xref: /freebsd/sys/netinet/tcp_stacks/rack_bbr_common.c (revision 120e232f1ae386f0ce413b27d60b3d52c568c58e)
1 /*-
2  * Copyright (c) 2016-2020 Netflix, Inc.
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  * 1. Redistributions of source code must retain the above copyright
8  *    notice, this list of conditions and the following disclaimer.
9  * 2. Redistributions in binary form must reproduce the above copyright
10  *    notice, this list of conditions and the following disclaimer in the
11  *    documentation and/or other materials provided with the distribution.
12  *
13  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
14  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
17  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
23  * SUCH DAMAGE.
24  *
25  */
26 /*
27  * Author: Randall Stewart <rrs@netflix.com>
28  * This work is based on the ACM Queue paper
29  * BBR - Congestion Based Congestion Control
30  * and also numerous discussions with Neal, Yuchung and Van.
31  */
32 
33 #include <sys/cdefs.h>
34 #include "opt_inet.h"
35 #include "opt_inet6.h"
36 #include "opt_ipsec.h"
37 #include "opt_ratelimit.h"
38 #include <sys/param.h>
39 #include <sys/arb.h>
40 #include <sys/module.h>
41 #include <sys/kernel.h>
42 #ifdef TCP_HHOOK
43 #include <sys/hhook.h>
44 #endif
45 #include <sys/malloc.h>
46 #include <sys/mbuf.h>
47 #include <sys/proc.h>
48 #include <sys/qmath.h>
49 #include <sys/socket.h>
50 #include <sys/socketvar.h>
51 #include <sys/sysctl.h>
52 #include <sys/systm.h>
53 #include <sys/tree.h>
54 #ifdef NETFLIX_STATS
55 #include <sys/stats.h> /* Must come after qmath.h and tree.h */
56 #endif
57 #include <sys/refcount.h>
58 #include <sys/queue.h>
59 #include <sys/smp.h>
60 #include <sys/kthread.h>
61 #include <sys/lock.h>
62 #include <sys/mutex.h>
63 #include <sys/tim_filter.h>
64 #include <sys/time.h>
65 #include <vm/uma.h>
66 #include <sys/kern_prefetch.h>
67 
68 #include <net/route.h>
69 #include <net/vnet.h>
70 #include <net/ethernet.h>
71 #include <net/bpf.h>
72 
73 #define TCPSTATES		/* for logging */
74 
75 #include <netinet/in.h>
76 #include <netinet/in_kdtrace.h>
77 #include <netinet/in_pcb.h>
78 #include <netinet/ip.h>
79 #include <netinet/ip_var.h>
80 #include <netinet/ip6.h>
81 #include <netinet6/in6_pcb.h>
82 #include <netinet6/ip6_var.h>
83 #include <netinet/tcp.h>
84 #include <netinet/tcp_fsm.h>
85 #include <netinet/tcp_seq.h>
86 #include <netinet/tcp_timer.h>
87 #include <netinet/tcp_var.h>
88 #include <netinet/tcpip.h>
89 #include <netinet/tcp_ecn.h>
90 #include <netinet/tcp_hpts.h>
91 #include <netinet/tcp_lro.h>
92 #include <netinet/cc/cc.h>
93 #include <netinet/tcp_log_buf.h>
94 #ifdef TCP_OFFLOAD
95 #include <netinet/tcp_offload.h>
96 #endif
97 #ifdef INET6
98 #include <netinet6/tcp6_var.h>
99 #endif
100 #include <netinet/tcp_fastopen.h>
101 
102 #include <netipsec/ipsec_support.h>
103 #include <net/if.h>
104 #include <net/if_var.h>
105 #include <net/if_private.h>
106 
107 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
108 #include <netipsec/ipsec.h>
109 #include <netipsec/ipsec6.h>
110 #endif				/* IPSEC */
111 
112 #include <netinet/udp.h>
113 #include <netinet/udp_var.h>
114 #include <machine/in_cksum.h>
115 
116 #ifdef MAC
117 #include <security/mac/mac_framework.h>
118 #endif
119 #include "rack_bbr_common.h"
120 
121 /*
122  * Common TCP Functions - These are shared by borth
123  * rack and BBR.
