D29564.id86781.diff
No OneTemporary
Actions

Size

78 KB

Referenced Files

None

Subscribers

None

D29564.id86781.diff
View Options

	diff --git a/sys/netinet/tcp_lro.h b/sys/netinet/tcp_lro.h
	--- a/sys/netinet/tcp_lro.h
	+++ b/sys/netinet/tcp_lro.h
	@@ -3,7 +3,7 @@
	*
	* Copyright (c) 2006, Myricom Inc.
	* Copyright (c) 2008, Intel Corporation.
	- * Copyright (c) 2016 Mellanox Technologies.
	+ * Copyright (c) 2016-2021 Mellanox Technologies.
	* All rights reserved.
	*
	* Redistribution and use in source and binary forms, with or without
	@@ -57,69 +57,91 @@
	#define TSTMP_HDWR 0x0200
	#define HAS_TSTMP 0x0400

	-/* Flags in LRO entry */
	-#define CAN_USE_ACKCMP 0x0001
	-#define HAS_COMP_ENTRIES 0x0002
	-
	struct inpcb;

	+union lro_address {
	+ u_long raw[1];
	+ struct {
	+ uint16_t lro_type; /* internal */
	+#define LRO_TYPE_NONE 0
	+#define LRO_TYPE_IPV4_TCP 1
	+#define LRO_TYPE_IPV6_TCP 2
	+#define LRO_TYPE_IPV4_UDP 3
	+#define LRO_TYPE_IPV6_UDP 4
	+ uint16_t vlan_id; /* VLAN identifier */
	+ uint16_t s_port; /* source TCP/UDP port */
	+ uint16_t d_port; /* source TCP/UDP port */
	+ uint32_t vxlan_vni; /* VXLAN virtual network identifier */
	+ union {
	+#ifdef INET
	+ struct in_addr v4;
	+#endif
	+#ifdef INET6
	+ struct in6_addr v6;
	+#endif
	+ } s_addr; /* source IPv4/IPv6 address */
	+ union {
	+#ifdef INET
	+ struct in_addr v4;
	+#endif
	+#ifdef INET6
	+ struct in6_addr v6;
	+#endif
	+ } d_addr; /* destination IPv4/IPv6 address */
	+ };
	+} __aligned(sizeof(uint64_t));
	+
	+#define LRO_RAW_ADDRESS_MAX \
	+ (sizeof(union lro_address) / sizeof(u_long))
	+
	+/* Optimize address comparison by comparing 64-bits at a time: */
	+
	+static inline bool
	+lro_address_compare(const union lro_address pa, const union lro_address pb)
	+{
	+ if (pa->lro_type == LRO_TYPE_NONE && pb->lro_type == LRO_TYPE_NONE) {
	+ return (true);
	+ } else for (unsigned i = 0; i < LRO_RAW_ADDRESS_MAX; i++) {
	+ if (pa->raw[i] != pb->raw[i])
	+ return (false);
	+ }
	+ return (true);
	+}
	+
	+struct lro_parser {
	+ union lro_address data;
	+ union {
	+ uint8_t *l3;
	+ struct ip *ip4;
	+ struct ip6_hdr *ip6;
	+ };
	+ union {
	+ uint8_t *l4;
	+ struct tcphdr *tcp;
	+ struct udphdr *udp;
	+ };
	+ uint16_t total_hdr_len;
	+};
	+
	+/* This structure is zeroed frequently, try to keep it small. */
	struct lro_entry {
	LIST_ENTRY(lro_entry) next;
	LIST_ENTRY(lro_entry) hash_next;
	struct mbuf *m_head;
	struct mbuf *m_tail;
	struct mbuf *m_last_mbuf;
	- struct mbuf *m_prev_last;
	- struct inpcb *inp;
	- union {
	- struct ip *ip4;
	- struct ip6_hdr *ip6;
	- } leip;
	- union {
	- in_addr_t s_ip4;
	- struct in6_addr s_ip6;
	- } lesource;
	- union {
	- in_addr_t d_ip4;
	- struct in6_addr d_ip6;
	- } ledest;
	- uint16_t source_port;
	- uint16_t dest_port;
	- uint16_t eh_type; /* EthernetHeader type. */
	- uint16_t append_cnt;
	- uint32_t p_len; /* IP header payload length. */
	- uint32_t ulp_csum; /* TCP, etc. checksum. */
	+ struct lro_parser outer;
	+ struct lro_parser inner;
	uint32_t next_seq; /* tcp_seq */
	uint32_t ack_seq; /* tcp_seq */
	uint32_t tsval;
	uint32_t tsecr;
	- uint32_t tcp_tot_p_len; /* TCP payload length of chain */
	uint16_t window;
	uint16_t timestamp; /* flag, not a TCP hdr field. */
	- uint16_t need_wakeup;
	- uint16_t mbuf_cnt; /* Count of mbufs collected see note */
	- uint16_t mbuf_appended;
	- uint16_t cmp_ack_cnt;
	- uint16_t flags;
	- uint16_t strip_cnt;
	- struct timeval mtime;
	+ sbintime_t alloc_time; /* time when entry was allocated */
	};
	-/*
	- * Note: The mbuf_cnt field tracks our number of mbufs added to the m_next
	- * list. Each mbuf counted can have data and of course it will
	- * have an ack as well (by defintion any inbound tcp segment will
	- * have an ack value. We use this count to tell us how many ACK's
	- * are present for our ack-count threshold. If we exceed that or
	- * the data threshold we will wake up the endpoint.
	- */
	-LIST_HEAD(lro_head, lro_entry);

	-#define le_ip4 leip.ip4
	-#define le_ip6 leip.ip6
	-#define source_ip4 lesource.s_ip4
	-#define dest_ip4 ledest.d_ip4
	-#define source_ip6 lesource.s_ip6
	-#define dest_ip6 ledest.d_ip6
	+LIST_HEAD(lro_head, lro_entry);

	struct lro_mbuf_sort {
	uint64_t seq;
	@@ -130,7 +152,7 @@
	struct lro_ctrl {
	struct ifnet *ifp;
	struct lro_mbuf_sort *lro_mbuf_data;
	- struct timeval lro_last_flush;
	+ sbintime_t lro_last_queue_time; /* last time data was queued */
	uint64_t lro_queued;
	uint64_t lro_flushed;
	uint64_t lro_bad_csum;
	@@ -139,7 +161,6 @@
	unsigned lro_mbuf_max;
	unsigned short lro_ackcnt_lim; /* max # of aggregated ACKs */
	unsigned lro_length_lim; /* max len of aggregated data */
	-
	u_long lro_hashsz;
	struct lro_head *lro_hash;
	struct lro_head lro_active;
	@@ -163,7 +184,7 @@
	#define M_ACKCMP M_PROTO4 /* Indicates LRO is sending in a Ack-compression mbuf */
	#define M_LRO_EHDRSTRP M_PROTO6 /* Indicates that LRO has stripped the etherenet header */

	-#define TCP_LRO_LENGTH_MAX 65535
	+#define TCP_LRO_LENGTH_MAX (65535 - 255) /* safe value with room for outer headers */
	#define TCP_LRO_ACKCNT_MAX 65535 /* unlimited */

	int tcp_lro_init(struct lro_ctrl *);
	diff --git a/sys/netinet/tcp_lro.c b/sys/netinet/tcp_lro.c
	--- a/sys/netinet/tcp_lro.c
	+++ b/sys/netinet/tcp_lro.c
	@@ -4,7 +4,7 @@
	* Copyright (c) 2007, Myricom Inc.
	* Copyright (c) 2008, Intel Corporation.
	* Copyright (c) 2012 The FreeBSD Foundation
	- * Copyright (c) 2016 Mellanox Technologies.
	+ * Copyright (c) 2016-2021 Mellanox Technologies.
	* All rights reserved.
	*
	* Portions of this software were developed by Bjoern Zeeb
	@@ -68,25 +68,29 @@
	#include <netinet/tcpip.h>
	#include <netinet/tcp_hpts.h>
	#include <netinet/tcp_log_buf.h>
	+#include <netinet/udp.h>
	#include <netinet6/ip6_var.h>

	#include <machine/in_cksum.h>

	static MALLOC_DEFINE(M_LRO, "LRO", "LRO control structures");

	-#define TCP_LRO_UPDATE_CSUM 1
	-#ifndef TCP_LRO_UPDATE_CSUM
	-#define TCP_LRO_INVALID_CSUM 0x0000
	-#endif
	+#define TCP_LRO_TS_OPTION \
	+ ntohl((TCPOPT_NOP << 24) \| (TCPOPT_NOP << 16) \| \
	+ (TCPOPT_TIMESTAMP << 8) \| TCPOLEN_TIMESTAMP)

	static void tcp_lro_rx_done(struct lro_ctrl *lc);
	-static int tcp_lro_rx2(struct lro_ctrl lc, struct mbuf m,
	- uint32_t csum, int use_hash);
	+static int tcp_lro_rx_common(struct lro_ctrl lc, struct mbuf m,
	+ uint32_t csum, bool use_hash);
	+#ifdef TCPHPTS
	+static bool do_bpf_strip_and_compress(struct inpcb , struct lro_ctrl ,
	+ struct lro_entry , struct mbuf , unsigned , bool);
	+#endif

	SYSCTL_NODE(_net_inet_tcp, OID_AUTO, lro, CTLFLAG_RW \| CTLFLAG_MPSAFE, 0,
	"TCP LRO");

	-static long tcplro_stacks_wanting_mbufq = 0;
	+static long tcplro_stacks_wanting_mbufq;
	counter_u64_t tcp_inp_lro_direct_queue;
	counter_u64_t tcp_inp_lro_wokeup_queue;
	counter_u64_t tcp_inp_lro_compressed;
	@@ -101,7 +105,6 @@
	counter_u64_t tcp_csum_hardware_w_ph;
	counter_u64_t tcp_csum_software;

	-
	static unsigned tcp_lro_entries = TCP_LRO_ENTRIES;
	SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, entries,
	CTLFLAG_RDTUN \| CTLFLAG_MPSAFE, &tcp_lro_entries, 0,
	@@ -134,8 +137,6 @@
	SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, csum_sw, CTLFLAG_RD,
	&tcp_csum_software, "Number of checksums processed in software");

