git: 530c2c30b0c7 - main - ip6_output: Reduce cache misses on pktopts
- Go to: [ bottom of page ] [ top of archives ] [ this month ]
Date: Wed, 20 Mar 2024 19:51:13 UTC
The branch main has been updated by gallatin:
URL: https://cgit.FreeBSD.org/src/commit/?id=530c2c30b0c75f1a71df637ae1e09b352f8256cb
commit 530c2c30b0c75f1a71df637ae1e09b352f8256cb
Author: Andrew Gallatin <gallatin@FreeBSD.org>
AuthorDate: 2024-03-20 19:46:01 +0000
Commit: Andrew Gallatin <gallatin@FreeBSD.org>
CommitDate: 2024-03-20 19:50:57 +0000
ip6_output: Reduce cache misses on pktopts
When profiling an IP6 heavy workload, I noticed that we were
getting a lot of cache misses in ip6_output() around
ip6_pktopts. This was happening because the TCP stack passes
inp->in6p_outputopts even if all options are unused. So in the
common case of no options present, pkt_opts is not null, and is
checked repeatedly for different options. Since ip6_pktopts is
large (4 cachelines), and every field is checked, we take 4
cache misses (2 of which tend to be hidden by the adjacent line
prefetcher).
To fix this common case, I introduced a new flag in ip6_pktopts
(ip6po_valid) which tracks which options have been set. In the
common case where nothing is set, this causes just a single
cache miss to load. It also eliminates a test for some options
(if (opt != NULL && opt->val >= const) vs if ((optvalid & flag) !=0 )
To keep the struct the same size in 64-bit kernels, and to keep
the integer values (like ip6po_hlim, ip6po_tclass, etc) on the
same cacheline, I moved them to the top.
As suggested by zlei, the null check in MAKE_EXTHDR() becomes
redundant, and can be removed.
For our web server workload (with the ip6po_tclass option set),
this drops the CPI from 2.9 to 2.4 for ip6_output
Differential Revision: https://reviews.freebsd.org/D44204
Reviewed by: bz, glebius, zlei
No Objection from: melifaro
Sponsored by: Netflix Inc.
---
sys/netinet6/ip6_output.c | 67 ++++++++++++++++++++++++++++++++---------------
sys/netinet6/ip6_var.h | 56 +++++++++++++++++++++++++--------------
2 files changed, 83 insertions(+), 40 deletions(-)
diff --git a/sys/netinet6/ip6_output.c b/sys/netinet6/ip6_output.c
index a2c3efad749b..530f86c36689 100644
--- a/sys/netinet6/ip6_output.c
+++ b/sys/netinet6/ip6_output.c
@@ -159,14 +159,12 @@ static int copypktopts(struct ip6_pktopts *, struct ip6_pktopts *, int);
*/
#define MAKE_EXTHDR(hp, mp, _ol) \
do { \
- if (hp) { \
- struct ip6_ext *eh = (struct ip6_ext *)(hp); \
- error = ip6_copyexthdr((mp), (caddr_t)(hp), \
- ((eh)->ip6e_len + 1) << 3); \
- if (error) \
- goto freehdrs; \
- (_ol) += (*(mp))->m_len; \
- } \
+ struct ip6_ext *eh = (struct ip6_ext *)(hp); \
+ error = ip6_copyexthdr((mp), (caddr_t)(hp), \
+ ((eh)->ip6e_len + 1) << 3); \
+ if (error) \
+ goto freehdrs; \
+ (_ol) += (*(mp))->m_len; \
} while (/*CONSTCOND*/ 0)
/*
@@ -431,6 +429,7 @@ ip6_output(struct mbuf *m0, struct ip6_pktopts *opt,
uint32_t fibnum;
struct m_tag *fwd_tag = NULL;
uint32_t id;
+ uint32_t optvalid;
NET_EPOCH_ASSERT();
@@ -491,14 +490,17 @@ ip6_output(struct mbuf *m0, struct ip6_pktopts *opt,
* Keep the length of the unfragmentable part for fragmentation.
