From e44c1887fdb21e349a778007fcd6d72d9cf45552 Mon Sep 17 00:00:00 2001 From: Steven Hartland Date: Mon, 10 Apr 2017 08:19:35 +0000 Subject: [PATCH] Use estimated RTT for receive buffer auto resizing instead of timestamps Switched from using timestamps to RTT estimates when performing TCP receive buffer auto resizing, as not all hosts support / enable TCP timestamps. Disabled reset of receive buffer auto scaling when not in bulk receive mode, which gives an extra 20% performance increase. Also extracted auto resizing to a common method shared between standard and fastpath modules. With this AWS S3 downloads at ~17ms latency on a 1Gbps connection jump from ~3MB/s to ~100MB/s using the default settings. Reviewed by: lstewart, gnn MFC after: 2 weeks Relnotes: Yes Sponsored by: Multiplay Differential Revision: https://reviews.freebsd.org/D9668 --- sys/netinet/in_kdtrace.c | 8 ++ sys/netinet/in_kdtrace.h | 1 + sys/netinet/tcp_input.c | 123 +++++++++++++++--------------- sys/netinet/tcp_output.c | 10 ++- sys/netinet/tcp_stacks/fastpath.c | 62 +-------------- sys/netinet/tcp_var.h | 2 + 6 files changed, 82 insertions(+), 124 deletions(-) diff --git a/sys/netinet/in_kdtrace.c b/sys/netinet/in_kdtrace.c index df217ee6ab4..79ae2599bc2 100644 --- a/sys/netinet/in_kdtrace.c +++ b/sys/netinet/in_kdtrace.c @@ -132,6 +132,14 @@ SDT_PROBE_DEFINE6_XLATE(tcp, , , state__change, "void *", "void *", "int", "tcplsinfo_t *"); +SDT_PROBE_DEFINE6_XLATE(tcp, , , receive__autoresize, + "void *", "void *", + "struct tcpcb *", "csinfo_t *", + "struct mbuf *", "ipinfo_t *", + "struct tcpcb *", "tcpsinfo_t *" , + "struct tcphdr *", "tcpinfoh_t *", + "int", "int"); + SDT_PROBE_DEFINE5_XLATE(udp, , , receive, "void *", "pktinfo_t *", "struct inpcb *", "csinfo_t *", diff --git a/sys/netinet/in_kdtrace.h b/sys/netinet/in_kdtrace.h index a36991ef14c..0825c7dff35 100644 --- a/sys/netinet/in_kdtrace.h +++ b/sys/netinet/in_kdtrace.h @@ -65,6 +65,7 @@ SDT_PROBE_DECLARE(tcp, , , debug__input); SDT_PROBE_DECLARE(tcp, , , debug__output); SDT_PROBE_DECLARE(tcp, , , debug__user); SDT_PROBE_DECLARE(tcp, , , debug__drop); +SDT_PROBE_DECLARE(tcp, , , receive__autoresize); SDT_PROBE_DECLARE(udp, , , receive); SDT_PROBE_DECLARE(udp, , , send); diff --git a/sys/netinet/tcp_input.c b/sys/netinet/tcp_input.c index 85410b67948..367e45ed1ac 100644 --- a/sys/netinet/tcp_input.c +++ b/sys/netinet/tcp_input.c @@ -1486,6 +1486,68 @@ drop: return (IPPROTO_DONE); } +/* + * Automatic sizing of receive socket buffer. Often the send + * buffer size is not optimally adjusted to the actual network + * conditions at hand (delay bandwidth product). Setting the + * buffer size too small limits throughput on links with high + * bandwidth and high delay (eg. trans-continental/oceanic links). + * + * On the receive side the socket buffer memory is only rarely + * used to any significant extent. This allows us to be much + * more aggressive in scaling the receive socket buffer. For + * the case that the buffer space is actually used to a large + * extent and we run out of kernel memory we can simply drop + * the new segments; TCP on the sender will just retransmit it + * later. Setting the buffer size too big may only consume too + * much kernel memory if the application doesn't read() from + * the socket or packet loss or reordering makes use of the + * reassembly queue. + * + * The criteria to step up the receive buffer one notch are: + * 1. Application has not set receive buffer size with + * SO_RCVBUF. Setting SO_RCVBUF clears SB_AUTOSIZE. + * 2. the number of bytes received during the time it takes + * one timestamp to be reflected back to us (the RTT); + * 3. received bytes per RTT is within seven eighth of the + * current socket buffer size; + * 4. receive buffer size has not hit maximal automatic size; + * + * This algorithm does one step per RTT at most and only if + * we receive a bulk stream w/o packet losses or reorderings. + * Shrinking the buffer during idle times is not necessary