	-
	-
	void
	tcp_lro_reg_mbufq(void)
	{
	@@ -226,34 +227,220 @@
	return (0);
	}

	-static struct tcphdr *
	-tcp_lro_get_th(struct lro_entry le, struct mbuf m)
	+struct vxlan_header {
	+ uint32_t vxlh_flags;
	+ uint32_t vxlh_vni;
	+};
	+
	+static inline void *
	+tcp_lro_low_level_parser(void ptr, struct lro_parser parser, bool update_data, bool is_vxlan)
	{
	- struct ether_header *eh;
	- struct tcphdr *th = NULL;
	-#ifdef INET6
	- struct ip6_hdr ip6 = NULL; / Keep compiler happy. */
	-#endif
	+ const struct ether_vlan_header *eh;
	+ void *old;
	+ uint16_t eth_type;
	+
	+ memset(parser, 0, sizeof(*parser));
	+ old = ptr;
	+
	+ if (is_vxlan) {
	+ const struct vxlan_header *vxh;
	+ vxh = ptr;
	+ ptr = (uint8_t )ptr + sizeof(vxh);
	+ if (update_data) {
	+ parser->data.vxlan_vni =
	+ vxh->vxlh_vni & htonl(0xffffff00);
	+ }
	+ }
	+
	+ eh = ptr;
	+ if (__predict_false(eh->evl_encap_proto == htons(ETHERTYPE_VLAN))) {
	+ eth_type = eh->evl_proto;
	+ if (update_data) {
	+ /* strip priority and keep VLAN ID only */
	+ parser->data.vlan_id = eh->evl_tag & htons(EVL_VLID_MASK);
	+ }
	+ /* advance to next header */
	+ ptr = (uint8_t *)ptr + ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
	+ } else {
	+ eth_type = eh->evl_encap_proto;
	+ /* advance to next header */
	+ ptr = (uint8_t *)ptr + ETHER_HDR_LEN;
	+ }
	+
	+ switch (eth_type) {
	#ifdef INET
	- struct ip ip4 = NULL; / Keep compiler happy. */
	+ case htons(ETHERTYPE_IP):
	+ parser->ip4 = ptr;
	+ /* Ensure there are no IPv4 options. */
	+ if ((parser->ip4->ip_hl << 2) != sizeof (*parser->ip4))
	+ break;
	+ /* .. and the packet is not fragmented. */
	+ if (parser->ip4->ip_off & htons(IP_MF\|IP_OFFMASK))
	+ break;
	+ ptr = (uint8_t *)ptr + (parser->ip4->ip_hl << 2);
	+ if (update_data) {
	+ parser->data.s_addr.v4 = parser->ip4->ip_src;
	+ parser->data.d_addr.v4 = parser->ip4->ip_dst;
	+ }
	+ switch (parser->ip4->ip_p) {
	+ case IPPROTO_UDP:
	+ parser->udp = ptr;
	+ if (update_data) {
	+ parser->data.lro_type = LRO_TYPE_IPV4_UDP;
	+ parser->data.s_port = parser->udp->uh_sport;
	+ parser->data.d_port = parser->udp->uh_dport;
	+ }
	+ ptr = ((uint8_t )ptr + sizeof(parser->udp));
	+ parser->total_hdr_len = (uint8_t )ptr - (uint8_t )old;
	+ return (ptr);
	+ case IPPROTO_TCP:
	+ parser->tcp = ptr;
	+ if (update_data) {
	+ parser->data.lro_type = LRO_TYPE_IPV4_TCP;
	+ parser->data.s_port = parser->tcp->th_sport;
	+ parser->data.d_port = parser->tcp->th_dport;
	+ }
	+ ptr = (uint8_t *)ptr + (parser->tcp->th_off << 2);
	+ parser->total_hdr_len = (uint8_t )ptr - (uint8_t )old;
	+ return (ptr);
	+ default:
	+ break;
	+ }
	+ break;
	#endif
	-
	- eh = mtod(m, struct ether_header *);
	- switch (le->eh_type) {
	#ifdef INET6
	- case ETHERTYPE_IPV6:
	- ip6 = (struct ip6_hdr *)(eh + 1);
	- th = (struct tcphdr *)(ip6 + 1);
	+ case htons(ETHERTYPE_IPV6):
	+ parser->ip6 = ptr;
	+ ptr = (uint8_t )ptr + sizeof(parser->ip6);
	+ if (update_data) {
	+ parser->data.s_addr.v6 = parser->ip6->ip6_src;
	+ parser->data.d_addr.v6 = parser->ip6->ip6_dst;
	+ }
	+ switch (parser->ip6->ip6_nxt) {
	+ case IPPROTO_UDP:
	+ parser->udp = ptr;
	+ if (update_data) {
	+ parser->data.lro_type = LRO_TYPE_IPV6_UDP;
	+ parser->data.s_port = parser->udp->uh_sport;
	+ parser->data.d_port = parser->udp->uh_dport;
	+ }
	+ ptr = (uint8_t )ptr + sizeof(parser->udp);
	+ parser->total_hdr_len = (uint8_t )ptr - (uint8_t )old;
	+ return (ptr);
	+ case IPPROTO_TCP:
	+ parser->tcp = ptr;
	+ if (update_data) {
	+ parser->data.lro_type = LRO_TYPE_IPV6_TCP;
	+ parser->data.s_port = parser->tcp->th_sport;
	+ parser->data.d_port = parser->tcp->th_dport;
	+ }
	+ ptr = (uint8_t *)ptr + (parser->tcp->th_off << 2);
	+ parser->total_hdr_len = (uint8_t )ptr - (uint8_t )old;
	+ return (ptr);
	+ default:
	+ break;
	+ }
	break;
	#endif
	+ default:
	+ break;
	+ }
	+ /* Invalid packet - cannot parse */
	+ return (NULL);
	+}
	+
	+static const int vxlan_csum = CSUM_INNER_L3_CALC \| CSUM_INNER_L3_VALID \|
	+ CSUM_INNER_L4_CALC \| CSUM_INNER_L4_VALID;
	+
	+static inline struct lro_parser *
	+tcp_lro_parser(struct mbuf m, struct lro_parser po, struct lro_parser *pi, bool update_data)
	+{
	+ void *data_ptr;
	+
	+ /* Try to parse outer headers first. */
	+ data_ptr = tcp_lro_low_level_parser(m->m_data, po, update_data, false);
	+ if (data_ptr == NULL \|\| po->total_hdr_len > m->m_len)
	+ return (NULL);
	+
	+ if (update_data) {
	+ /* Store VLAN ID, if any. */
	+ if (__predict_false(m->m_flags & M_VLANTAG)) {
	+ po->data.vlan_id =
	+ htons(m->m_pkthdr.ether_vtag) & htons(EVL_VLID_MASK);
	+ }
	+ }
	+
	+ switch (po->data.lro_type) {
	+ case LRO_TYPE_IPV4_UDP:
	+ case LRO_TYPE_IPV6_UDP:
	+ /* Check for VXLAN headers. */
	+ if ((m->m_pkthdr.csum_flags & vxlan_csum) != vxlan_csum)
	+ break;
	+
	+ /* Try to parse inner headers. */
	+ data_ptr = tcp_lro_low_level_parser(data_ptr, pi, update_data, true);
	+ if (data_ptr == NULL \|\| pi->total_hdr_len > m->m_len)
	+ break;
	+
	+ /* Verify supported header types. */
	+ switch (pi->data.lro_type) {
	+ case LRO_TYPE_IPV4_TCP:
	+ case LRO_TYPE_IPV6_TCP:
	+ return (pi);
	+ default:
	+ break;
	+ }
	+ break;
	+ case LRO_TYPE_IPV4_TCP:
	+ case LRO_TYPE_IPV6_TCP:
	+ if (update_data)
	+ memset(pi, 0, sizeof(*pi));
	+ return (po);
	+ default:
	+ break;
	+ }
	+ return (NULL);
	+}
	+
	+static inline int
	+tcp_lro_trim_mbuf_chain(struct mbuf m, const struct lro_parser po)
	+{
	+ int len;
	+
	+ switch (po->data.lro_type) {
	#ifdef INET
	- case ETHERTYPE_IP:
	- ip4 = (struct ip *)(eh + 1);
	- th = (struct tcphdr *)(ip4 + 1);
	+ case LRO_TYPE_IPV4_TCP:
	+ len = ((uint8_t )po->ip4 - (uint8_t )m->m_data) +
	+ ntohs(po->ip4->ip_len);
	+ break;
	+#endif
	+#ifdef INET6
	+ case LRO_TYPE_IPV6_TCP:
	+ len = ((uint8_t )po->ip6 - (uint8_t )m->m_data) +
	+ ntohs(po->ip6->ip6_plen) + sizeof(*po->ip6);
	break;
	#endif
	+ default:
	+ return (TCP_LRO_CANNOT);
	+ }
	+
	+ /*
	+ * If the frame is padded beyond the end of the IP packet,
	+ * then trim the extra bytes off:
	+ */
	+ if (__predict_true(m->m_pkthdr.len == len)) {
	+ return (0);
	+ } else if (m->m_pkthdr.len > len) {
	+ m_adj(m, len - m->m_pkthdr.len);
	+ return (0);
	}
	- return (th);
	+ return (TCP_LRO_CANNOT);
	+}
	+
	+static struct tcphdr *
	+tcp_lro_get_th(struct mbuf *m)
	+{
	+ return ((struct tcphdr )((uint8_t )m->m_data + m->m_pkthdr.lro_tcp_h_off));
	}

	static void
	@@ -300,86 +487,67 @@
	}

	static uint16_t
	-tcp_lro_csum_th(struct tcphdr *th)
	+tcp_lro_rx_csum_tcphdr(const struct tcphdr *th)
	{
	- uint32_t ch;
	- uint16_t *p, l;
	-
	- ch = th->th_sum = 0x0000;
	- l = th->th_off;
	- p = (uint16_t *)th;
	- while (l > 0) {
	- ch += *p;
	- p++;
	- ch += *p;
	- p++;
	- l--;
	+ const uint16_t *ptr;
	+ uint32_t csum;
	+ uint16_t len;
	+
	+ csum = -th->th_sum; /* exclude checksum field */
	+ len = th->th_off;
	+ ptr = (const uint16_t *)th;
	+ while (len--) {
	+ csum += *ptr;
	+ ptr++;
	+ csum += *ptr;
	+ ptr++;
	}
	- while (ch > 0xffff)
	- ch = (ch >> 16) + (ch & 0xffff);
	+ while (csum > 0xffff)
	+ csum = (csum >> 16) + (csum & 0xffff);

	- return (ch & 0xffff);
	+ return (csum);
	}

	static uint16_t
	-tcp_lro_rx_csum_fixup(struct lro_entry le, void l3hdr, struct tcphdr *th,
	- uint16_t tcp_data_len, uint16_t csum)
	+tcp_lro_rx_csum_data(const struct lro_parser *pa, uint16_t tcp_csum)
	{
	uint32_t c;
	uint16_t cs;

	- c = csum;
	+ c = tcp_csum;

	- /* Remove length from checksum. */
	- switch (le->eh_type) {
	+ switch (pa->data.lro_type) {
	#ifdef INET6
	- case ETHERTYPE_IPV6:
	- {
	- struct ip6_hdr *ip6;
	-
	- ip6 = (struct ip6_hdr *)l3hdr;
	- if (le->append_cnt == 0)
	- cs = ip6->ip6_plen;
	- else {
	- uint32_t cx;
	-
	- cx = ntohs(ip6->ip6_plen);
	- cs = in6_cksum_pseudo(ip6, cx, ip6->ip6_nxt, 0);
	- }
	+ case LRO_TYPE_IPV6_TCP:
	+ /* Compute full pseudo IPv6 header checksum. */
	+ cs = in6_cksum_pseudo(pa->ip6, ntohs(pa->ip6->ip6_plen), pa->ip6->ip6_nxt, 0);
	break;
	- }
	#endif
	#ifdef INET
	- case ETHERTYPE_IP:
	- {
	- struct ip *ip4;
	-
	- ip4 = (struct ip *)l3hdr;
	- if (le->append_cnt == 0)
	- cs = ip4->ip_len;
	- else {
	- cs = in_addword(ntohs(ip4->ip_len) - sizeof(*ip4),
	- IPPROTO_TCP);
	- cs = in_pseudo(ip4->ip_src.s_addr, ip4->ip_dst.s_addr,
	- htons(cs));
	- }
	+ case LRO_TYPE_IPV4_TCP:
	+ /* Compute full pseudo IPv4 header checsum. */
	+ cs = in_addword(ntohs(pa->ip4->ip_len) - sizeof(*pa->ip4), IPPROTO_TCP);
	+ cs = in_pseudo(pa->ip4->ip_src.s_addr, pa->ip4->ip_dst.s_addr, htons(cs));
	break;
	- }
	#endif
	default:
	cs = 0; /* Keep compiler happy. */
	+ break;
	}

	+ /* Complement checksum. */
	cs = ~cs;
	c += cs;

	- /* Remove TCP header csum. */
	- cs = ~tcp_lro_csum_th(th);
	+ /* Remove TCP header checksum. */
	+ cs = ~tcp_lro_rx_csum_tcphdr(pa->tcp);
	c += cs;
	+
	+ /* Compute checksum remainder. */
	while (c > 0xffff)
	c = (c >> 16) + (c & 0xffff);

	- return (c & 0xffff);
	+ return (c);
	}

	static void
	@@ -397,84 +565,51 @@
	tcp_lro_flush_inactive(struct lro_ctrl lc, const struct timeval timeout)
	{
	struct lro_entry le, le_tmp;
	- struct timeval tv;
	+ sbintime_t sbt;

	if (LIST_EMPTY(&lc->lro_active))
	return;

	- getmicrouptime(&tv);
	- timevalsub(&tv, timeout);
	+ /* get timeout time */
	+ sbt = getsbinuptime() - tvtosbt(*timeout);
	+
	LIST_FOREACH_SAFE(le, &lc->lro_active, next, le_tmp) {
	- if (timevalcmp(&tv, &le->mtime, >=)) {
	+ if (sbt >= le->alloc_time) {
	tcp_lro_active_remove(le);
	tcp_lro_flush(lc, le);
	}
	}
	}

	-#ifdef INET6
	-static int
	-tcp_lro_rx_ipv6(struct lro_ctrl lc, struct mbuf m, struct ip6_hdr *ip6,
	- struct tcphdr **th)
	-{
	-
	- /* XXX-BZ we should check the flow-label. */
	-
	- /* XXX-BZ We do not yet support ext. hdrs. */
	- if (ip6->ip6_nxt != IPPROTO_TCP)
	- return (TCP_LRO_NOT_SUPPORTED);
	-
	- /* Find the TCP header. */
	- th = (struct tcphdr )(ip6 + 1);
	-
	- return (0);
	-}
	-#endif
	-
	#ifdef INET
	static int
	-tcp_lro_rx_ipv4(struct lro_ctrl lc, struct mbuf m, struct ip *ip4,
	- struct tcphdr **th)
	+tcp_lro_rx_ipv4(struct lro_ctrl lc, struct mbuf m, struct ip *ip4)
	{
	- int csum_flags;
	uint16_t csum;

	- if (ip4->ip_p != IPPROTO_TCP)
	- return (TCP_LRO_NOT_SUPPORTED);
	-
	- /* Ensure there are no options. */
	- if ((ip4->ip_hl << 2) != sizeof (*ip4))
	- return (TCP_LRO_CANNOT);
	-
	- /* .. and the packet is not fragmented. */
	- if (ip4->ip_off & htons(IP_MF\|IP_OFFMASK))
	- return (TCP_LRO_CANNOT);
	-
	/* Legacy IP has a header checksum that needs to be correct. */
	- csum_flags = m->m_pkthdr.csum_flags;
	- if (csum_flags & CSUM_IP_CHECKED) {
	- if (__predict_false((csum_flags & CSUM_IP_VALID) == 0)) {
	+ if (m->m_pkthdr.csum_flags & CSUM_IP_CHECKED) {
	+ if (__predict_false((m->m_pkthdr.csum_flags & CSUM_IP_VALID) == 0)) {
	lc->lro_bad_csum++;
	return (TCP_LRO_CANNOT);
	}
	} else {
	csum = in_cksum_hdr(ip4);
	- if (__predict_false((csum) != 0)) {
	+ if (__predict_false(csum != 0)) {
	lc->lro_bad_csum++;
	return (TCP_LRO_CANNOT);
	}
	}
	- /* Find the TCP header (we assured there are no IP options). */
	- th = (struct tcphdr )(ip4 + 1);
	return (0);
	}
	#endif

	#ifdef TCPHPTS
	static void
	-tcp_lro_log(struct tcpcb tp, struct lro_ctrl lc,
	- struct lro_entry le, struct mbuf m, int frm, int32_t tcp_data_len,
	- uint32_t th_seq , uint32_t th_ack, uint16_t th_win)
	+tcp_lro_log(struct tcpcb tp, const struct lro_ctrl lc,
	+ const struct lro_entry le, const struct mbuf m,
	+ int frm, int32_t tcp_data_len, uint32_t th_seq,
	+ uint32_t th_ack, uint16_t th_win)
	{
	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
	union tcp_log_stackspecific log;
	@@ -489,8 +624,8 @@
	log.u_bbr.flex2 = m->m_pkthdr.len;
	else
	log.u_bbr.flex2 = 0;
	- log.u_bbr.flex3 = le->append_cnt;
	- log.u_bbr.flex4 = le->p_len;
	+ log.u_bbr.flex3 = le->m_head->m_pkthdr.lro_nsegs;
	+ log.u_bbr.flex4 = le->m_head->m_pkthdr.lro_tcp_d_len;
	if (le->m_head) {
	log.u_bbr.flex5 = le->m_head->m_pkthdr.len;
	log.u_bbr.delRate = le->m_head->m_flags;
	@@ -505,11 +640,7 @@
	log.u_bbr.cwnd_gain = le->window;
	log.u_bbr.cur_del_rate = (uintptr_t)m;
	log.u_bbr.bw_inuse = (uintptr_t)le->m_head;
	- log.u_bbr.pkts_out = le->mbuf_cnt; /* Total mbufs added */
	- log.u_bbr.applimited = le->ulp_csum;
	- log.u_bbr.lost = le->mbuf_appended;
	- log.u_bbr.pkt_epoch = le->cmp_ack_cnt;
	- log.u_bbr.flex6 = tcp_tv_to_usectick(&lc->lro_last_flush);
	+ log.u_bbr.flex6 = sbttous(lc->lro_last_queue_time);
	if (in_epoch(net_epoch_preempt))
	log.u_bbr.inhpts = 1;
	else
	@@ -523,198 +654,292 @@
	}
	#endif

	-static void
	-tcp_flush_out_le(struct tcpcb tp, struct lro_ctrl lc, struct lro_entry *le)
	+static inline void
	+tcp_lro_assign_and_checksum_16(uint16_t ptr, uint16_t value, uint16_t psum)
	{
	- if (le->append_cnt > 1) {
	- struct tcphdr *th;
	- uint16_t p_len;
	+ uint32_t csum;