124  */
125 static int
ctf_get_enet_type(struct ifnet * ifp,struct mbuf * m)126 ctf_get_enet_type(struct ifnet *ifp, struct mbuf *m)
127 {
128 	struct ether_header *eh;
129 #ifdef INET6
130 	struct ip6_hdr *ip6 = NULL;	/* Keep compiler happy. */
131 #endif
132 #ifdef INET
133 	struct ip *ip = NULL;		/* Keep compiler happy. */
134 #endif
135 #if defined(INET) || defined(INET6)
136 	struct tcphdr *th;
137 	int32_t tlen;
138 	uint16_t drop_hdrlen;
139 #endif
140 	uint16_t etype;
141 #ifdef INET
142 	uint8_t iptos;
143 #endif
144 
145 	/* Is it the easy way? */
146 	if (m->m_flags & M_LRO_EHDRSTRP)
147 		return (m->m_pkthdr.lro_etype);
148 	/*
149 	 * Ok this is the old style call, the ethernet header is here.
150 	 * This also means no checksum or BPF were done. This
151 	 * can happen if the race to setup the inp fails and
152 	 * LRO sees no INP at packet input, but by the time
153 	 * we queue the packets an INP gets there. Its rare
154 	 * but it can occur so we will handle it. Note that
155 	 * this means duplicated work but with the rarity of it
156 	 * its not worth worrying about.
157 	 */
158 	/* Let the BPF see the packet */
159 	if (bpf_peers_present(ifp->if_bpf))
160 		ETHER_BPF_MTAP(ifp, m);
161 	/* Now the csum */
162 	eh = mtod(m, struct ether_header *);
163 	etype = ntohs(eh->ether_type);
164 	m_adj(m,  sizeof(*eh));
165 	switch (etype) {
166 #ifdef INET6
167 		case ETHERTYPE_IPV6:
168 		{
169 			if (m->m_len < (sizeof(*ip6) + sizeof(*th))) {
170 				m = m_pullup(m, sizeof(*ip6) + sizeof(*th));
171 				if (m == NULL) {
172 					KMOD_TCPSTAT_INC(tcps_rcvshort);
173 					return (-1);
174 				}
175 			}
176 			ip6 = (struct ip6_hdr *)(eh + 1);
177 			th = (struct tcphdr *)(ip6 + 1);
178 			drop_hdrlen = sizeof(*ip6);
179 			tlen = ntohs(ip6->ip6_plen);
180 			if (m->m_pkthdr.csum_flags & CSUM_DATA_VALID_IPV6) {
181 				if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR)
182 					th->th_sum = m->m_pkthdr.csum_data;
183 				else
184 					th->th_sum = in6_cksum_pseudo(ip6, tlen,
185 								      IPPROTO_TCP,
186 								      m->m_pkthdr.csum_data);
187 				th->th_sum ^= 0xffff;
188 			} else
189 				th->th_sum = in6_cksum(m, IPPROTO_TCP, drop_hdrlen, tlen);
190 			if (th->th_sum) {
191 				KMOD_TCPSTAT_INC(tcps_rcvbadsum);
192 				m_freem(m);
193 				return (-1);
194 			}
195 			return (etype);
196 		}
197 #endif
198 #ifdef INET
199 		case ETHERTYPE_IP:
200 		{
201 			if (m->m_len < sizeof (struct tcpiphdr)) {
202 				m = m_pullup(m, sizeof (struct tcpiphdr));
203 				if (m == NULL) {
204 					KMOD_TCPSTAT_INC(tcps_rcvshort);
205 					return (-1);
206 				}
207 			}
208 			ip = (struct ip *)(eh + 1);
209 			th = (struct tcphdr *)(ip + 1);
210 			drop_hdrlen = sizeof(*ip);
211 			iptos = ip->ip_tos;
212 			tlen = ntohs(ip->ip_len) - sizeof(struct ip);
213 			if (m->m_pkthdr.csum_flags & CSUM_DATA_VALID) {
214 				if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR)
215 					th->th_sum = m->m_pkthdr.csum_data;
216 				else
217 					th->th_sum = in_pseudo(ip->ip_src.s_addr,
218 							       ip->ip_dst.s_addr,
219 							       htonl(m->m_pkthdr.csum_data + tlen + IPPROTO_TCP));
220 				th->th_sum ^= 0xffff;
221 			} else {
222 				int len;
223 				struct ipovly *ipov = (struct ipovly *)ip;
224 				/*
225 				 * Checksum extended TCP header and data.