*/
bzero(&exthdrs, sizeof(exthdrs));
- optlen = 0;
+ optlen = optvalid = 0;
unfragpartlen = sizeof(struct ip6_hdr);
if (opt) {
+ optvalid = opt->ip6po_valid;
+
/* Hop-by-Hop options header. */
- MAKE_EXTHDR(opt->ip6po_hbh, &exthdrs.ip6e_hbh, optlen);
+ if ((optvalid & IP6PO_VALID_HBH) != 0)
+ MAKE_EXTHDR(opt->ip6po_hbh, &exthdrs.ip6e_hbh, optlen);
/* Destination options header (1st part). */
- if (opt->ip6po_rthdr) {
+ if ((optvalid & IP6PO_VALID_RHINFO) != 0) {
#ifndef RTHDR_SUPPORT_IMPLEMENTED
/*
* If there is a routing header, discard the packet
@@ -524,11 +526,13 @@ ip6_output(struct mbuf *m0, struct ip6_pktopts *opt,
* options, which might automatically be inserted in
* the kernel.
*/
- MAKE_EXTHDR(opt->ip6po_dest1, &exthdrs.ip6e_dest1,
- optlen);
+ if ((optvalid & IP6PO_VALID_DEST1) != 0)
+ MAKE_EXTHDR(opt->ip6po_dest1, &exthdrs.ip6e_dest1,
+ optlen);
}
/* Routing header. */
- MAKE_EXTHDR(opt->ip6po_rthdr, &exthdrs.ip6e_rthdr, optlen);
+ if ((optvalid & IP6PO_VALID_RHINFO) != 0)
+ MAKE_EXTHDR(opt->ip6po_rthdr, &exthdrs.ip6e_rthdr, optlen);
unfragpartlen += optlen;
@@ -538,7 +542,8 @@ ip6_output(struct mbuf *m0, struct ip6_pktopts *opt,
*/
/* Destination options header (2nd part). */
- MAKE_EXTHDR(opt->ip6po_dest2, &exthdrs.ip6e_dest2, optlen);
+ if ((optvalid & IP6PO_VALID_DEST2) != 0)
+ MAKE_EXTHDR(opt->ip6po_dest2, &exthdrs.ip6e_dest2, optlen);
}
/*
@@ -627,7 +632,7 @@ ip6_output(struct mbuf *m0, struct ip6_pktopts *opt,
/* Route packet. */
ro_pmtu = ro;
- if (opt && opt->ip6po_rthdr)
+ if ((optvalid & IP6PO_VALID_RHINFO) != 0)
ro = &opt->ip6po_route;
if (ro != NULL)
dst = (struct sockaddr_in6 *)&ro->ro_dst;
@@ -641,7 +646,7 @@ again:
* Do not override if a non-zero value is already set.
* We check the diffserv field and the ECN field separately.
*/
- if (opt && opt->ip6po_tclass >= 0) {
+ if ((optvalid & IP6PO_VALID_TC) != 0){
int mask = 0;
if (IPV6_DSCP(ip6) == 0)
@@ -653,7 +658,7 @@ again:
}
/* Fill in or override the hop limit field, if necessary. */
- if (opt && opt->ip6po_hlim != -1)
+ if ((optvalid & IP6PO_VALID_HLIM) != 0)
ip6->ip6_hlim = opt->ip6po_hlim & 0xff;
else if (IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst)) {
if (im6o != NULL)
@@ -855,7 +860,7 @@ nonh6lookup:
/* All scope ID checks are successful. */
if (nh && !IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst)) {
- if (opt && opt->ip6po_nextroute.ro_nh) {
+ if ((optvalid & IP6PO_VALID_NHINFO) != 0) {
/*
* The nexthop is explicitly specified by the
* application. We assume the next hop is an IPv6
@@ -2648,10 +2653,14 @@ ip6_clearpktopts(struct ip6_pktopts *pktopt, int optname)
free(pktopt->ip6po_pktinfo, M_IP6OPT);
pktopt->ip6po_pktinfo = NULL;
}
- if (optname == -1 || optname == IPV6_HOPLIMIT)
+ if (optname == -1 || optname == IPV6_HOPLIMIT) {
pktopt->ip6po_hlim = -1;
- if (optname == -1 || optname == IPV6_TCLASS)