	- p_len = htons(le->p_len);
	- switch (le->eh_type) {
	-#ifdef INET6
	- case ETHERTYPE_IPV6:
	- {
	- struct ip6_hdr *ip6;
	+ csum = 0xffff - *ptr + value;
	+ while (csum > 0xffff)
	+ csum = (csum >> 16) + (csum & 0xffff);
	+ *ptr = value;
	+ *psum = csum;
	+}
	+
	+static uint16_t
	+tcp_lro_update_checksum(const struct lro_parser pa, const struct lro_entry le,
	+ uint16_t payload_len, uint16_t delta_sum)
	+{
	+ uint32_t csum;
	+ uint16_t tlen;
	+ uint16_t temp[5] = {};
	+
	+ switch (pa->data.lro_type) {
	+ case LRO_TYPE_IPV4_TCP:
	+ /* Compute new IPv4 length. */
	+ tlen = (pa->ip4->ip_hl << 2) + (pa->tcp->th_off << 2) + payload_len;
	+ tcp_lro_assign_and_checksum_16(&pa->ip4->ip_len, htons(tlen), &temp[0]);
	+
	+ /* Subtract delta from current IPv4 checksum. */
	+ csum = pa->ip4->ip_sum + 0xffff - temp[0];
	+ while (csum > 0xffff)
	+ csum = (csum >> 16) + (csum & 0xffff);
	+ tcp_lro_assign_and_checksum_16(&pa->ip4->ip_sum, csum, &temp[1]);
	+ goto update_tcp_header;
	+
	+ case LRO_TYPE_IPV6_TCP:
	+ /* Compute new IPv6 length. */
	+ tlen = (pa->tcp->th_off << 2) + payload_len;
	+ tcp_lro_assign_and_checksum_16(&pa->ip6->ip6_plen, htons(tlen), &temp[0]);
	+ goto update_tcp_header;
	+
	+ case LRO_TYPE_IPV4_UDP:
	+ /* Compute new IPv4 length. */
	+ tlen = (pa->ip4->ip_hl << 2) + sizeof(*pa->udp) + payload_len;
	+ tcp_lro_assign_and_checksum_16(&pa->ip4->ip_len, htons(tlen), &temp[0]);
	+
	+ /* Subtract delta from current IPv4 checksum. */
	+ csum = pa->ip4->ip_sum + 0xffff - temp[0];
	+ while (csum > 0xffff)
	+ csum = (csum >> 16) + (csum & 0xffff);
	+ tcp_lro_assign_and_checksum_16(&pa->ip4->ip_sum, csum, &temp[1]);
	+ goto update_udp_header;
	+
	+ case LRO_TYPE_IPV6_UDP:
	+ /* Compute new IPv6 length. */
	+ tlen = sizeof(*pa->udp) + payload_len;
	+ tcp_lro_assign_and_checksum_16(&pa->ip6->ip6_plen, htons(tlen), &temp[0]);
	+ goto update_udp_header;
	+
	+ default:
	+ return (0);
	+ }
	+
	+update_tcp_header:
	+ /* Compute current TCP header checksum. */
	+ temp[2] = tcp_lro_rx_csum_tcphdr(pa->tcp);
	+
	+ /* Incorporate the latest ACK into the TCP header. */
	+ pa->tcp->th_ack = le->ack_seq;
	+ pa->tcp->th_win = le->window;
	+
	+ /* Incorporate latest timestamp into the TCP header. */
	+ if (le->timestamp != 0) {
	+ uint32_t *ts_ptr;
	+
	+ ts_ptr = (uint32_t *)(pa->tcp + 1);
	+ ts_ptr[1] = htonl(le->tsval);
	+ ts_ptr[2] = le->tsecr;
	+ }
	+
	+ /* Compute new TCP header checksum. */
	+ temp[3] = tcp_lro_rx_csum_tcphdr(pa->tcp);
	+
	+ /* Compute new TCP checksum. */
	+ csum = pa->tcp->th_sum + 0xffff - delta_sum +
	+ 0xffff - temp[0] + 0xffff - temp[3] + temp[2];
	+ while (csum > 0xffff)
	+ csum = (csum >> 16) + (csum & 0xffff);
	+
	+ /* Assign new TCP checksum. */
	+ tcp_lro_assign_and_checksum_16(&pa->tcp->th_sum, csum, &temp[4]);
	+
	+ /* Compute all modififications affecting next checksum. */
	+ csum = temp[0] + temp[1] + 0xffff - temp[2] +
	+ temp[3] + temp[4] + delta_sum;
	+ while (csum > 0xffff)
	+ csum = (csum >> 16) + (csum & 0xffff);
	+
	+ /* Return delta checksum to next stage, if any. */
	+ return (csum);
	+
	+update_udp_header:
	+ tlen = sizeof(*pa->udp) + payload_len;
	+ /* Assign new UDP length and compute checksum delta. */
	+ tcp_lro_assign_and_checksum_16(&pa->udp->uh_ulen, htons(tlen), &temp[2]);
	+
	+ /* Check if there is a UDP checksum. */
	+ if (__predict_false(pa->udp->uh_sum != 0)) {
	+ /* Compute new UDP checksum. */
	+ csum = pa->udp->uh_sum + 0xffff - delta_sum +
	+ 0xffff - temp[0] + 0xffff - temp[2];
	+ while (csum > 0xffff)
	+ csum = (csum >> 16) + (csum & 0xffff);
	+ /* Assign new UDP checksum. */
	+ tcp_lro_assign_and_checksum_16(&pa->udp->uh_sum, csum, &temp[3]);
	+ }

	- ip6 = le->le_ip6;
	- ip6->ip6_plen = p_len;
	- th = (struct tcphdr *)(ip6 + 1);
	+ /* Compute all modififications affecting next checksum. */
	+ csum = temp[0] + temp[1] + temp[2] + temp[3] + delta_sum;
	+ while (csum > 0xffff)
	+ csum = (csum >> 16) + (csum & 0xffff);
	+
	+ /* Return delta checksum to next stage, if any. */
	+ return (csum);
	+}
	+
	+static void
	+tcp_flush_out_entry(struct lro_ctrl lc, struct lro_entry le)
	+{
	+ /* Check if we need to recompute any checksums. */
	+ if (le->m_head->m_pkthdr.lro_nsegs > 1) {
	+ uint16_t csum;
	+
	+ switch (le->inner.data.lro_type) {
	+ case LRO_TYPE_IPV4_TCP:
	+ csum = tcp_lro_update_checksum(&le->inner, le,
	+ le->m_head->m_pkthdr.lro_tcp_d_len,
	+ le->m_head->m_pkthdr.lro_tcp_d_csum);
	+ csum = tcp_lro_update_checksum(&le->outer, NULL,
	+ le->m_head->m_pkthdr.lro_tcp_d_len +
	+ le->inner.total_hdr_len, csum);
	le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID \|
	- CSUM_PSEUDO_HDR;
	- le->p_len += ETHER_HDR_LEN + sizeof(*ip6);
	+ CSUM_PSEUDO_HDR \| CSUM_IP_CHECKED \| CSUM_IP_VALID;
	+ le->m_head->m_pkthdr.csum_data = 0xffff;
	break;
	- }
	-#endif
	-#ifdef INET
	- case ETHERTYPE_IP:
	- {
	- struct ip *ip4;
	- uint32_t cl;
	- uint16_t c;
	-
	- ip4 = le->le_ip4;
	- /* Fix IP header checksum for new length. */
	- c = ~ip4->ip_sum;
	- cl = c;
	- c = ~ip4->ip_len;
	- cl += c + p_len;
	- while (cl > 0xffff)
	- cl = (cl >> 16) + (cl & 0xffff);
	- c = cl;
	- ip4->ip_sum = ~c;
	- ip4->ip_len = p_len;
	- th = (struct tcphdr *)(ip4 + 1);
	+ case LRO_TYPE_IPV6_TCP:
	+ csum = tcp_lro_update_checksum(&le->inner, le,
	+ le->m_head->m_pkthdr.lro_tcp_d_len,
	+ le->m_head->m_pkthdr.lro_tcp_d_csum);
	+ csum = tcp_lro_update_checksum(&le->outer, NULL,
	+ le->m_head->m_pkthdr.lro_tcp_d_len +
	+ le->inner.total_hdr_len, csum);
	le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID \|
	- CSUM_PSEUDO_HDR \| CSUM_IP_CHECKED \| CSUM_IP_VALID;
	- le->p_len += ETHER_HDR_LEN;
	+ CSUM_PSEUDO_HDR;
	+ le->m_head->m_pkthdr.csum_data = 0xffff;
	+ break;
	+ case LRO_TYPE_NONE:
	+ switch (le->outer.data.lro_type) {
	+ case LRO_TYPE_IPV4_TCP:
	+ csum = tcp_lro_update_checksum(&le->outer, le,
	+ le->m_head->m_pkthdr.lro_tcp_d_len,
	+ le->m_head->m_pkthdr.lro_tcp_d_csum);
	+ le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID \|
	+ CSUM_PSEUDO_HDR \| CSUM_IP_CHECKED \| CSUM_IP_VALID;
	+ le->m_head->m_pkthdr.csum_data = 0xffff;
	+ break;
	+ case LRO_TYPE_IPV6_TCP:
	+ csum = tcp_lro_update_checksum(&le->outer, le,
	+ le->m_head->m_pkthdr.lro_tcp_d_len,
	+ le->m_head->m_pkthdr.lro_tcp_d_csum);
	+ le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID \|
	+ CSUM_PSEUDO_HDR;
	+ le->m_head->m_pkthdr.csum_data = 0xffff;
	+ break;
	+ default:
	+ break;
	+ }
	break;
	- }
	-#endif
	default:
	- th = NULL; /* Keep compiler happy. */
	- }
	- le->m_head->m_pkthdr.csum_data = 0xffff;
	- le->m_head->m_pkthdr.len = le->p_len;
	-
	- /* Incorporate the latest ACK into the TCP header. */
	- th->th_ack = le->ack_seq;
	- th->th_win = le->window;
	- /* Incorporate latest timestamp into the TCP header. */
	- if (le->timestamp != 0) {
	- uint32_t *ts_ptr;
	-
	- ts_ptr = (uint32_t *)(th + 1);
	- ts_ptr[1] = htonl(le->tsval);
	- ts_ptr[2] = le->tsecr;
	+ break;
	}
	- /* Update the TCP header checksum. */
	- le->ulp_csum += p_len;
	- le->ulp_csum += tcp_lro_csum_th(th);
	- while (le->ulp_csum > 0xffff)
	- le->ulp_csum = (le->ulp_csum >> 16) +
	- (le->ulp_csum & 0xffff);
	- th->th_sum = (le->ulp_csum & 0xffff);
	- th->th_sum = ~th->th_sum;
	}
	+
	/*
	* Break any chain, this is not set to NULL on the singleton
	* case m_nextpkt points to m_head. Other case set them
	* m_nextpkt to NULL in push_and_replace.
	*/
	le->m_head->m_nextpkt = NULL;
	- le->m_head->m_pkthdr.lro_nsegs = le->append_cnt;
	+ lc->lro_queued += le->m_head->m_pkthdr.lro_nsegs;
	(*lc->ifp->if_input)(lc->ifp, le->m_head);
	- lc->lro_queued += le->append_cnt;
	}

	static void
	-tcp_set_le_to_m(struct lro_ctrl lc, struct lro_entry le, struct mbuf *m)
	+tcp_set_entry_to_mbuf(struct lro_ctrl lc, struct lro_entry le,
	+ struct mbuf m, struct tcphdr th)
	{
	- struct ether_header *eh;
	- void l3hdr = NULL; / Keep compiler happy. */
	- struct tcphdr *th;
	-#ifdef INET6
	- struct ip6_hdr ip6 = NULL; / Keep compiler happy. */
	-#endif
	-#ifdef INET
	- struct ip ip4 = NULL; / Keep compiler happy. */
	-#endif
	uint32_t *ts_ptr;
	- int error, l, ts_failed = 0;
	uint16_t tcp_data_len;
	- uint16_t csum;
	+ uint16_t tcp_opt_len;

	- error = -1;
	- eh = mtod(m, struct ether_header *);
	- /*
	- * We must reset the other pointers since the mbuf
	- * we were pointing too is about to go away.
	- */
	- switch (le->eh_type) {
	-#ifdef INET6
	- case ETHERTYPE_IPV6:
	- l3hdr = ip6 = (struct ip6_hdr *)(eh + 1);
	- error = tcp_lro_rx_ipv6(lc, m, ip6, &th);
	- le->le_ip6 = ip6;
	- le->source_ip6 = ip6->ip6_src;
	- le->dest_ip6 = ip6->ip6_dst;
	- le->p_len = m->m_pkthdr.len - ETHER_HDR_LEN - sizeof(*ip6);
	- break;
	-#endif
	-#ifdef INET
	- case ETHERTYPE_IP:
	- l3hdr = ip4 = (struct ip *)(eh + 1);
	- error = tcp_lro_rx_ipv4(lc, m, ip4, &th);
	- le->le_ip4 = ip4;
	- le->source_ip4 = ip4->ip_src.s_addr;
	- le->dest_ip4 = ip4->ip_dst.s_addr;
	- le->p_len = m->m_pkthdr.len - ETHER_HDR_LEN;
	- break;
	-#endif
	- }
	- KASSERT(error == 0, ("%s: le=%p tcp_lro_rx_xxx failed\n",
	- __func__, le));
	ts_ptr = (uint32_t *)(th + 1);
	- l = (th->th_off << 2);
	- l -= sizeof(*th);
	- if (l != 0 &&
	- (__predict_false(l != TCPOLEN_TSTAMP_APPA) \|\|
	- (*ts_ptr != ntohl(TCPOPT_NOP<<24\|TCPOPT_NOP<<16\|
	- TCPOPT_TIMESTAMP<<8\|TCPOLEN_TIMESTAMP)))) {
	- /* We have failed to find a timestamp some other option? */
	- ts_failed = 1;
	- }
	- if ((l != 0) && (ts_failed == 0)) {
	+ tcp_opt_len = (th->th_off << 2);
	+ tcp_opt_len -= sizeof(*th);
	+
	+ /* Check if there is a timestamp option. */
	+ if (tcp_opt_len == 0 \|\|
	+ __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA \|\|
	+ *ts_ptr != TCP_LRO_TS_OPTION)) {
	+ /* We failed to find the timestamp option. */
	+ le->timestamp = 0;
	+ } else {
	le->timestamp = 1;
	le->tsval = ntohl(*(ts_ptr + 1));
	le->tsecr = *(ts_ptr + 2);
	- } else
	- le->timestamp = 0;
	- le->source_port = th->th_sport;
	- le->dest_port = th->th_dport;
	- /* Pull out the csum */
	- tcp_data_len = m->m_pkthdr.lro_len;
	+ }
	+
	+ tcp_data_len = m->m_pkthdr.lro_tcp_d_len;
	+
	+ /* Pull out TCP sequence numbers and window size. */
	le->next_seq = ntohl(th->th_seq) + tcp_data_len;
	le->ack_seq = th->th_ack;
	le->window = th->th_win;
	- csum = th->th_sum;
	- /* Setup the data pointers */
	+
	+ /* Setup new data pointers. */
	le->m_head = m;
	le->m_tail = m_last(m);
	- le->append_cnt = 0;
	- le->ulp_csum = tcp_lro_rx_csum_fixup(le, l3hdr, th, tcp_data_len,
	- ~csum);
	- le->append_cnt++;
	- th->th_sum = csum; /* Restore checksum on first packet. */
	}

	static void
	-tcp_push_and_replace(struct tcpcb tp, struct lro_ctrl lc, struct lro_entry le, struct mbuf m)
	+tcp_push_and_replace(struct lro_ctrl lc, struct lro_entry le, struct mbuf *m)
	{
	+ struct lro_parser *pa;
	+
	/*
	- * Push up the stack the current le and replace
	- * it with m.
	+ * Push up the stack of the current entry
	+ * and replace it with "m".
	*/
	struct mbuf *msave;

