Hi,
Claudio@ mentioned the idea to use TSO and LRO on the loopback
interface to transfer TCP faster.
I see a performance effect with this diff, but more importantly it
gives us more test coverage. Currently LRO on lo(4) is default
off.
Future plan is:
- Fix some corner cases for LRO/TSO with TCP path-MTU discovery
and IP forwarding when LRO is enabled.
- Enable LRO/TSO for lo(4) and ix(4) per default.
- Jan@ commits his ixl(4) TSO diff.
ok for lo(4) LRO/TSO with default off?
bluhm
Index: sys/net/if.c
===================================================================
RCS file: /data/mirror/openbsd/cvs/src/sys/net/if.c,v
retrieving revision 1.700
diff -u -p -r1.700 if.c
--- sys/net/if.c 12 Jun 2023 21:19:54 -0000 1.700
+++ sys/net/if.c 23 Jun 2023 15:48:27 -0000
@@ -106,6 +106,9 @@
#ifdef MROUTING
#include <netinet/ip_mroute.h>
#endif
+#include <netinet/tcp.h>
+#include <netinet/tcp_timer.h>
+#include <netinet/tcp_var.h>
#ifdef INET6
#include <netinet6/in6_var.h>
@@ -802,12 +805,29 @@ if_input_local(struct ifnet *ifp, struct
* is now incorrect, will be calculated before sending.
*/
keepcksum = m->m_pkthdr.csum_flags & (M_IPV4_CSUM_OUT |
- M_TCP_CSUM_OUT | M_UDP_CSUM_OUT | M_ICMP_CSUM_OUT);
+ M_TCP_CSUM_OUT | M_UDP_CSUM_OUT | M_ICMP_CSUM_OUT |
+ M_TCP_TSO);
m_resethdr(m);
m->m_flags |= M_LOOP | keepflags;
m->m_pkthdr.csum_flags = keepcksum;
m->m_pkthdr.ph_ifidx = ifp->if_index;
m->m_pkthdr.ph_rtableid = ifp->if_rdomain;
+
+ if (ISSET(keepcksum, M_TCP_TSO) && m->m_pkthdr.len > ifp->if_mtu) {
+ if (ifp->if_mtu > 0 &&
+ ((af == AF_INET &&
+ ISSET(ifp->if_capabilities, IFCAP_TSOv4)) ||
+ (af == AF_INET6 &&
+ ISSET(ifp->if_capabilities, IFCAP_TSOv6)))) {
+ tcpstat_inc(tcps_inswlro);
+ tcpstat_add(tcps_inpktlro,
+ (m->m_pkthdr.len + ifp->if_mtu - 1) / ifp->if_mtu);
+ } else {
+ tcpstat_inc(tcps_inbadlro);
+ m_freem(m);
+ return (EPROTONOSUPPORT);
+ }
+ }
if (ISSET(keepcksum, M_TCP_CSUM_OUT))
m->m_pkthdr.csum_flags |= M_TCP_CSUM_IN_OK;
Index: sys/net/if_loop.c
===================================================================
RCS file: /data/mirror/openbsd/cvs/src/sys/net/if_loop.c,v
retrieving revision 1.94
diff -u -p -r1.94 if_loop.c
--- sys/net/if_loop.c 5 Jun 2023 11:35:46 -0000 1.94
+++ sys/net/if_loop.c 23 Jun 2023 15:48:27 -0000
@@ -175,7 +175,8 @@ loop_clone_create(struct if_clone *ifc,
ifp->if_xflags = IFXF_CLONED;
ifp->if_capabilities = IFCAP_CSUM_IPv4 |
IFCAP_CSUM_TCPv4 | IFCAP_CSUM_UDPv4 |
- IFCAP_CSUM_TCPv6 | IFCAP_CSUM_UDPv6;
+ IFCAP_CSUM_TCPv6 | IFCAP_CSUM_UDPv6 |
+ IFCAP_LRO;
ifp->if_rtrequest = lortrequest;
ifp->if_ioctl = loioctl;
ifp->if_input = loinput;
@@ -281,6 +282,10 @@ loioctl(struct ifnet *ifp, u_long cmd, c
switch (cmd) {
case SIOCSIFFLAGS:
+ if (ISSET(ifp->if_xflags, IFXF_LRO))