+ pktopt->ip6po_valid &= ~IP6PO_VALID_HLIM;
+ }
+ if (optname == -1 || optname == IPV6_TCLASS) {
pktopt->ip6po_tclass = -1;
+ pktopt->ip6po_valid &= ~IP6PO_VALID_TC;
+ }
if (optname == -1 || optname == IPV6_NEXTHOP) {
if (pktopt->ip6po_nextroute.ro_nh) {
NH_FREE(pktopt->ip6po_nextroute.ro_nh);
@@ -2660,16 +2669,19 @@ ip6_clearpktopts(struct ip6_pktopts *pktopt, int optname)
if (pktopt->ip6po_nexthop)
free(pktopt->ip6po_nexthop, M_IP6OPT);
pktopt->ip6po_nexthop = NULL;
+ pktopt->ip6po_valid &= ~IP6PO_VALID_NHINFO;
}
if (optname == -1 || optname == IPV6_HOPOPTS) {
if (pktopt->ip6po_hbh)
free(pktopt->ip6po_hbh, M_IP6OPT);
pktopt->ip6po_hbh = NULL;
+ pktopt->ip6po_valid &= ~IP6PO_VALID_HBH;
}
if (optname == -1 || optname == IPV6_RTHDRDSTOPTS) {
if (pktopt->ip6po_dest1)
free(pktopt->ip6po_dest1, M_IP6OPT);
pktopt->ip6po_dest1 = NULL;
+ pktopt->ip6po_valid &= ~IP6PO_VALID_DEST1;
}
if (optname == -1 || optname == IPV6_RTHDR) {
if (pktopt->ip6po_rhinfo.ip6po_rhi_rthdr)
@@ -2679,11 +2691,13 @@ ip6_clearpktopts(struct ip6_pktopts *pktopt, int optname)
NH_FREE(pktopt->ip6po_route.ro_nh);
pktopt->ip6po_route.ro_nh = NULL;
}
+ pktopt->ip6po_valid &= ~IP6PO_VALID_RHINFO;
}
if (optname == -1 || optname == IPV6_DSTOPTS) {
if (pktopt->ip6po_dest2)
free(pktopt->ip6po_dest2, M_IP6OPT);
pktopt->ip6po_dest2 = NULL;
+ pktopt->ip6po_valid &= ~IP6PO_VALID_DEST2;
}
}
@@ -2730,6 +2744,7 @@ copypktopts(struct ip6_pktopts *dst, struct ip6_pktopts *src, int canwait)
PKTOPT_EXTHDRCPY(ip6po_dest1);
PKTOPT_EXTHDRCPY(ip6po_dest2);
PKTOPT_EXTHDRCPY(ip6po_rthdr); /* not copy the cached route */
+ dst->ip6po_valid = src->ip6po_valid;
return (0);
bad:
@@ -2959,6 +2974,7 @@ ip6_setpktopt(int optname, u_char *buf, int len, struct ip6_pktopts *opt,
return (ENOBUFS);
}
bcopy(pktinfo, opt->ip6po_pktinfo, sizeof(*pktinfo));
+ opt->ip6po_valid |= IP6PO_VALID_PKTINFO;
break;
}
@@ -2981,6 +2997,7 @@ ip6_setpktopt(int optname, u_char *buf, int len, struct ip6_pktopts *opt,
return (EINVAL);
opt->ip6po_hlim = *hlimp;
+ opt->ip6po_valid |= IP6PO_VALID_HLIM;
break;
}
@@ -2995,6 +3012,7 @@ ip6_setpktopt(int optname, u_char *buf, int len, struct ip6_pktopts *opt,
return (EINVAL);
opt->ip6po_tclass = tclass;
+ opt->ip6po_valid |= IP6PO_VALID_TC;
break;
}
@@ -3045,6 +3063,7 @@ ip6_setpktopt(int optname, u_char *buf, int len, struct ip6_pktopts *opt,
if (opt->ip6po_nexthop == NULL)
return (ENOBUFS);
bcopy(buf, opt->ip6po_nexthop, *buf);
+ opt->ip6po_valid |= IP6PO_VALID_NHINFO;
break;
case IPV6_2292HOPOPTS:
@@ -3083,6 +3102,7 @@ ip6_setpktopt(int optname, u_char *buf, int len, struct ip6_pktopts *opt,
if (opt->ip6po_hbh == NULL)
return (ENOBUFS);
bcopy(hbh, opt->ip6po_hbh, hbhlen);
+ opt->ip6po_valid |= IP6PO_VALID_HBH;
break;
}
@@ -3150,6 +3170,10 @@ ip6_setpktopt(int optname, u_char *buf, int len, struct ip6_pktopts *opt,
if (*newdest == NULL)
return (ENOBUFS);
bcopy(dest, *newdest, destlen);