	/* Grab off the next and save it */
	msave = le->m_head->m_nextpkt;
	le->m_head->m_nextpkt = NULL;
	- /* Now push out the old le entry */
	- tcp_flush_out_le(tp, lc, le);
	+
	+ /* Now push out the old entry */
	+ tcp_flush_out_entry(lc, le);
	+
	+ /* Re-parse new header, should not fail. */
	+ pa = tcp_lro_parser(m, &le->outer, &le->inner, false);
	+ KASSERT(pa != NULL,
	+ ("tcp_push_and_replace: LRO parser failed on m=%p\n", m));
	+
	/*
	- * Now to replace the data properly in the le
	- * we have to reset the tcp header and
	+ * Now to replace the data properly in the entry
	+ * we have to reset the TCP header and
	* other fields.
	*/
	- tcp_set_le_to_m(lc, le, m);
	+ tcp_set_entry_to_mbuf(lc, le, m, pa->tcp);
	+
	/* Restore the next list */
	m->m_nextpkt = msave;
	}

	static void
	-tcp_lro_condense(struct tcpcb tp, struct lro_ctrl lc, struct lro_entry *le)
	+tcp_lro_mbuf_append_pkthdr(struct mbuf m, const struct mbuf p)
	+{
	+ uint32_t csum;
	+
	+ if (m->m_pkthdr.lro_nsegs == 1) {
	+ /* Compute relative checksum. */
	+ csum = p->m_pkthdr.lro_tcp_d_csum;
	+ } else {
	+ /* Merge TCP data checksums. */
	+ csum = (uint32_t)m->m_pkthdr.lro_tcp_d_csum +
	+ (uint32_t)p->m_pkthdr.lro_tcp_d_csum;
	+ while (csum > 0xffff)
	+ csum = (csum >> 16) + (csum & 0xffff);
	+ }
	+
	+ /* Update various counters. */
	+ m->m_pkthdr.len += p->m_pkthdr.lro_tcp_d_len;
	+ m->m_pkthdr.lro_tcp_d_csum = csum;
	+ m->m_pkthdr.lro_tcp_d_len += p->m_pkthdr.lro_tcp_d_len;
	+ m->m_pkthdr.lro_nsegs += p->m_pkthdr.lro_nsegs;
	+}
	+
	+static void
	+tcp_lro_condense(struct lro_ctrl lc, struct lro_entry le)
	{
	/*
	* Walk through the mbuf chain we
	@@ -724,8 +949,10 @@
	uint32_t *ts_ptr;
	struct mbuf *m;
	struct tcphdr *th;
	- uint16_t tcp_data_len, csum_upd;
	- int l;
	+ uint32_t tcp_data_len_total;
	+ uint32_t tcp_data_seg_total;
	+ uint16_t tcp_data_len;
	+ uint16_t tcp_opt_len;

	/*
	* First we must check the lead (m_head)
	@@ -734,30 +961,25 @@
	* right away (sack etc).
	*/
	again:
	-
	m = le->m_head->m_nextpkt;
	if (m == NULL) {
	- /* Just the one left */
	+ /* Just one left. */
	return;
	}
	- if (m->m_flags & M_ACKCMP)
	- panic("LRO condense lc:%p le:%p reaches with mbuf:%p ackcmp",
	- lc, le, m);
	- th = tcp_lro_get_th(le, le->m_head);
	- KASSERT(th != NULL,
	- ("le:%p m:%p th comes back NULL?", le, le->m_head));
	- l = (th->th_off << 2);
	- l -= sizeof(*th);
	+
	+ th = tcp_lro_get_th(m);
	+ tcp_opt_len = (th->th_off << 2);
	+ tcp_opt_len -= sizeof(*th);
	ts_ptr = (uint32_t *)(th + 1);
	- if (l != 0 && (__predict_false(l != TCPOLEN_TSTAMP_APPA) \|\|
	- (*ts_ptr != ntohl(TCPOPT_NOP<<24\|TCPOPT_NOP<<16\|
	- TCPOPT_TIMESTAMP<<8\|TCPOLEN_TIMESTAMP)))) {
	+
	+ if (tcp_opt_len != 0 && __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA \|\|
	+ *ts_ptr != TCP_LRO_TS_OPTION)) {
	/*
	* Its not the timestamp. We can't
	* use this guy as the head.
	*/
	le->m_head->m_nextpkt = m->m_nextpkt;
	- tcp_push_and_replace(tp, lc, le, m);
	+ tcp_push_and_replace(lc, le, m);
	goto again;
	}
	if ((th->th_flags & ~(TH_ACK \| TH_PUSH)) != 0) {
	@@ -766,9 +988,7 @@
	* before this segment, e.g. FIN.
	*/
	le->m_head->m_nextpkt = m->m_nextpkt;
	- KASSERT(((m->m_flags & M_LRO_EHDRSTRP) == 0) ,
	- ("tp:%p mbuf:%p has stripped ethernet flags:0x%x", tp, m, m->m_flags));
	- tcp_push_and_replace(tp, lc, le, m);
	+ tcp_push_and_replace(lc, le, m);
	goto again;
	}
	while((m = le->m_head->m_nextpkt) != NULL) {
	@@ -776,32 +996,26 @@
	* condense m into le, first
	* pull m out of the list.
	*/
	- KASSERT(((m->m_flags & M_LRO_EHDRSTRP) == 0) ,
	- ("tp:%p mbuf:%p has stripped ethernet flags:0x%x", tp, m, m->m_flags));
	- KASSERT(((m->m_flags & M_ACKCMP) == 0),
	- ("LRO condense lc:%p le:%p reaches with mbuf:%p ackcmp", lc, le, m));
	le->m_head->m_nextpkt = m->m_nextpkt;
	m->m_nextpkt = NULL;
	/* Setup my data */
	- tcp_data_len = m->m_pkthdr.lro_len;
	- th = tcp_lro_get_th(le, m);
	- KASSERT(th != NULL,
	- ("le:%p m:%p th comes back NULL?", le, m));
	+ tcp_data_len = m->m_pkthdr.lro_tcp_d_len;
	+ th = tcp_lro_get_th(m);
	ts_ptr = (uint32_t *)(th + 1);
	- l = (th->th_off << 2);
	- l -= sizeof(*th);
	- if (le->append_cnt >= lc->lro_ackcnt_lim) {
	- tcp_push_and_replace(tp, lc, le, m);
	- goto again;
	- }
	- if (le->p_len > (lc->lro_length_lim - tcp_data_len)) {
	+ tcp_opt_len = (th->th_off << 2);
	+ tcp_opt_len -= sizeof(*th);
	+ tcp_data_len_total = le->m_head->m_pkthdr.lro_tcp_d_len + tcp_data_len;
	+ tcp_data_seg_total = le->m_head->m_pkthdr.lro_nsegs + m->m_pkthdr.lro_nsegs;
	+
	+ if (tcp_data_seg_total >= lc->lro_ackcnt_lim \|\|
	+ tcp_data_len_total >= lc->lro_length_lim) {
	/* Flush now if appending will result in overflow. */
	- tcp_push_and_replace(tp, lc, le, m);
	+ tcp_push_and_replace(lc, le, m);
	goto again;
	}
	- if (l != 0 && (__predict_false(l != TCPOLEN_TSTAMP_APPA) \|\|
	- (*ts_ptr != ntohl(TCPOPT_NOP<<24\|TCPOPT_NOP<<16\|
	- TCPOPT_TIMESTAMP<<8\|TCPOLEN_TIMESTAMP)))) {
	+ if (tcp_opt_len != 0 &&
	+ __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA \|\|
	+ *ts_ptr != TCP_LRO_TS_OPTION)) {
	/*
	* Maybe a sack in the new one? We need to
	* start all over after flushing the
	@@ -809,18 +1023,18 @@
	* and flush it (calling the replace again possibly
	* or just returning).
	*/
	- tcp_push_and_replace(tp, lc, le, m);
	+ tcp_push_and_replace(lc, le, m);
	goto again;
	}
	if ((th->th_flags & ~(TH_ACK \| TH_PUSH)) != 0) {
	- tcp_push_and_replace(tp, lc, le, m);
	+ tcp_push_and_replace(lc, le, m);
	goto again;
	}
	- if (l != 0) {
	+ if (tcp_opt_len != 0) {
	uint32_t tsval = ntohl(*(ts_ptr + 1));
	/* Make sure timestamp values are increasing. */
	if (TSTMP_GT(le->tsval, tsval)) {
	- tcp_push_and_replace(tp, lc, le, m);
	+ tcp_push_and_replace(lc, le, m);
	goto again;
	}
	le->tsval = tsval;
	@@ -832,27 +1046,26 @@
	le->ack_seq == th->th_ack &&
	le->window == th->th_win))) {
	/* Out of order packet or duplicate ACK. */
	- tcp_push_and_replace(tp, lc, le, m);
	+ tcp_push_and_replace(lc, le, m);
	goto again;
	}
	- if (tcp_data_len \|\| SEQ_GT(ntohl(th->th_ack), ntohl(le->ack_seq))) {
	+ if (tcp_data_len != 0 \|\|
	+ SEQ_GT(ntohl(th->th_ack), ntohl(le->ack_seq))) {
	le->next_seq += tcp_data_len;
	le->ack_seq = th->th_ack;
	le->window = th->th_win;
	} else if (th->th_ack == le->ack_seq) {
	le->window = WIN_MAX(le->window, th->th_win);
	}
	- csum_upd = m->m_pkthdr.lro_csum;
	- le->ulp_csum += csum_upd;
	+
	if (tcp_data_len == 0) {
	- le->append_cnt++;
	- le->mbuf_cnt--;
	m_freem(m);
	continue;
	}
	- le->append_cnt++;
	- le->mbuf_appended++;
	- le->p_len += tcp_data_len;
	+
	+ /* Merge TCP data checksum and length to head mbuf. */
	+ tcp_lro_mbuf_append_pkthdr(le->m_head, m);
	+
	/*
	* Adjust the mbuf so that m_data points to the first byte of
	* the ULP payload. Adjust the mbuf to avoid complications and
	@@ -883,269 +1096,247 @@
	}

	static struct mbuf *
	-tcp_lro_get_last_if_ackcmp(struct lro_ctrl lc, struct lro_entry le, struct inpcb *inp)
	+tcp_lro_get_last_if_ackcmp(struct lro_ctrl lc, struct lro_entry le,
	+ struct inpcb inp, unsigned packs)
	{
	- struct mbuf *m = NULL;
	struct tcpcb *tp;
	+ struct mbuf *m;

	tp = intotcpcb(inp);
	- if (tp) {
	- /* Look at the last mbuf if any in queue */
	- if ((tp->t_tail_pkt) &&
	- (tp->t_tail_pkt->m_flags & M_ACKCMP)) {
	- if (M_TRAILINGSPACE(tp->t_tail_pkt) >= sizeof(struct tcp_ackent)) {
	- tcp_lro_log(tp, lc, le, NULL, 23, 0, 0, 0, 0);
	- m = tp->t_tail_pkt;
	- } else {
	- if ((inp->inp_flags2 & INP_MBUF_L_ACKS) == 0) {
	- counter_u64_add(tcp_would_have_but, 1);
	- inp->inp_flags2 \|= INP_MBUF_L_ACKS;
	- }
	- }
	+ if (__predict_false(tp == NULL))
	+ return (NULL);
	+
	+ /* Look at the last mbuf if any in queue */
	+ m = tp->t_tail_pkt;
	+ if (m != NULL && (m->m_flags & M_ACKCMP) != 0) {
	+ if (M_TRAILINGSPACE(m) >= sizeof(struct tcp_ackent)) {
	+ tcp_lro_log(tp, lc, le, NULL, 23, 0, 0, 0, 0);
	+ counter_u64_add(tcp_extra_mbuf, 1);
	+ return (m);
	}
	}
	+
	+ /* Decide mbuf size. */
	+ if ((packs sizeof(struct tcp_ackent)) > MLEN)
	+ m = m_getcl(M_NOWAIT, MT_DATA, M_ACKCMP \| M_PKTHDR);
	+ else
	+ m = m_gethdr(M_NOWAIT, MT_DATA);
	+
	+ if (__predict_false(m == NULL)) {
	+ counter_u64_add(tcp_would_have_but, 1);
	+ return (NULL);
	+ }
	+
	+ m->m_flags \|= M_ACKCMP;
	+
	+ /* Insert compressed ACK mbuf last. */
	+ tp->t_tail_pkt->m_nextpkt = m;
	+ tp->t_tail_pkt = m;
	+
	return (m);
	}

	static struct inpcb *
	-tcp_lro_lookup(struct lro_ctrl lc, struct lro_entry le)
	+tcp_lro_lookup(struct ifnet ifp, struct lro_parser pa)
	{
	- struct inpcb *inp = NULL;
	+ struct inpcb *inp;

	NET_EPOCH_ASSERT();
	- switch (le->eh_type) {
	+
	+ switch (pa->data.lro_type) {
	#ifdef INET6
	- case ETHERTYPE_IPV6:
	- inp = in6_pcblookup(&V_tcbinfo, &le->source_ip6,
	- le->source_port, &le->dest_ip6,le->dest_port,
	- INPLOOKUP_WLOCKPCB,
	- lc->ifp);
	+ case LRO_TYPE_IPV6_TCP:
	+ inp = in6_pcblookup(&V_tcbinfo,
	+ &pa->data.s_addr.v6,
	+ pa->data.s_port,
	+ &pa->data.d_addr.v6,
	+ pa->data.d_port,
	+ INPLOOKUP_WLOCKPCB,
	+ ifp);
	break;
	#endif
	#ifdef INET
	- case ETHERTYPE_IP:
	- inp = in_pcblookup(&V_tcbinfo, le->le_ip4->ip_src,
	- le->source_port, le->le_ip4->ip_dst, le->dest_port,
	- INPLOOKUP_WLOCKPCB,
	- lc->ifp);
	+ case LRO_TYPE_IPV4_TCP:
	+ inp = in_pcblookup(&V_tcbinfo,
	+ pa->data.s_addr.v4,
	+ pa->data.s_port,
	+ pa->data.d_addr.v4,
	+ pa->data.d_port,
	+ INPLOOKUP_WLOCKPCB,
	+ ifp);
	break;
	#endif
	+ default:
	+ inp = NULL;
	+ break;
	}
	return (inp);
	}