+ SET(ifp->if_capabilities, IFCAP_TSOv4 | IFCAP_TSOv6);
+ else
+ CLR(ifp->if_capabilities, IFCAP_TSOv4 | IFCAP_TSOv6);
break;
case SIOCSIFADDR:
Index: sys/netinet/tcp_usrreq.c
===================================================================
RCS file: /data/mirror/openbsd/cvs/src/sys/netinet/tcp_usrreq.c,v
retrieving revision 1.219
diff -u -p -r1.219 tcp_usrreq.c
--- sys/netinet/tcp_usrreq.c 23 May 2023 09:16:16 -0000 1.219
+++ sys/netinet/tcp_usrreq.c 23 Jun 2023 15:48:27 -0000
@@ -1340,6 +1340,7 @@ tcp_sysctl_tcpstat(void *oldp, size_t *o
ASSIGN(tcps_outhwtso);
ASSIGN(tcps_outpkttso);
ASSIGN(tcps_outbadtso);
+ ASSIGN(tcps_inswlro);
ASSIGN(tcps_inhwlro);
ASSIGN(tcps_inpktlro);
ASSIGN(tcps_inbadlro);
Index: sys/netinet/tcp_var.h
===================================================================
RCS file: /data/mirror/openbsd/cvs/src/sys/netinet/tcp_var.h,v
retrieving revision 1.167
diff -u -p -r1.167 tcp_var.h
--- sys/netinet/tcp_var.h 23 May 2023 09:16:16 -0000 1.167
+++ sys/netinet/tcp_var.h 23 Jun 2023 15:48:27 -0000
@@ -447,6 +447,7 @@ struct tcpstat {
u_int32_t tcps_outhwtso; /* output tso processed by hardware */
u_int32_t tcps_outpkttso; /* packets generated by tso */
u_int32_t tcps_outbadtso; /* output tso failed, packet dropped */
+ u_int32_t tcps_inswlro; /* input lro on pseudo device */
u_int32_t tcps_inhwlro; /* input lro from hardware */
u_int32_t tcps_inpktlro; /* packets coalesced by hardware lro */
u_int32_t tcps_inbadlro; /* input bad lro packets */
@@ -628,6 +629,7 @@ enum tcpstat_counters {
tcps_outhwtso,
tcps_outpkttso,
tcps_outbadtso,
+ tcps_inswlro,
tcps_inhwlro,
tcps_inpktlro,
tcps_inbadlro,
Index: usr.bin/netstat/inet.c
===================================================================
RCS file: /data/mirror/openbsd/cvs/src/usr.bin/netstat/inet.c,v
retrieving revision 1.176
diff -u -p -r1.176 inet.c
--- usr.bin/netstat/inet.c 23 May 2023 09:16:16 -0000 1.176
+++ usr.bin/netstat/inet.c 23 Jun 2023 15:55:55 -0000
@@ -439,9 +439,12 @@ tcp_stats(char *name)
p(tcps_inswcsum, "\t\t%u packet%s software-checksummed\n");
p(tcps_rcvbadsig, "\t\t%u bad/missing md5 checksum%s\n");
p(tcps_rcvgoodsig, "\t\t%llu good md5 checksum%s\n");
+ p(tcps_inswlro,
+ "\t\t%u input LRO packet%s passed through pseudo device\n");
p(tcps_inhwlro, "\t\t%u input LRO generated packet%s from hardware\n");
- p(tcps_inpktlro, "\t\t%u input LRO coalesced packet%s by hardware\n");
- p(tcps_inbadlro, "\t\t%u input bad LRO packet%s\n");
+ p(tcps_inpktlro,
+ "\t\t%u input LRO coalesced packet%s by network device\n");
+ p(tcps_inbadlro, "\t\t%u input bad LRO packet%s dropped\n");
p(tcps_connattempt, "\t%u connection request%s\n");
p(tcps_accepts, "\t%u connection accept%s\n");
p(tcps_connects, "\t%u connection%s established (including accepts)\n");