+ if (newdest == &opt->ip6po_dest1)
+ opt->ip6po_valid |= IP6PO_VALID_DEST1;
+ else
+ opt->ip6po_valid |= IP6PO_VALID_DEST2;
break;
}
@@ -3192,6 +3216,7 @@ ip6_setpktopt(int optname, u_char *buf, int len, struct ip6_pktopts *opt,
if (opt->ip6po_rthdr == NULL)
return (ENOBUFS);
bcopy(rth, opt->ip6po_rthdr, rthlen);
+ opt->ip6po_valid |= IP6PO_VALID_RHINFO;
break;
}
diff --git a/sys/netinet6/ip6_var.h b/sys/netinet6/ip6_var.h
index aeeda14f281f..f3ac13a9caad 100644
--- a/sys/netinet6/ip6_var.h
+++ b/sys/netinet6/ip6_var.h
@@ -130,27 +130,26 @@ struct ip6po_nhinfo {
#define ip6po_nexthop ip6po_nhinfo.ip6po_nhi_nexthop
#define ip6po_nextroute ip6po_nhinfo.ip6po_nhi_route
+/*
+ * Note that fields with valid data must be flagged in ip6po_valid.
+ * This is done to reduce cache misses in ip6_output(). Before
+ * ip6po_valid, ip6_output needed to check all the individual fields
+ * of ip6_pktopts needed to be checked themselves, and they are spread
+ * across 4 cachelines. ip6_output() is currently the only consumer of
+ * these flags, as it is in the critical path of every packet sent.
+ */
struct ip6_pktopts {
- struct mbuf *ip6po_m; /* Pointer to mbuf storing the data */
- int ip6po_hlim; /* Hoplimit for outgoing packets */
-
- /* Outgoing IF/address information */
- struct in6_pktinfo *ip6po_pktinfo;
-
- /* Next-hop address information */
- struct ip6po_nhinfo ip6po_nhinfo;
-
- struct ip6_hbh *ip6po_hbh; /* Hop-by-Hop options header */
-
- /* Destination options header (before a routing header) */
- struct ip6_dest *ip6po_dest1;
-
- /* Routing header related info. */
- struct ip6po_rhinfo ip6po_rhinfo;
-
- /* Destination options header (after a routing header) */
- struct ip6_dest *ip6po_dest2;
+ uint32_t ip6po_valid;
+#define IP6PO_VALID_HLIM 0x0001
+#define IP6PO_VALID_PKTINFO 0x0002
+#define IP6PO_VALID_NHINFO 0x0004
+#define IP6PO_VALID_HBH 0x0008
+#define IP6PO_VALID_DEST1 0x0010
+#define IP6PO_VALID_RHINFO 0x0020
+#define IP6PO_VALID_DEST2 0x0040
+#define IP6PO_VALID_TC 0x0080
+ int ip6po_hlim; /* Hoplimit for outgoing packets */
int ip6po_tclass; /* traffic class */
int ip6po_minmtu; /* fragment vs PMTU discovery policy */
@@ -171,6 +170,25 @@ struct ip6_pktopts {
#endif
#define IP6PO_DONTFRAG 0x04 /* disable fragmentation (IPV6_DONTFRAG) */
#define IP6PO_USECOA 0x08 /* use care of address */
+
+ struct mbuf *ip6po_m; /* Pointer to mbuf storing the data */
+
+ /* Outgoing IF/address information */
+ struct in6_pktinfo *ip6po_pktinfo;
+
+ /* Next-hop address information */
+ struct ip6po_nhinfo ip6po_nhinfo;
+
+ struct ip6_hbh *ip6po_hbh; /* Hop-by-Hop options header */
+
+ /* Destination options header (before a routing header) */
+ struct ip6_dest *ip6po_dest1;
+
+ /* Routing header related info. */
+ struct ip6po_rhinfo ip6po_rhinfo;
+
+ /* Destination options header (after a routing header) */
+ struct ip6_dest *ip6po_dest2;
};
/*