	-#endif
	-#ifdef NO
	-static void
	-stack_guard_prep(uint32_t *sg, int len)
	+static inline bool
	+tcp_lro_ack_valid(struct tcphdr th, uint32_t *ppts)
	{
	- int i;
	-
	- for (i = 0; i < len; i++) {
	- sg[i] = 0xdeadc0de;
	+ /* For ACKCMP we only accept ACK, PUSH, ECE and CWR. */
	+ if ((th->th_flags & ~(TH_ACK \| TH_PUSH \| TH_ECE \| TH_CWR)) != 0)
	+ return (false);
	+
	+ /* ACK flag must be set. */
	+ if (!(th->th_flags & TH_ACK))
	+ return (false);
	+
	+ switch (th->th_off << 2) {
	+ case (sizeof(*th) + TCPOLEN_TSTAMP_APPA):
	+ ppts = (uint32_t )(th + 1);
	+ /* Check if we have only one timestamp option. */
	+ return (**ppts == TCP_LRO_TS_OPTION);
	+ case (sizeof(*th)):
	+ /* No options. */
	+ *ppts = NULL;
	+ return (true);
	+ default:
	+ return (false);
	}
	}

	-static void
	-stack_guard_check(struct lro_ctrl lc, struct lro_entry le, uint32_t *sg, int len)
	+static unsigned
	+tcp_lro_count_acks(struct mbuf *mb)
	{
	- int i;
	+ uint32_t *ts_ptr;
	+ unsigned count;

	- for (i = 0; i < len; i++) {
	- if (sg[i] != 0xdeadc0de)
	- panic("Stack guard fails sg[%d] = 0x%x le:%p lc:%p sg:%p\n",
	- i, sg[i], le, lc, sg);
	+ for (count = 0; mb != NULL; mb = mb->m_nextpkt) {
	+ if (mb->m_pkthdr.lro_tcp_d_len != 0)
	+ continue;
	+ count += tcp_lro_ack_valid(tcp_lro_get_th(mb), &ts_ptr);
	}
	+ return (count);
	}
	-#endif

	-void
	-tcp_lro_flush(struct lro_ctrl lc, struct lro_entry le)
	+static int
	+tcp_lro_flush_tcphpts(struct lro_ctrl lc, struct lro_entry le)
	{
	- struct tcpcb *tp = NULL;
	-#ifdef TCPHPTS
	- struct inpcb *inp = NULL;
	- int need_wakeup = 0, can_queue = 0;
	+ struct inpcb *inp;
	+ struct tcpcb *tp;
	+ struct mbuf **pp;
	+ unsigned nacks;
	+ bool bpf_req;
	+
	+ /* Check if packet doesn't belongs to our network interface. */
	+ if ((tcplro_stacks_wanting_mbufq == 0) \|\|
	+ (le->outer.data.vlan_id != 0) \|\|
	+ (le->inner.data.lro_type != LRO_TYPE_NONE))
	+ return (TCP_LRO_CANNOT);

	- /* Now lets lookup the inp first */
	- CURVNET_SET(lc->ifp->if_vnet);
	/*
	- * XXXRRS Currently the common input handler for
	- * mbuf queuing cannot handle VLAN Tagged. This needs
	- * to be fixed and the or condition removed (i.e. the
	- * common code should do the right lookup for the vlan
	- * tag and anything else that the vlan_input() does).
	+ * Be proactive about unspecified IPv6 address in source. As
	+ * we use all-zero to indicate unbounded/unconnected pcb,
	+ * unspecified IPv6 address can be used to confuse us.
	+ *
	+ * Note that packets with unspecified IPv6 destination is
	+ * already dropped in ip6_input.
	*/
	- if (le->m_head == NULL) {
	- /*
	- * Everything was pushed up to the stack nothing to do
	- * but release the reference and be done.
	- */
	- if (le->inp) {
	- INP_WLOCK(le->inp);
	- if (in_pcbrele_wlocked(le->inp) == 0) {
	- /*
	- * We released it and still
	- * have the lock.
	- */
	- INP_WUNLOCK(le->inp);
	- }
	- }
	- goto done;
	+ if (__predict_false(le->outer.data.lro_type == LRO_TYPE_IPV6_TCP &&
	+ IN6_IS_ADDR_UNSPECIFIED(&le->outer.data.s_addr.v6)))
	+ return (TCP_LRO_CANNOT);
	+
	+ if (__predict_false(le->inner.data.lro_type == LRO_TYPE_IPV6_TCP &&
	+ IN6_IS_ADDR_UNSPECIFIED(&le->inner.data.s_addr.v6)))
	+ return (TCP_LRO_CANNOT);
	+
	+ /* Lookup inp, if any. */
	+ inp = tcp_lro_lookup(lc->ifp,
	+ (le->inner.data.lro_type == LRO_TYPE_NONE) ? &le->outer : &le->inner);
	+ if (inp == NULL)
	+ return (TCP_LRO_CANNOT);
	+
	+ counter_u64_add(tcp_inp_lro_locks_taken, 1);
	+
	+ /* Get TCP control structure. */
	+ tp = intotcpcb(inp);
	+
	+ /* Check if the inp is dead, Jim. */
	+ if (tp == NULL \|\|
	+ (inp->inp_flags & (INP_DROPPED \| INP_TIMEWAIT)) \|\|
	+ (inp->inp_flags2 & INP_FREED)) {
	+ INP_WUNLOCK(inp);
	+ return (TCP_LRO_CANNOT);
	}
	- if ((tcplro_stacks_wanting_mbufq == 0) \|\| (le->m_head->m_flags & M_VLANTAG))
	- goto skip_lookup;

	- if (le->inp == NULL) {
	- le->inp = inp = tcp_lro_lookup(lc, le);
	- if (inp && ((inp->inp_flags & (INP_DROPPED\|INP_TIMEWAIT)) \|\|
	- (inp->inp_flags2 & INP_FREED))) {
	- /*
	- * We can't present these to the inp since
	- * it will not support the stripped ethernet
	- * header that these have nor if a compressed
	- * ack is presnet.
	- */
	- INP_WUNLOCK(inp);
	- lro_free_mbuf_chain(le->m_head);
	- goto done;
	- }
	- if ((le->flags & HAS_COMP_ENTRIES) &&
	- ((inp->inp_flags2 & INP_MBUF_ACKCMP) == 0)) {
	- /*
	- * It swapped to off, must be a stack
	- * switch. We need to ditch all the packets
	- * and the peer will just have to retransmit.
	- */
	- INP_WUNLOCK(inp);
	- lro_free_mbuf_chain(le->m_head);
	- goto done;
	- }
	+ /* Check if the transport doesn't support the needed optimizations. */
	+ if ((inp->inp_flags2 & (INP_SUPPORTS_MBUFQ \| INP_MBUF_ACKCMP)) == 0) {
	+ INP_WUNLOCK(inp);
	+ return (TCP_LRO_CANNOT);
	+ }
	+
	+ /* Check if packets should be tapped to BPF. */
	+ bpf_req = bpf_peers_present(lc->ifp->if_bpf);
	+
	+ /* Check if the "inp" supports ACK compression. */
	+ if (inp->inp_flags2 & INP_MBUF_ACKCMP) {
	+ /* Count number of ACKs in there. */
	+ nacks = tcp_lro_count_acks(le->m_head);
	} else {
	- /* We have a reference on the inp lets lock and release it */
	- inp = le->inp;
	- INP_WLOCK(inp);
	- if (in_pcbrele_wlocked(inp)) {
	- /*
	- * We lost the inp. We can't present these to the inp since
	- * it will not support the stripped off etherent header.
	- */
	- lro_free_mbuf_chain(le->m_head);
	- goto done;
	- }
	- if (inp && ((inp->inp_flags & (INP_DROPPED\|INP_TIMEWAIT)) \|\|
	- (inp->inp_flags2 & INP_FREED))) {
	- /*
	- * We can't present these to the inp since
	- * it may not support them.
	- */
	- INP_WUNLOCK(inp);
	- lro_free_mbuf_chain(le->m_head);
	- goto done;
	- }
	- if ((le->flags & HAS_COMP_ENTRIES) &&
	- ((inp->inp_flags2 & INP_MBUF_ACKCMP) == 0)) {
	- /*
	- * It swapped to off, must be a stack
	- * switch. We need to ditch all the packets
	- * and the peer will just have to retransmit.
	- */
	- INP_WUNLOCK(inp);
	- lro_free_mbuf_chain(le->m_head);
	- goto done;
	- }
	+ nacks = 0;
	}
	- if (inp && ((inp->inp_flags2 & INP_SUPPORTS_MBUFQ) \|\|
	- (inp->inp_flags2 & INP_MBUF_ACKCMP))) {
	- /* The transport supports mbuf queuing */
	- can_queue = 1;
	- if (le->need_wakeup \|\|
	- ((inp->inp_in_input == 0) &&
	- ((inp->inp_flags2 & INP_MBUF_QUEUE_READY) == 0))) {
	- /*
	- * Either the transport is off on a keep-alive
	- * (it has the queue_ready flag clear and its
	- * not already been woken) or the entry has
	- * some urgent thing (FIN or possibly SACK blocks).
	- * This means we need to wake the transport up by
	- * putting it on the input pacer.
	- */
	- need_wakeup = 1;
	- if ((inp->inp_flags2 & INP_DONT_SACK_QUEUE) &&
	- (le->need_wakeup != 1)) {
	- /*
	- * Prohibited from a sack wakeup.
	- */
	- need_wakeup = 0;
	- }
	+
	+ /* Strip and compress all the incoming packets. */
	+ for (pp = &le->m_head; *pp != NULL; ) {
	+ if (do_bpf_strip_and_compress(inp, lc, le, pp, &nacks, bpf_req) == false) {
	+ /* Advance to next mbuf. */
	+ pp = &(*pp)->m_nextpkt;
	}
	- /* Do we need to be awoken due to lots of data or acks? */
	- if ((le->tcp_tot_p_len >= lc->lro_length_lim) \|\|
	- (le->mbuf_cnt >= lc->lro_ackcnt_lim))
	- need_wakeup = 1;
	}
	- if (inp)
	- tp = intotcpcb(inp);
	+
	+ /* Update "m_last_mbuf", if any. */
	+ if (pp == &le->m_head)
	+ le->m_last_mbuf = *pp;
	else
	- tp = NULL;
	- if (can_queue) {
	+ le->m_last_mbuf = __containerof(pp, struct mbuf, m_nextpkt);
	+
	+ /* Check if any data mbufs left. */
	+ if (le->m_head != NULL) {
	counter_u64_add(tcp_inp_lro_direct_queue, 1);
	- tcp_lro_log(tp, lc, le, NULL, 22, need_wakeup,
	- inp->inp_flags2, inp->inp_in_input, le->need_wakeup);
	+ tcp_lro_log(tp, lc, le, NULL, 22, 1,
	+ inp->inp_flags2, inp->inp_in_input, 1);
	tcp_queue_pkts(tp, le);
	- if (need_wakeup) {
	- /*
	- * We must get the guy to wakeup via
	- * hpts.
	- */
	- NET_EPOCH_ASSERT();
	- if (le->need_wakeup == 2) {
	- /*
	- * The value 2 is set if the
	- * options are unrecognized i.e.
	- * not just a timestamp. So really
	- * sack is usually what it is but
	- * it might be some other option (CWR
	- * etc).
	- */
	- counter_u64_add(tcp_inp_lro_sack_wake, 1);
	- }
	- counter_u64_add(tcp_inp_lro_wokeup_queue, 1);
	- if ((*tp->t_fb->tfb_do_queued_segments)(inp->inp_socket, tp, 0)) {
	- inp = NULL;
	- }
	- }
	}
	- if (inp) {
	- /* Unlock it */
	- tp = NULL;
	- counter_u64_add(tcp_inp_lro_locks_taken, 1);
	+
	+ /* Wakeup */
	+ counter_u64_add(tcp_inp_lro_wokeup_queue, 1);
	+ if ((*tp->t_fb->tfb_do_queued_segments)(inp->inp_socket, tp, 0))
	+ inp = NULL;
	+
	+ if (inp != NULL)
	INP_WUNLOCK(inp);
	- }
	- if (can_queue == 0) {
	-skip_lookup:
	- if (le->strip_cnt) {
	- /*
	- * We have stripped mbufs, the connection
	- * must have changed underneath us. You
	- * loose the packets as a penalty.
	- */
	- lro_free_mbuf_chain(le->m_head);
	- goto done;
	- }
	-#endif /* TCPHPTS */
	- /* Old fashioned lro method */
	- if (le->m_head != le->m_last_mbuf) {
	+ return (0); /* Success. */
	+}
	+#endif
	+
	+void
	+tcp_lro_flush(struct lro_ctrl lc, struct lro_entry le)
	+{
	+ /* Only optimise if there are multiple packets waiting. */
	+ if (le->m_head != le->m_last_mbuf) {
	+#ifdef TCPHPTS
	+ int error;
	+
	+ CURVNET_SET(lc->ifp->if_vnet);
	+ error = tcp_lro_flush_tcphpts(lc, le);
	+ CURVNET_RESTORE();
	+
	+ if (error != 0) {
	+#endif
	counter_u64_add(tcp_inp_lro_compressed, 1);
	- tcp_lro_condense(tp, lc, le);
	- } else
	- counter_u64_add(tcp_inp_lro_single_push, 1);
	- tcp_flush_out_le(tp, lc, le);
	+ tcp_lro_condense(lc, le);
	+ tcp_flush_out_entry(lc, le);
	#ifdef TCPHPTS
	- }
	-done:
	- CURVNET_RESTORE();
	+ }
	#endif
	+ } else {
	+ counter_u64_add(tcp_inp_lro_single_push, 1);
	+ tcp_flush_out_entry(lc, le);
	+ }
	+
	lc->lro_flushed++;
	bzero(le, sizeof(*le));
	LIST_INSERT_HEAD(&lc->lro_free, le, next);
	@@ -1246,11 +1437,15 @@
	uint64_t nseq;
	unsigned x;

	+ CURVNET_SET(lc->ifp->if_vnet);
	+
	/* check if no mbufs to flush */
	if (lc->lro_mbuf_count == 0)
	goto done;

	- microuptime(&lc->lro_last_flush);
	+ /* get current time */
	+ lc->lro_last_queue_time = getsbinuptime();
	+
	/* sort all mbufs according to stream */
	tcp_lro_sort(lc->lro_mbuf_data, lc->lro_mbuf_count);

