,6 +7188,7 @@ void sctp_copy_sock(struct sock *newsk, struct sock *sk, > > newsk->sk_lingertime = sk->sk_lingertime; > > newsk->sk_rcvtimeo = sk->sk_rcvtimeo; > > newsk->sk_sndtimeo = sk->sk_sndtimeo; > > + newsk->sk_rxhash = sk->sk_rxhash; > > > > newinet = inet_sk(newsk); > > > HI Marcelo, > > sock_rps_record_flow should probably be in sctp_poll also (like it is > in tcp_poll). > > Thanks, > Tom Hi Tom, Yes, that's the middle chunk on this patch, no? Note that for udp it's done after the actual poll, while for tcp it's being saved before it. I went more udp-style on this one. Thanks, Marcelo
On Tue, Apr 12, 2016 at 2:57 PM, Marcelo Ricardo Leitner <marcelo.leitner@gmail.com> wrote: > On Tue, Apr 12, 2016 at 02:50:45PM -0700, Tom Herbert wrote: >> On Tue, Apr 12, 2016 at 2:11 PM, Marcelo Ricardo Leitner >> <marcelo.leitner@gmail.com> wrote: >> > This patch adds what's missing to properly support RPS and RFS on SCTP, >> > as some of it is already implemented in common calls. >> > >> > Having support for RPS and RFS allows better scaling specially because >> > not all NICs support hashing SCTP headers. >> > >> > Save the hash right when we dequeue a skb from inqueue so we do it only >> > once per skb instead of per chunk. New sockets will then inherit the >> > hash through sctp_copy_sock(). >> > >> > Signed-off-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com> >> > --- >> > net/sctp/inqueue.c | 3 +++ >> > net/sctp/socket.c | 3 +++ >> > 2 files changed, 6 insertions(+) >> > >> > diff --git a/net/sctp/inqueue.c b/net/sctp/inqueue.c >> > index 7e8a16c77039e1b70ef89f3e862dbb332bcc614f..b335ffcef0b901b0e71bf0843057dbf5877da31b 100644 >> > --- a/net/sctp/inqueue.c >> > +++ b/net/sctp/inqueue.c >> > @@ -163,6 +163,9 @@ struct sctp_chunk *sctp_inq_pop(struct sctp_inq *queue) >> > chunk->singleton = 1; >> > ch = (sctp_chunkhdr_t *) chunk->skb->data; >> > chunk->data_accepted = 0; >> > + >> > + if (chunk->asoc) >> > + sock_rps_save_rxhash(chunk->asoc->base.sk, chunk->skb); >> > } >> > >> > chunk->chunk_hdr = ch; >> > diff --git a/net/sctp/socket.c b/net/sctp/socket.c >> > index 878d28eda1a68dc639d7b9f3f663d7518f21bb32..36697f85ce48be39f41ddedd9f53369c7f9e28d8 100644 >> > --- a/net/sctp/socket.c >> > +++ b/net/sctp/socket.c >> > @@ -6430,6 +6430,8 @@ unsigned int sctp_poll(struct file *file, struct socket *sock, poll_table *wait) >> > >> > poll_wait(file, sk_sleep(sk), wait); >> > >> > + sock_rps_record_flow(sk); >> > + >> > /* A TCP-style listening socket becomes readable when the accept queue >> > * is not empty. >> > */ >> > @@ -7186,6 +7188,7 @@ void sctp_copy_sock(struct sock *newsk, struct sock *sk, >> > newsk->sk_lingertime = sk->sk_lingertime; >> > newsk->sk_rcvtimeo = sk->sk_rcvtimeo; >> > newsk->sk_sndtimeo = sk->sk_sndtimeo; >> > + newsk->sk_rxhash = sk->sk_rxhash; >> > >> > newinet = inet_sk(newsk); >> > >> HI Marcelo, >> >> sock_rps_record_flow should probably be in sctp_poll also (like it is >> in tcp_poll). >> >> Thanks, >> Tom > > Hi Tom, > > Yes, that's the middle chunk on this patch, no? > > Note that for udp it's done after the actual poll, while for tcp it's > being saved before it. I went more udp-style on this one. > Yes, I missed that. > Thanks, > Marcelo
diff --git a/net/sctp/inqueue.c b/net/sctp/inqueue.c index 7e8a16c77039e1b70ef89f3e862dbb332bcc614f..b335ffcef0b901b0e71bf0843057dbf5877da31b 100644 --- a/net/sctp/inqueue.c +++ b/net/sctp/inqueue.c @@ -163,6 +163,9 @@ struct sctp_chunk *sctp_inq_pop(struct sctp_inq *queue) chunk->singleton = 1; ch = (sctp_chunkhdr_t *) chunk->skb->data; chunk->data_accepted = 0; + + if (chunk->asoc) + sock_rps_save_rxhash(chunk->asoc->base.sk, chunk->skb); } chunk->chunk_hdr = ch; diff --git a/net/sctp/socket.c b/net/sctp/socket.c index 878d28eda1a68dc639d7b9f3f663d7518f21bb32..36697f85ce48be39f41ddedd9f53369c7f9e28d8 100644 --- a/net/sctp/socket.c +++ b/net/sctp/socket.c @@ -6430,6 +6430,8 @@ unsigned int sctp_poll(struct file *file, struct socket *sock, poll_table *wait) poll_wait(file, sk_sleep(sk), wait); + sock_rps_record_flow(sk); + /* A TCP-style listening socket becomes readable when the accept queue * is not empty. */ @@ -7186,6 +7188,7 @@ void sctp_copy_sock(struct sock *newsk, struct sock *sk, newsk->sk_lingertime = sk->sk_lingertime; newsk->sk_rcvtimeo = sk->sk_rcvtimeo; newsk->sk_sndtimeo = sk->sk_sndtimeo; + newsk->sk_rxhash = sk->sk_rxhash; newinet = inet_sk(newsk);
This patch adds what's missing to properly support RPS and RFS on SCTP, as some of it is already implemented in common calls. Having support for RPS and RFS allows better scaling specially because not all NICs support hashing SCTP headers. Save the hash right when we dequeue a skb from inqueue so we do it only once per skb instead of per chunk. New sockets will then inherit the hash through sctp_copy_sock(). Signed-off-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com> --- net/sctp/inqueue.c | 3 +++ net/sctp/socket.c | 3 +++ 2 files changed, 6 insertions(+)