as + * it doesn't consume any memory when idle. + * + * TODO: Only step up if the application is actually serving + * the buffer to better manage the socket buffer resources. + */ +int +tcp_autorcvbuf(struct mbuf *m, struct tcphdr *th, struct socket *so, + struct tcpcb *tp, int tlen) +{ + int newsize = 0; + + if (V_tcp_do_autorcvbuf && (so->so_rcv.sb_flags & SB_AUTOSIZE) && + tp->t_srtt != 0 && tp->rfbuf_ts != 0 && + TCP_TS_TO_TICKS(tcp_ts_getticks() - tp->rfbuf_ts) > + (tp->t_srtt >> TCP_RTT_SHIFT)) { + if (tp->rfbuf_cnt > (so->so_rcv.sb_hiwat / 8 * 7) && + so->so_rcv.sb_hiwat < V_tcp_autorcvbuf_max) { + newsize = min(so->so_rcv.sb_hiwat + + V_tcp_autorcvbuf_inc, V_tcp_autorcvbuf_max); + } + TCP_PROBE6(receive__autoresize, NULL, tp, m, tp, th, newsize); + + /* Start over with next RTT. */ + tp->rfbuf_ts = 0; + tp->rfbuf_cnt = 0; + } else { + tp->rfbuf_cnt += tlen; /* add up */ + } + + return (newsize); +} + void tcp_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int drop_hdrlen, int tlen, uint8_t iptos, @@ -1849,62 +1911,7 @@ tcp_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so, #endif TCP_PROBE3(debug__input, tp, th, m); - /* - * Automatic sizing of receive socket buffer. Often the send - * buffer size is not optimally adjusted to the actual network - * conditions at hand (delay bandwidth product). Setting the - * buffer size too small limits throughput on links with high - * bandwidth and high delay (eg. trans-continental/oceanic links). - * - * On the receive side the socket buffer memory is only rarely - * used to any significant extent. This allows us to be much - * more aggressive in scaling the receive socket buffer. For - * the case that the buffer space is actually used to a large - * extent and we run out of kernel memory we can simply drop - * the new segments; TCP on the sender will just retransmit it - * later. Setting the buffer size too big may only consume too - * much kernel memory if the application doesn't read() from - * the socket or packet loss or reordering makes use of the - * reassembly queue. - * - * The criteria to step up the receive buffer one notch are: - * 1. Application has not set receive buffer size with - * SO_RCVBUF. Setting SO_RCVBUF clears SB_AUTOSIZE. - * 2. the number of bytes received during the time it takes - * one timestamp to be reflected back to us (the RTT); - * 3. received bytes per RTT is within seven eighth of the - * current socket buffer size; - * 4. receive buffer size has not hit maximal automatic size; - * - * This algorithm does one step per RTT at most and only if - * we receive a bulk stream w/o packet losses or reorderings. - * Shrinking the buffer during idle times is not necessary as - * it doesn't consume any memory when idle. - * - * TODO: Only step up if the application is actually serving - * the buffer to better manage the socket buffer resources. - */ - if (V_tcp_do_autorcvbuf && - (to.to_flags & TOF_TS) && - to.to_tsecr && - (so->so_rcv.sb_flags & SB_AUTOSIZE)) { - if (TSTMP_GT(to.to_tsecr, tp->rfbuf_ts) && - to.to_tsecr - tp->rfbuf_ts < hz) { - if (tp->rfbuf_cnt > - (so->so_rcv.sb_hiwat / 8 * 7) && - so->so_rcv.sb_hiwat < - V_tcp_autorcvbuf_max) { - newsize = - min(so->so_rcv.sb_hiwat + - V_tcp_autorcvbuf_inc, - V_tcp_autorcvbuf_max); - } - /* Start over with next RTT. */ - tp->rfbuf_ts = 0; - tp->rfbuf_cnt = 0; - } else - tp->rfbuf_cnt += tlen; /* add up */ - } + newsize = tcp_autorcvbuf(m, th, so, tp, tlen); /* Add data to socket buffer. */ SOCKBUF_LOCK(&so->so_rcv); @@ -1945,10 +1952,6 @@ tcp_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so, win = 0; tp->rcv_wnd = imax(win, (int)(tp->rcv_adv - tp->rcv_nxt)); - /* Reset receive buffer auto scaling when not in bulk receive mode. */ - tp->rfbuf_ts = 0; - tp->rfbuf_cnt = 0; - switch (tp->t_state) { /* diff --git a/sys/netinet/tcp_output.c b/sys/netinet/tcp_output.c index f5c8c835e80..d997b271ea3 100644 --- a/sys/netinet/tcp_output.c +++ b/sys/netinet/tcp_output.c @@ -831,11 +831,13 @@ send: to.to_tsval = tcp_ts_getticks() + tp->ts_offset; to.to_tsecr = tp->ts_recent; to.to_flags |= TOF_TS; - /* Set receive buffer autosizing timestamp. */ - if (tp->rfbuf_ts == 0 && - (so->so_rcv.sb_flags & SB_AUTOSIZE)) - tp->rfbuf_ts = tcp_ts_getticks(); } + + /* Set receive buffer autosizing timestamp. */ + if (tp->rfbuf_ts == 0 && + (so->so_rcv.sb_flags & SB_AUTOSIZE)) + tp->rfbuf_ts = tcp_ts_getticks(); + /* Selective ACK's. */ if (tp->t_flags & TF_SACK_PERMIT) { if (flags & TH_SYN) diff --git a/sys/netinet/tcp_stacks/fastpath.c b/sys/netinet/tcp_stacks/fastpath.c index 7a3961f8da1..aadd9b17238 100644 --- a/sys/netinet/tcp_stacks/fastpath.c +++ b/sys/netinet/tcp_stacks/fastpath.c @@ -399,62 +399,8 @@ tcp_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so, (void *)tcp_saveipgen, &tcp_savetcp, 0); #endif TCP_PROBE3(debug__input, tp, th, m); - /* - * Automatic sizing of receive socket buffer. Often the send - * buffer size is not optimally adjusted to the actual network - * conditions at hand (delay bandwidth product). Setting the - * buffer size too small limits throughput on links with high - * bandwidth and high delay (eg. trans-continental/oceanic links). - * - * On the receive side the socket buffer memory is only rarely - * used to any significant extent. This allows us to be much - * more aggressive in scaling the receive socket buffer. For - * the case that the buffer space is actually used to a large - * extent and we run out of kernel memory we can simply drop - * the new segments; TCP on the sender will just retransmit it - * later. Setting the buffer size too big may only consume too - * much kernel memory if the application doesn't read() from - * the socket or packet loss or reordering makes use of the - * reassembly queue. - * - * The criteria to step up the receive buffer one notch are: - * 1. Application has not set receive buffer size with - * SO_RCVBUF. Setting SO_RCVBUF clears SB_AUTOSIZE. - * 2. the number of bytes received during the time it takes - * one timestamp to be reflected back to us (the RTT); - * 3. received bytes per RTT is within seven eighth of the - * current socket buffer size; - * 4. receive buffer size has not hit maximal automatic size; - * - * This algorithm does one step per RTT at most and only if - * we receive a bulk stream w/o packet losses or reorderings. - * Shrinking the buffer during idle times is not necessary as - * it doesn't consume any memory when idle. - * - * TODO: Only step up if the application is actually serving - * the buffer to better manage the socket buffer resources. - */ - if (V_tcp_do_autorcvbuf && - (to->to_flags & TOF_TS) && - to->to_tsecr && - (so->so_rcv.sb_flags & SB_AUTOSIZE)) { - if (TSTMP_GT(to->to_tsecr, tp->rfbuf_ts) && - to->to_tsecr - tp->rfbuf_ts < hz) { - if (tp->rfbuf_cnt > - (so->so_rcv.sb_hiwat / 8 * 7) && - so->so_rcv.sb_hiwat < - V_tcp_autorcvbuf_max) { - newsize = - min(so->so_rcv.sb_hiwat + - V_tcp_autorcvbuf_inc, - V_tcp_autorcvbuf_max); - } - /* Start over with next RTT. */ - tp->rfbuf_ts = 0; - tp->rfbuf_cnt = 0; - } else - tp->rfbuf_cnt += tlen; /* add up */ - } + + newsize = tcp_autorcvbuf(m, th, so, tp, tlen); /* Add data to socket buffer. */ SOCKBUF_LOCK(&so->so_rcv); @@ -532,10 +478,6 @@ tcp_do_slowpath(struct mbuf *m, struct tcphdr *th, struct socket *so, win = 0; tp->rcv_wnd = imax(win, (int)(tp->rcv_adv - tp->rcv_nxt)); - /* Reset receive buffer auto scaling when not in bulk receive mode. */ - tp->rfbuf_ts = 0; - tp->rfbuf_cnt = 0; - switch (tp->t_state) { /* diff --git a/sys/netinet/tcp_var.h b/sys/netinet/tcp_var.h index 5705e553d1d..d298c9dd6c6 100644 --- a/sys/netinet/tcp_var.h +++ b/sys/netinet/tcp_var.h @@ -778,6 +778,8 @@ void hhook_run_tcp_est_in(struct tcpcb *tp, #endif int tcp_input(struct mbuf **, int *, int); +int tcp_autorcvbuf(struct mbuf *, struct tcphdr *, struct socket *, + struct tcpcb *, int); void tcp_do_segment(struct mbuf *, struct tcphdr *, struct socket *, struct tcpcb *, int, int, uint8_t, int);