	@@ -1274,7 +1469,7 @@
	}

	/* add packet to LRO engine */
	- if (tcp_lro_rx2(lc, mb, 0, 0) != 0) {
	+ if (tcp_lro_rx_common(lc, mb, 0, false) != 0) {
	/* input packet to network layer */
	(*lc->ifp->if_input)(lc->ifp, mb);
	lc->lro_queued++;
	@@ -1286,28 +1481,19 @@
	tcp_lro_rx_done(lc);

	lc->lro_mbuf_count = 0;
	-}

	-static void
	-lro_set_mtime(struct timeval tv, struct timespec ts)
	-{
	- tv->tv_sec = ts->tv_sec;
	- tv->tv_usec = ts->tv_nsec / 1000;
	+ CURVNET_RESTORE();
	}

	#ifdef TCPHPTS
	static void
	-build_ack_entry(struct tcp_ackent ae, struct tcphdr th, struct mbuf *m, uint16_t hdr_len, uint16_t iptos)
	+build_ack_entry(struct tcp_ackent ae, struct tcphdr th, struct mbuf *m,
	+ uint32_t *ts_ptr, uint16_t iptos)
	{
	/*
	- * Given a TCP ack, summarize it down into the small tcp
	- * ack entry.
	+ * Given a TCP ACK, summarize it down into the small TCP ACK
	+ * entry.
	*/
	- u_char *cp;
	-
	-
	- KASSERT(((th->th_flags & ~(TH_ACK \| TH_PUSH \| TH_CWR \| TH_ECE)) == 0),
	- ("tcphdr:%p mbuf:%p has unallowed bits %x", th, m, th->th_flags));
	ae->timestamp = m->m_pkthdr.rcv_tstmp;
	if (m->m_flags & M_TSTMP_LRO)
	ae->flags = TSTMP_LRO;
	@@ -1316,572 +1502,258 @@
	ae->seq = ntohl(th->th_seq);
	ae->ack = ntohl(th->th_ack);
	ae->flags \|= th->th_flags;
	- if (hdr_len) {
	- /* We have a timestamp options get out the bits */
	- cp = (u_char *)(th + 1);
	- /* Skip the two NOP's at the front */
	- while (*cp == TCPOPT_NOP)
	- cp++;
	- KASSERT(((*cp == TCPOPT_TIMESTAMP) &&
	- (cp[1] == TCPOLEN_TIMESTAMP)),
	- ("At %p in tcphdr:%p options of %d not timestamp",
	- cp, th, hdr_len));
	- bcopy((char *)cp + 2,
	- (char *)&ae->ts_value, sizeof(uint32_t));
	- ae->ts_value = ntohl(ae->ts_value);
	- bcopy((char *)cp + 6,
	- (char *)&ae->ts_echo, sizeof(uint32_t));
	- ae->ts_echo = ntohl(ae->ts_echo);
	+ if (ts_ptr != NULL) {
	+ ae->ts_value = ntohl(ts_ptr[1]);
	+ ae->ts_echo = ntohl(ts_ptr[2]);
	ae->flags \|= HAS_TSTMP;
	}
	ae->win = ntohs(th->th_win);
	ae->codepoint = iptos;
	}

	-static struct mbuf *
	-do_bpf_and_csum(struct inpcb inp, struct lro_ctrl lc, struct lro_entry *le,
	- struct ether_header eh, struct mbuf m, int bpf_req, int locked)
	+/*
	+ * Do BPF tap for either ACK_CMP packets or MBUF QUEUE type packets
	+ * and strip all but the IPv4/IPv6 header.
	+ */
	+static bool
	+do_bpf_strip_and_compress(struct inpcb inp, struct lro_ctrl lc,
	+ struct lro_entry le, struct mbuf pp, unsigned packs, bool bpf_req)
	{
	- /*
	- * Do TCP/IP checksum and BPF tap for either ACK_CMP packets or
	- * MBUF QUEUE type packets.
	- */
	+ union {
	+ void *ptr;
	+ struct ip *ip4;
	+ struct ip6_hdr *ip6;
	+ } l3;
	+ struct mbuf *m;
	+ struct mbuf *nm;
	struct tcphdr *th;
	-#ifdef INET6
	- struct ip6_hdr ip6 = NULL; / Keep compiler happy. */
	-#endif
	-#ifdef INET
	- struct ip ip = NULL; / Keep compiler happy. */
	-#endif
	+ struct tcp_ackent *ack_ent;
	+ uint32_t *ts_ptr;
	+ uint16_t lro_type;
	+ uint16_t iptos;
	+ int tcp_hdr_offset;
	+ int idx;

	- uint16_t drop_hdrlen;
	- int etype, tlen;
	-#ifdef INET
	- uint8_t iptos;
	-#endif
	+ /* Get current mbuf. */
	+ m = *pp;

	/* Let the BPF see the packet */
	- if (bpf_req && lc->ifp)
	+ if (__predict_false(bpf_req))
	ETHER_BPF_MTAP(lc->ifp, m);
	- /* Get type and Trim off the ethernet header */
	- m->m_pkthdr.lro_etype = etype = ntohs(eh->ether_type);
	- m_adj(m, sizeof(*eh));
	- m->m_flags \|= M_LRO_EHDRSTRP;
	- switch (etype) {
	-#ifdef INET6
	- case ETHERTYPE_IPV6:
	- {
	- if (m->m_len < (sizeof(ip6) + sizeof(th))) {
	- m = m_pullup(m, sizeof(ip6) + sizeof(th));
	- if (m == NULL) {
	- TCPSTAT_INC(tcps_rcvshort);
	- m_freem(m);
	- return (NULL);
	- }
	- }
	- ip6 = (struct ip6_hdr *)(eh + 1);
	- th = (struct tcphdr *)(ip6 + 1);
	- tlen = ntohs(ip6->ip6_plen);
	- drop_hdrlen = sizeof(*ip6);
	- if (m->m_pkthdr.csum_flags & CSUM_DATA_VALID_IPV6) {
	- if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR) {
	- counter_u64_add(tcp_csum_hardware_w_ph, 1);
	- th->th_sum = m->m_pkthdr.csum_data;
	- } else {
	- counter_u64_add(tcp_csum_hardware, 1);
	- th->th_sum = in6_cksum_pseudo(ip6, tlen,
	- IPPROTO_TCP, m->m_pkthdr.csum_data);
	- }
	- th->th_sum ^= 0xffff;
	- } else {
	- counter_u64_add(tcp_csum_software, 1);
	- th->th_sum = in6_cksum(m, IPPROTO_TCP, drop_hdrlen, tlen);
	- }
	- if (th->th_sum) {
	- TCPSTAT_INC(tcps_rcvbadsum);
	- if (locked) {
	- /* Log the bad news */
	- struct tcpcb *tp = intotcpcb(inp);

	- tcp_lro_log(tp, lc, le, m, 13, tlen, m->m_pkthdr.csum_flags, drop_hdrlen, th->th_sum);
	- }
	- m_freem(m);
	- return (NULL);
	+ tcp_hdr_offset = m->m_pkthdr.lro_tcp_h_off;
	+
	+ lro_type = le->inner.data.lro_type;
	+ switch (lro_type) {
	+ case LRO_TYPE_NONE:
	+ lro_type = le->outer.data.lro_type;
	+ switch (lro_type) {
	+ case LRO_TYPE_IPV4_TCP:
	+ tcp_hdr_offset -= sizeof(*le->outer.ip4);
	+ break;
	+ case LRO_TYPE_IPV6_TCP:
	+ tcp_hdr_offset -= sizeof(*le->outer.ip6);
	+ break;
	+ default:
	+ goto compressed;
	}
	+ break;
	+ case LRO_TYPE_IPV4_TCP:
	+ tcp_hdr_offset -= sizeof(*le->outer.ip4);
	+ break;
	+ case LRO_TYPE_IPV6_TCP:
	+ tcp_hdr_offset -= sizeof(*le->outer.ip6);
	+ break;
	+ default:
	+ goto compressed;
	+ }

	+ MPASS(tcp_hdr_offset >= 0);

	+ m_adj(m, tcp_hdr_offset);
	+ m->m_flags \|= M_LRO_EHDRSTRP;
	+ m->m_flags &= ~M_ACKCMP;
	+ m->m_pkthdr.lro_tcp_h_off -= tcp_hdr_offset;

	+ th = tcp_lro_get_th(m);

	- /*
	- * Be proactive about unspecified IPv6 address in source.
	- * As we use all-zero to indicate unbounded/unconnected pcb,
	- * unspecified IPv6 address can be used to confuse us.
	- *
	- * Note that packets with unspecified IPv6 destination is
	- * already dropped in ip6_input.
	- */
	- if (IN6_IS_ADDR_UNSPECIFIED(&ip6->ip6_src)) {
	- /* XXX stat */
	- m_freem(m);
	- return (NULL);
	- }
	+ th->th_sum = 0; /* TCP checksum is valid. */
	+
	+ /* No more ACKs, just return. */
	+ if (*packs == 0)
	+ goto done;
	+
	+ /* Check if there is TCP data. */
	+ if (m->m_pkthdr.lro_tcp_d_len != 0)
	+ goto done;
	+
	+ /* Check if ACK can be compressed */
	+ if (tcp_lro_ack_valid(th, &ts_ptr) == false)
	+ goto done;
	+
	+ l3.ptr = mtod(m, void *);
	+
	+ switch (lro_type) {
	+ case LRO_TYPE_IPV4_TCP:
	+ iptos = l3.ip4->ip_tos;
	+ break;
	+ case LRO_TYPE_IPV6_TCP:
	+ iptos = IPV6_TRAFFIC_CLASS(l3.ip6);
	+ break;
	+ default:
	+ iptos = 0; /* Keep compiler happy. */
	break;
	}
	+
	+ nm = tcp_lro_get_last_if_ackcmp(lc, le, inp, packs);
	+ (packs)--; / We've consumed an ACK. */
	+
	+ if (__predict_false(nm == NULL))
	+ goto done;
	+
	+ MPASS(M_TRAILINGSPACE(nm) >= sizeof(struct tcp_ackent));
	+
	+ /* We can add in to the one on the tail */
	+ ack_ent = mtod(nm, struct tcp_ackent *);
	+ idx = (nm->m_len / sizeof(struct tcp_ackent));
	+ build_ack_entry(&ack_ent[idx], th, m, ts_ptr, iptos);
	+
	+ /* Bump the size of both pkt-hdr and len */
	+ nm->m_len += sizeof(struct tcp_ackent);
	+ nm->m_pkthdr.len += sizeof(struct tcp_ackent);
	+
	+ counter_u64_add(tcp_comp_total, 1);
	+compressed:
	+ /* Advance to next mbuf before freeing. */
	+ *pp = m->m_nextpkt;
	+ m_freem(m);
	+ return (true);
	+done:
	+ counter_u64_add(tcp_uncomp_total, 1);
	+ return (false);
	+}
	#endif
	-#ifdef INET
	- case ETHERTYPE_IP:
	- {
	- if (m->m_len < sizeof (struct tcpiphdr)) {
	- if ((m = m_pullup(m, sizeof (struct tcpiphdr)))
	- == NULL) {
	- TCPSTAT_INC(tcps_rcvshort);
	- m_freem(m);
	- return (NULL);
	- }
	- }
	- ip = (struct ip *)(eh + 1);
	- th = (struct tcphdr *)(ip + 1);
	- iptos = ip->ip_tos;
	- drop_hdrlen = sizeof(*ip);
	- tlen = ntohs(ip->ip_len) - sizeof(struct ip);
	- if (m->m_pkthdr.csum_flags & CSUM_DATA_VALID) {
	- if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR) {
	- counter_u64_add(tcp_csum_hardware_w_ph, 1);
	- th->th_sum = m->m_pkthdr.csum_data;
	- } else {
	- counter_u64_add(tcp_csum_hardware, 1);
	- th->th_sum = in_pseudo(ip->ip_src.s_addr,
	- ip->ip_dst.s_addr,
	- htonl(m->m_pkthdr.csum_data + tlen +
	- IPPROTO_TCP));
	- }
	- th->th_sum ^= 0xffff;
	- } else {
	- int len;
	- struct ipovly ipov = (struct ipovly )ip;
	- /*
	- * Checksum extended TCP header and data.
	- */
	- counter_u64_add(tcp_csum_software, 1);
	- len = drop_hdrlen + tlen;
	- bzero(ipov->ih_x1, sizeof(ipov->ih_x1));
	- ipov->ih_len = htons(tlen);
	- th->th_sum = in_cksum(m, len);
	- /* Reset length for SDT probes. */
	- ip->ip_len = htons(len);
	- /* Reset TOS bits */
	- ip->ip_tos = iptos;
	- /* Re-initialization for later version check */
	- ip->ip_v = IPVERSION;
	- ip->ip_hl = sizeof(*ip) >> 2;
	- }
	- if (th->th_sum) {
	- TCPSTAT_INC(tcps_rcvbadsum);
	- if (locked) {
	- /* Log the bad news */
	- struct tcpcb *tp = intotcpcb(inp);

	- tcp_lro_log(tp, lc, le, m, 13, tlen, m->m_pkthdr.csum_flags, drop_hdrlen, th->th_sum);
	- }
	- m_freem(m);
	- return (NULL);
	- }
	- break;
	+static struct lro_head *
	+tcp_lro_rx_get_bucket(struct lro_ctrl lc, struct mbuf m, struct lro_parser *parser)
	+{
	+ u_long hash;
	+
	+ if (M_HASHTYPE_ISHASH(m)) {
	+ hash = m->m_pkthdr.flowid;
	+ } else {
	+ for (unsigned i = hash = 0; i != LRO_RAW_ADDRESS_MAX; i++)
	+ hash += parser->data.raw[i];
	}
	-#endif
	- } /* end switch */
	- return (m);
	+ return (&lc->lro_hash[hash % lc->lro_hashsz]);
	}
	-#endif

	static int
	-tcp_lro_rx2(struct lro_ctrl lc, struct mbuf m, uint32_t csum, int use_hash)
	+tcp_lro_rx_common(struct lro_ctrl lc, struct mbuf m, uint32_t csum, bool use_hash)
	{
	+ struct lro_parser pi; /* inner address data */
	+ struct lro_parser po; /* outer address data */
	+ struct lro_parser pa; / current parser for TCP stream */
	struct lro_entry *le;
	- struct ether_header *eh;
	-#ifdef INET6
	- struct ip6_hdr ip6 = NULL; / Keep compiler happy. */
	-#endif
	-#ifdef INET
	- struct ip ip4 = NULL; / Keep compiler happy. */
	-#endif
	- struct tcphdr *th;
	- void l3hdr = NULL; / Keep compiler happy. */
	- uint32_t *ts_ptr;
	- tcp_seq seq;
	- int error, ip_len, hdr_len, locked = 0;
	- uint16_t eh_type, tcp_data_len, need_flush;
	-#ifdef TCPHPTS
	- uint16_t iptos;
	-#endif
	struct lro_head *bucket;
	- struct timespec arrv;
	+ struct tcphdr *th;
	+ int tcp_data_len;
	+ int tcp_opt_len;
	+ int error;
	+ uint16_t tcp_data_sum;

	- /* Clear the flags we may use to communicate with TCP */
	- m->m_flags &= ~(M_ACKCMP\|M_LRO_EHDRSTRP);
	+ /* Quickly decide if packet cannot be LRO'ed */
	+ if (__predict_false(V_ip6_forwarding != 0))
	+ return (TCP_LRO_CANNOT);

	/* We expect a contiguous header [eh, ip, tcp]. */
	- if ((m->m_flags & (M_TSTMP_LRO\|M_TSTMP)) == 0) {
	- /* If no hardware or arrival stamp on the packet add arrival */
	- nanouptime(&arrv);
	- m->m_pkthdr.rcv_tstmp = (arrv.tv_sec * 1000000000) + arrv.tv_nsec;
	- m->m_flags \|= M_TSTMP_LRO;
	- }
	- eh = mtod(m, struct ether_header *);
	- eh_type = ntohs(eh->ether_type);
	- switch (eh_type) {
	-#ifdef INET6
	- case ETHERTYPE_IPV6:
	- {
	- CURVNET_SET(lc->ifp->if_vnet);
	- if (V_ip6_forwarding != 0) {
	- /* XXX-BZ stats but changing lro_ctrl is a problem. */
	- CURVNET_RESTORE();
	- return (TCP_LRO_CANNOT);
	- }
	- CURVNET_RESTORE();
	- l3hdr = ip6 = (struct ip6_hdr *)(eh + 1);
	- error = tcp_lro_rx_ipv6(lc, m, ip6, &th);
	- if (error != 0)
	- return (error);
	- tcp_data_len = ntohs(ip6->ip6_plen);
	-#ifdef TCPHPTS
	- iptos = IPV6_TRAFFIC_CLASS(ip6);
	-#endif
	- ip_len = sizeof(*ip6) + tcp_data_len;
	- break;
	- }
	-#endif
	+ pa = tcp_lro_parser(m, &po, &pi, true);
	+ if (__predict_false(pa == NULL))
	+ return (TCP_LRO_NOT_SUPPORTED);
	+
	+ /* We don't expect any padding. */
	+ error = tcp_lro_trim_mbuf_chain(m, pa);
	+ if (__predict_false(error != 0))
	+ return (error);
	+
	#ifdef INET
	- case ETHERTYPE_IP:
	- {
	- CURVNET_SET(lc->ifp->if_vnet);
	- if (V_ipforwarding != 0) {
	- /* XXX-BZ stats but changing lro_ctrl is a problem. */
	- CURVNET_RESTORE();
	- return (TCP_LRO_CANNOT);
	- }
	- CURVNET_RESTORE();
	- l3hdr = ip4 = (struct ip *)(eh + 1);
	- error = tcp_lro_rx_ipv4(lc, m, ip4, &th);
	- if (error != 0)
	+ switch (pa->data.lro_type) {
	+ case LRO_TYPE_IPV4_TCP:
	+ error = tcp_lro_rx_ipv4(lc, m, pa->ip4);
	+ if (__predict_false(error != 0))
	return (error);
	- ip_len = ntohs(ip4->ip_len);
	-#ifdef TCPHPTS
	- iptos = ip4->ip_tos;
	-#endif
	- tcp_data_len = ip_len - sizeof(*ip4);
	+ break;
	+ default:
	break;
	}
	#endif
	- /* XXX-BZ what happens in case of VLAN(s)? */
	- default:
	- return (TCP_LRO_NOT_SUPPORTED);
	+ /* If no hardware or arrival stamp on the packet add timestamp */
	+ if ((m->m_flags & (M_TSTMP_LRO \| M_TSTMP)) == 0) {
	+ m->m_pkthdr.rcv_tstmp = sbttons(lc->lro_last_queue_time);
	+ m->m_flags \|= M_TSTMP_LRO;
	}

	- /*
	- * If the frame is padded beyond the end of the IP packet, then we must
	- * trim the extra bytes off.
	- */
	- hdr_len = m->m_pkthdr.len - (ETHER_HDR_LEN + ip_len);
	- if (hdr_len != 0) {
	- if (hdr_len < 0)
	- /* Truncated packet. */
	- return (TCP_LRO_CANNOT);
	+ /* Get pointer to TCP header. */
	+ th = pa->tcp;

	- m_adj(m, -hdr_len);
	- }
	- /*
	- * Check TCP header constraints.
	- */
	- hdr_len = (th->th_off << 2);
	- ts_ptr = (uint32_t *)(th + 1);
	- tcp_data_len -= hdr_len;
	- hdr_len -= sizeof(*th);
	- if (th->th_flags & TH_SYN)
	+ /* Don't process SYN packets. */
	+ if (__predict_false(th->th_flags & TH_SYN))
	return (TCP_LRO_CANNOT);
	- if ((th->th_flags & ~(TH_ACK \| TH_PUSH)) != 0) {
	- need_flush = 1;
	- } else
	- need_flush = 0;
	- if (hdr_len != 0 && (__predict_false(hdr_len != TCPOLEN_TSTAMP_APPA) \|\|
	- (*ts_ptr != ntohl(TCPOPT_NOP<<24\|TCPOPT_NOP<<16\|
	- TCPOPT_TIMESTAMP<<8\|TCPOLEN_TIMESTAMP)))) {
	- /*
	- * We have an option besides Timestamps, maybe
	- * it is a sack (most likely) which means we
	- * will probably need to wake up a sleeper (if
	- * the guy does queueing).
	- */
	- need_flush = 2;
	- }
	- /* If the driver did not pass in the checksum, set it now. */
	- if (csum == 0x0000)
	- csum = th->th_sum;
	- seq = ntohl(th->th_seq);
	+
	+ /* Get total TCP header length and compute payload length. */
	+ tcp_opt_len = (th->th_off << 2);
	+ tcp_data_len = m->m_pkthdr.len - ((uint8_t *)th -
	+ (uint8_t *)m->m_data) - tcp_opt_len;
	+ tcp_opt_len -= sizeof(*th);
	+
	+ /* Don't process invalid TCP headers. */
	+ if (__predict_false(tcp_opt_len < 0 \|\| tcp_data_len < 0))
	+ return (TCP_LRO_CANNOT);
	+
	+ /* Compute TCP data only checksum. */
	+ if (tcp_data_len == 0)
	+ tcp_data_sum = 0; /* no data, no checksum */
	+ else if (__predict_false(csum != 0))
	+ tcp_data_sum = tcp_lro_rx_csum_data(pa, ~csum);
	+ else
	+ tcp_data_sum = tcp_lro_rx_csum_data(pa, ~th->th_sum);
	+
	+ /* Save TCP info in mbuf. */
	+ m->m_nextpkt = NULL;
	+ m->m_pkthdr.rcvif = lc->ifp;
	+ m->m_pkthdr.lro_tcp_d_csum = tcp_data_sum;
	+ m->m_pkthdr.lro_tcp_d_len = tcp_data_len;
	+ m->m_pkthdr.lro_tcp_h_off = ((uint8_t )th - (uint8_t )m->m_data);
	+ m->m_pkthdr.lro_nsegs = 1;
	+
	+ /* Get hash bucket. */
	if (!use_hash) {
	bucket = &lc->lro_hash[0];
	- } else if (M_HASHTYPE_ISHASH(m)) {
	- bucket = &lc->lro_hash[m->m_pkthdr.flowid % lc->lro_hashsz];
	} else {
	- uint32_t hash;
	-
	- switch (eh_type) {
	-#ifdef INET
	- case ETHERTYPE_IP:
	- hash = ip4->ip_src.s_addr + ip4->ip_dst.s_addr;
	- break;
	-#endif
	-#ifdef INET6
	- case ETHERTYPE_IPV6:
	- hash = ip6->ip6_src.s6_addr32[0] +
	- ip6->ip6_dst.s6_addr32[0];
	- hash += ip6->ip6_src.s6_addr32[1] +
	- ip6->ip6_dst.s6_addr32[1];
	- hash += ip6->ip6_src.s6_addr32[2] +
	- ip6->ip6_dst.s6_addr32[2];
	- hash += ip6->ip6_src.s6_addr32[3] +
	- ip6->ip6_dst.s6_addr32[3];
	- break;
	-#endif
	- default:
	- hash = 0;
	- break;
	- }
	- hash += th->th_sport + th->th_dport;
	- bucket = &lc->lro_hash[hash % lc->lro_hashsz];
	+ bucket = tcp_lro_rx_get_bucket(lc, m, pa);
	}

	/* Try to find a matching previous segment. */
	LIST_FOREACH(le, bucket, hash_next) {
	- if (le->eh_type != eh_type)
	- continue;
	- if (le->source_port != th->th_sport \|\|
	- le->dest_port != th->th_dport)
	+ /* Compare addresses and ports. */
	+ if (lro_address_compare(&po.data, &le->outer.data) == false \|\|
	+ lro_address_compare(&pi.data, &le->inner.data) == false)
	continue;
	- switch (eh_type) {
	-#ifdef INET6
	- case ETHERTYPE_IPV6:
	- if (bcmp(&le->source_ip6, &ip6->ip6_src,
	- sizeof(struct in6_addr)) != 0 \|\|
	- bcmp(&le->dest_ip6, &ip6->ip6_dst,
	- sizeof(struct in6_addr)) != 0)
	- continue;
	- break;
	-#endif
	-#ifdef INET
	- case ETHERTYPE_IP:
	- if (le->source_ip4 != ip4->ip_src.s_addr \|\|
	- le->dest_ip4 != ip4->ip_dst.s_addr)
	- continue;
	- break;
	-#endif
	- }
	- if (tcp_data_len \|\| SEQ_GT(ntohl(th->th_ack), ntohl(le->ack_seq)) \|\|
	- (th->th_ack == le->ack_seq)) {
	- m->m_pkthdr.lro_len = tcp_data_len;
	- } else {
	- /* no data and old ack */
	+
	+ /* Check if no data and old ACK. */
	+ if (tcp_data_len == 0 &&
	+ SEQ_LT(ntohl(th->th_ack), ntohl(le->ack_seq))) {
	m_freem(m);
	return (0);
	}
	-#ifdef TCPHPTS
	- if ((tcplro_stacks_wanting_mbufq == 0) \|\| (m->m_flags & M_VLANTAG))
	- goto skip_lookup_a;
	- if (le->inp == NULL) {
	- CURVNET_SET(lc->ifp->if_vnet);
	- le->inp = tcp_lro_lookup(lc, le);
	- if (le->inp) {
	- in_pcbref(le->inp);
	- locked = 1;
	- }
	- CURVNET_RESTORE();
	- } else if (le->inp) {
	- INP_WLOCK(le->inp);
	- locked = 1;
	- }
	- if (locked && ((le->inp->inp_flags & (INP_DROPPED\|INP_TIMEWAIT)) \|\|
	- (le->inp->inp_flags2 & INP_FREED))) {
	- /* We can't present these to the inp since
	- * its dead Jim.
	- */
	- int ret;

	- ret = in_pcbrele_wlocked(le->inp);
	- if (ret == 0)
	- INP_WUNLOCK(le->inp);
	- le->inp = NULL;
	- locked = 0;
	- tcp_lro_active_remove(le);
	- if (le->strip_cnt && le->m_head) {
	- /*
	- * If we have any stripped packets we
	- * just dump the whole chain. The
	- * tcp_lro_flush code knows how
	- * to handle things when le->m_head is NULL
	- * and even le->inp is NULL.
	- */
	- lro_free_mbuf_chain(le->m_head);
	- le->m_head = NULL;
	- }
	- tcp_lro_flush(lc, le);
	- return (TCP_LRO_CANNOT);
	- }
	- /* See if it has been switched on */
	- if (le->inp && (le->inp->inp_flags2 & INP_MBUF_ACKCMP))
	- le->flags \|= CAN_USE_ACKCMP;
	-
	- if ((need_flush == 1) &&
	- le->inp &&
	- (le->inp->inp_flags2 & (INP_MBUF_ACKCMP\|INP_SUPPORTS_MBUFQ)) &&
	- ((th->th_flags & ~(TH_ACK \| TH_PUSH \| TH_ECE \| TH_CWR)) == 0)) {
	- /*
	- * For MBUF queuing or ACKCMP we can accept ECE and CWR
	- * since each packet is sent to the transport (or the
	- * compressed state including the ECN bits).
	- */
	- need_flush = 0;
	- }
	-skip_lookup_a:
	-#endif
	- if (need_flush)
	- le->need_wakeup = need_flush;
	- /* Save of the data only csum */
	- m->m_pkthdr.rcvif = lc->ifp;
	- m->m_pkthdr.lro_csum = tcp_lro_rx_csum_fixup(le, l3hdr, th,
	- tcp_data_len, ~csum);
	- th->th_sum = csum; /* Restore checksum */
	-#ifdef TCPHPTS
	- if ((le->flags & CAN_USE_ACKCMP) \|\|
	- (le->inp &&
	- (le->inp->inp_flags2 & (INP_MBUF_ACKCMP\|INP_SUPPORTS_MBUFQ)))) {
	- /*
	- * Mbuf queued and ACKCMP packets have their BPF and csum
	- * done here in LRO. They will still end up looking at the
	- * headers and such (IP/TCP) but we don't want to proceed
	- * with any bad csum!
	- */
	- m = do_bpf_and_csum(le->inp, lc, le, eh, m, bpf_peers_present(lc->ifp->if_bpf), locked);
	- if (m == NULL) {
	- /* Bad csum, accounting already done */
	- if (locked) {
	- INP_WUNLOCK(le->inp);
	- }
	- return (0);
	- }
	- le->strip_cnt++;
	- }
	- if ((need_flush == 0) &&
	- (th->th_flags & TH_ACK) &&
	- (tcp_data_len == 0) &&
	- (le->flags & CAN_USE_ACKCMP)) {
	- /*
	- * Ok this is a pure ack lets find out if our
	- * last packet already has one of these.
	- */
	- struct mbuf *nm;
	- struct tcp_ackent *ack_ent;
	- int idx;
	-
	- INP_WLOCK_ASSERT(le->inp);
	- if (le->m_head == NULL) {
	- /* Ok can we still use the end of the inp's? */
	- nm = tcp_lro_get_last_if_ackcmp(lc, le, le->inp);
	- if (nm == NULL) {
	- /* gone or full */
	- goto new_one;
	- }
	- /* We can add in to the one on the tail */
	- ack_ent = mtod(nm, struct tcp_ackent *);
	- idx = (nm->m_len / sizeof(struct tcp_ackent));
	- build_ack_entry(&ack_ent[idx], th, m, hdr_len, iptos);
	- /* Bump the size of both pkt-hdr and len */
	- nm->m_len += sizeof(struct tcp_ackent);
	- nm->m_pkthdr.len += sizeof(struct tcp_ackent);
	- le->ack_seq = th->th_ack;
	- le->window = th->th_win;
	- m_freem(m);
	- counter_u64_add(tcp_extra_mbuf, 1);
	- INP_WUNLOCK(le->inp);
	- return (0);
	- } else if (le->m_last_mbuf->m_flags & M_ACKCMP) {
	- /* Yes we might be able to be appended to */
	- nm = le->m_last_mbuf;
	- if (M_TRAILINGSPACE(nm) < sizeof(struct tcp_ackent)) {
	- if ((le->inp->inp_flags2 & INP_MBUF_L_ACKS) == 0) {
	- counter_u64_add(tcp_would_have_but, 1);
	- le->inp->inp_flags2 \|= INP_MBUF_L_ACKS;
	- }
	- goto new_one;
	- }
	- /* we have room */
	- ack_ent = mtod(nm, struct tcp_ackent *);
	- idx = (nm->m_len / sizeof(struct tcp_ackent));
	- build_ack_entry(&ack_ent[idx], th, m, hdr_len, iptos);
	- /* Bump the size of both pkt-hdr and len */
	- nm->m_len += sizeof(struct tcp_ackent);
	- nm->m_pkthdr.len += sizeof(struct tcp_ackent);
	- m_freem(m);
	- le->flags \|= HAS_COMP_ENTRIES;
	- le->cmp_ack_cnt++;
	- goto compressed;
	- } else {
	- /* Nope we need a new one */
	-new_one:
	- if (le->inp->inp_flags2 & INP_MBUF_L_ACKS)
	- nm = m_getcl(M_NOWAIT, MT_DATA, (M_ACKCMP\|M_PKTHDR));
	- else {
	- nm = m_gethdr(M_NOWAIT, MT_DATA);
	- nm->m_flags \|= M_ACKCMP;
	- }
	- if (nm) {
	- nm->m_pkthdr.rcvif = lc->ifp;
	- ack_ent = mtod(nm, struct tcp_ackent *);
	- build_ack_entry(ack_ent, th, m, hdr_len, iptos);
	- m_freem(m);
	- m = nm;
	- m->m_pkthdr.len = m->m_len = sizeof(struct tcp_ackent);
	- le->flags \|= HAS_COMP_ENTRIES;
	- le->cmp_ack_cnt++;
	- }
	- /* We fall through and append */
	- }
	- }
	- if (m->m_flags & M_ACKCMP) {
	- counter_u64_add(tcp_comp_total, 1);
	- } else {
	- counter_u64_add(tcp_uncomp_total, 1);
	- }
	-#endif
	- /* Save off the tail I am appending too (prev) */
	- m->m_nextpkt = NULL;
	- if (le->m_head == NULL) {
	- /*
	- * Case where we wer chaining off the inp
	- * and now no-longer can.
	- */
	- le->m_head = m;
	- le->m_tail = m_last(m);
	- le->m_last_mbuf = m;
	- le->m_prev_last = NULL;
	- } else {
	- le->m_prev_last = le->m_last_mbuf;
	- /* Mark me in the last spot */
	- le->m_last_mbuf->m_nextpkt = m;
	- /* Now set the tail to me */
	- le->m_last_mbuf = m;
	- le->tcp_tot_p_len += tcp_data_len;
	- }
	-#ifdef TCPHPTS
	-compressed:
	-#endif
	- le->mbuf_cnt++;
	- /* Add to the total size of data */
	- lro_set_mtime(&le->mtime, &arrv);
	- if (locked)
	- INP_WUNLOCK(le->inp);
	+ /* Mark "m" in the last spot. */
	+ le->m_last_mbuf->m_nextpkt = m;
	+ /* Now set the tail to "m". */
	+ le->m_last_mbuf = m;
	return (0);
	}
	+
	/* Try to find an empty slot. */
	if (LIST_EMPTY(&lc->lro_free))
	return (TCP_LRO_NO_ENTRIES);
	@@ -1890,181 +1762,40 @@
	le = LIST_FIRST(&lc->lro_free);
	LIST_REMOVE(le, next);
	tcp_lro_active_insert(lc, bucket, le);
	- lro_set_mtime(&le->mtime, &arrv);

	- /* Start filling in details. */
	- switch (eh_type) {
	-#ifdef INET6
	- case ETHERTYPE_IPV6:
	- le->le_ip6 = ip6;
	- le->source_ip6 = ip6->ip6_src;
	- le->dest_ip6 = ip6->ip6_dst;
	- le->eh_type = eh_type;
	- le->p_len = m->m_pkthdr.len - ETHER_HDR_LEN - sizeof(*ip6);
	- break;
	-#endif
	-#ifdef INET
	- case ETHERTYPE_IP:
	- le->le_ip4 = ip4;
	- le->source_ip4 = ip4->ip_src.s_addr;
	- le->dest_ip4 = ip4->ip_dst.s_addr;
	- le->eh_type = eh_type;
	- le->p_len = m->m_pkthdr.len - ETHER_HDR_LEN;
	- break;
	-#endif
	- }
	- le->source_port = th->th_sport;
	- le->dest_port = th->th_dport;
	- le->next_seq = seq + tcp_data_len;
	- le->ack_seq = th->th_ack;
	- le->window = th->th_win;
	- if (hdr_len != 0) {
	- le->timestamp = 1;
	- le->tsval = ntohl(*(ts_ptr + 1));
	- le->tsecr = *(ts_ptr + 2);
	- }
	- KASSERT(le->ulp_csum == 0, ("%s: le=%p le->ulp_csum=0x%04x\n",
	- __func__, le, le->ulp_csum));
	-
	- le->append_cnt = 0;
	- le->ulp_csum = tcp_lro_rx_csum_fixup(le, l3hdr, th, tcp_data_len,
	- ~csum);
	- le->append_cnt++;
	- th->th_sum = csum; /* Restore checksum */
	- m->m_pkthdr.rcvif = lc->ifp;
	- m->m_pkthdr.lro_len = tcp_data_len;
	- le->mbuf_cnt = 1;
	- le->cmp_ack_cnt = 0;
	- le->flags = 0;
	-#ifdef TCPHPTS
	- /*
	- * Lets find out if we can use the mbuf-compression.
	- */
	- if ((tcplro_stacks_wanting_mbufq == 0) \|\| (m->m_flags & M_VLANTAG))
	- goto skip_lookup_b;
	- CURVNET_SET(lc->ifp->if_vnet);
	- le->inp = tcp_lro_lookup(lc, le);
	- if (le->inp && ((le->inp->inp_flags & (INP_DROPPED\|INP_TIMEWAIT)) \|\|
	- (le->inp->inp_flags2 & INP_FREED))) {
	- INP_WUNLOCK(le->inp);
	- le->inp = NULL;
	- }
	- if (le->inp) {
	- if ((need_flush == 1) &&
	- (le->inp->inp_flags2 & (INP_MBUF_ACKCMP\|INP_SUPPORTS_MBUFQ)) &&
	- ((th->th_flags & ~(TH_ACK \| TH_PUSH \| TH_ECE \| TH_CWR)) == 0)) {
	- /*
	- * For MBUF queuing or ACKCMP we can accept ECE and CWR
	- * since each packet is sent to the transport (or the
	- * compressed state including the ECN bits).
	- */
	- need_flush = 0;
	- }
	- locked = 1;
	- if (le->inp->inp_flags2 & INP_MBUF_ACKCMP)
	- le->flags \|= CAN_USE_ACKCMP;
	- if ((le->flags & CAN_USE_ACKCMP) \|\|
	- (le->inp &&
	- (le->inp->inp_flags2 & (INP_MBUF_ACKCMP\|INP_SUPPORTS_MBUFQ)))) {
	- m = do_bpf_and_csum(le->inp, lc, le, eh, m, bpf_peers_present(lc->ifp->if_bpf), locked);
	- if (m == NULL) {
	- /* Bad csum, accounting already done */
	- INP_WUNLOCK(le->inp);
	- le->inp = NULL;
	- return (0);
	- }
	- le->strip_cnt++;
	- }
	- in_pcbref(le->inp);
	- }
	- CURVNET_RESTORE();
	- if ((need_flush == 0) &&
	- (th->th_flags & TH_ACK) &&
	- (tcp_data_len == 0) &&
	- (le->flags & CAN_USE_ACKCMP)) {
	- /* Ok this is a pure ack lets build our special COMPRESS mbuf */
	- struct mbuf *nm;
	- struct tcp_ackent *ack_ent;
	-
	- /* Question what is going on with the last mbuf on the inp queue, can we use it? */
	- INP_WLOCK_ASSERT(le->inp);
	- nm = tcp_lro_get_last_if_ackcmp(lc, le, le->inp);
	- if (nm) {
	- int idx;
	-
	- /* We can add in to the one on the tail */
	- ack_ent = mtod(nm, struct tcp_ackent *);
	- idx = (nm->m_len / sizeof(struct tcp_ackent));
	- build_ack_entry(&ack_ent[idx], th, m, hdr_len, iptos);
	- nm->m_len += sizeof(struct tcp_ackent);
	- nm->m_pkthdr.len += sizeof(struct tcp_ackent);
	- le->ack_seq = th->th_ack;
	- le->window = th->th_win;
	- m_freem(m);
	- counter_u64_add(tcp_extra_mbuf, 1);
	- le->m_head = NULL;
	- le->m_tail = NULL;
	- le->m_last_mbuf = NULL;
	- le->m_prev_last = NULL;
	- INP_WUNLOCK(le->inp);
	- return (0);
	- } else {
	- if (le->inp->inp_flags2 & INP_MBUF_L_ACKS)
	- nm = m_getcl(M_NOWAIT, MT_DATA, (M_ACKCMP\|M_PKTHDR));
	- else {
	- nm = m_gethdr(M_NOWAIT, MT_DATA);
	- nm->m_flags \|= M_ACKCMP;
	- }
	- if (nm) {
	- nm->m_pkthdr.rcvif = lc->ifp;
	- ack_ent = mtod(nm, struct tcp_ackent *);
	- build_ack_entry(ack_ent, th, m, hdr_len, iptos);
	- m_freem(m);
	- m = nm;
	- m->m_pkthdr.len = m->m_len = sizeof(struct tcp_ackent);
	- le->flags \|= HAS_COMP_ENTRIES;
	- le->cmp_ack_cnt++;
	- }
	- }
	- }
	- if (m->m_flags & M_ACKCMP) {
	- counter_u64_add(tcp_comp_total, 1);
	- } else {
	- counter_u64_add(tcp_uncomp_total, 1);
	- }
	-skip_lookup_b:
	-#endif
	- if (need_flush)
	- le->need_wakeup = need_flush;
	- else
	- le->need_wakeup = 0;
	- m->m_nextpkt = NULL;
	- le->m_head = m;
	- le->m_tail = m_last(m);
	+ /* Make sure the headers are set. */
	+ le->inner = pi;
	+ le->outer = po;
	+
	+ /* Store time this entry was allocated. */
	+ le->alloc_time = lc->lro_last_queue_time;
	+
	+ tcp_set_entry_to_mbuf(lc, le, m, th);
	+
	+ /* Now set the tail to "m". */
	le->m_last_mbuf = m;
	- le->m_prev_last = NULL;
	- /*
	- * We keep the total size here for cross checking when we may need
	- * to flush/wakeup in the MBUF_QUEUE case.
	- */
	- le->tcp_tot_p_len = tcp_data_len;
	- if (locked)
	- INP_WUNLOCK(le->inp);
	+
	return (0);
	}

	int
	tcp_lro_rx(struct lro_ctrl lc, struct mbuf m, uint32_t csum)
	{
	+ int error;
	+
	+ /* get current time */
	+ lc->lro_last_queue_time = getsbinuptime();
	+
	+ CURVNET_SET(lc->ifp->if_vnet);
	+ error = tcp_lro_rx_common(lc, m, csum, true);
	+ CURVNET_RESTORE();

	- return tcp_lro_rx2(lc, m, csum, 1);
	+ return (error);
	}

	void
	tcp_lro_queue_mbuf(struct lro_ctrl lc, struct mbuf mb)
	{
	- struct timespec arrv;
	-
	/* sanity checks */
	if (__predict_false(lc->ifp == NULL \|\| lc->lro_mbuf_data == NULL \|\|
	lc->lro_mbuf_max == 0)) {
	@@ -2080,15 +1811,7 @@
	(*lc->ifp->if_input) (lc->ifp, mb);
	return;
	}
	- /* Arrival Stamp the packet */
	-
	- if ((mb->m_flags & M_TSTMP) == 0) {
	- /* If no hardware or arrival stamp on the packet add arrival */
	- nanouptime(&arrv);
	- mb->m_pkthdr.rcv_tstmp = ((arrv.tv_sec * 1000000000) +
	- arrv.tv_nsec);
	- mb->m_flags \|= M_TSTMP_LRO;
	- }
	+
	/* create sequence number */
	lc->lro_mbuf_data[lc->lro_mbuf_count].seq =
	(((uint64_t)M_HASHTYPE_GET(mb)) << 56) \|
	diff --git a/sys/sys/mbuf.h b/sys/sys/mbuf.h
	--- a/sys/sys/mbuf.h
	+++ b/sys/sys/mbuf.h
	@@ -203,9 +203,9 @@
	#define tso_segsz PH_per.sixteen[1] /* inbound after LRO */
	#define lro_nsegs tso_segsz /* inbound after LRO */
	#define csum_data PH_per.thirtytwo[1] /* inbound from hardware up */
	-#define lro_len PH_loc.sixteen[0] /* inbound during LRO (no reassembly) */
	-#define lro_csum PH_loc.sixteen[1] /* inbound during LRO (no reassembly) */
	-#define lro_etype PH_loc.sixteen[2] /* inbound during LRO (no reassembly) */
	+#define lro_tcp_d_len PH_loc.sixteen[0] /* inbound during LRO (no reassembly) */
	+#define lro_tcp_d_csum PH_loc.sixteen[1] /* inbound during LRO (no reassembly) */
	+#define lro_tcp_h_off PH_loc.sixteen[2] /* inbound during LRO (no reassembly) */
	/* Note PH_loc is used during IP reassembly (all 8 bytes as a ptr) */

	/*

File Metadata

Mime Type: text/plain
Expires: Wed, Jul 22, 10:04 PM (2 h, 38 m)
Storage Engine: blob
Storage Format: Raw Data
Storage Handle: 35383813
Default Alt Text: D29564.id86781.diff (78 KB)

D29564.id86781.diffNo OneTemporaryActions

D29564.id86781.diffView Options

File Metadata

Event Timeline

D29564.id86781.diff
No OneTemporary
Actions

D29564.id86781.diff
View Options