226 				 */
227 				len = drop_hdrlen + tlen;
228 				bzero(ipov->ih_x1, sizeof(ipov->ih_x1));
229 				ipov->ih_len = htons(tlen);
230 				th->th_sum = in_cksum(m, len);
231 				/* Reset length for SDT probes. */
232 				ip->ip_len = htons(len);
233 				/* Reset TOS bits */
234 				ip->ip_tos = iptos;
235 				/* Re-initialization for later version check */
236 				ip->ip_v = IPVERSION;
237 				ip->ip_hl = sizeof(*ip) >> 2;
238 			}
239 			if (th->th_sum) {
240 				KMOD_TCPSTAT_INC(tcps_rcvbadsum);
241 				m_freem(m);
242 				return (-1);
243 			}
244 			break;
245 		}
246 #endif
247 	};
248 	return (etype);
249 }
250 
251 /*
252  * The function ctf_process_inbound_raw() is used by
253  * transport developers to do the steps needed to
254  * support MBUF Queuing i.e. the flags in
255  * inp->inp_flags2:
256  *
257  * - INP_SUPPORTS_MBUFQ
258  * - INP_MBUF_QUEUE_READY
259  * - INP_DONT_SACK_QUEUE
260  * - INP_MBUF_ACKCMP
261  *
262  * These flags help control how LRO will deliver
263  * packets to the transport. You first set in inp_flags2
264  * the INP_SUPPORTS_MBUFQ to tell the LRO code that you
265  * will gladly take a queue of packets instead of a compressed
266  * single packet. You also set in your t_fb pointer the
267  * tfb_do_queued_segments to point to ctf_process_inbound_raw.
268  *
269  * This then gets you lists of inbound ACK's/Data instead
270  * of a condensed compressed ACK/DATA packet. Why would you
271  * want that? This will get you access to all the arrival
272  * times of at least LRO and possibly at the Hardware (if
273  * the interface card supports that) of the actual ACK/DATA.
274  * In some transport designs this is important since knowing
275  * the actual time we got the packet is useful information.
276  *
277  * A new special type of mbuf may also be supported by the transport
278  * if it has set the INP_MBUF_ACKCMP flag. If its set, LRO will
279  * possibly create a M_ACKCMP type mbuf. This is a mbuf with
280  * an array of "acks". One thing also to note is that when this
281  * occurs a subsequent LRO may find at the back of the untouched
282  * mbuf queue chain a M_ACKCMP and append on to it. This means
283  * that until the transport pulls in the mbuf chain queued
284  * for it more ack's may get on the mbufs that were already
285  * delivered. There currently is a limit of 6 acks condensed
286  * into 1 mbuf which means often when this is occuring, we
287  * don't get that effect but it does happen.
288  *
289  * Now there are some interesting Caveats that the transport
290  * designer needs to take into account when using this feature.
291  *
292  * 1) It is used with HPTS and pacing, when the pacing timer
293  *    for output calls it will first call the input.
294  * 2) When you set INP_MBUF_QUEUE_READY this tells LRO
295  *    queue normal packets, I am busy pacing out data and
296  *    will process the queued packets before my tfb_tcp_output
297  *    call from pacing. If a non-normal packet arrives, (e.g. sack)
298  *    you will be awoken immediately.
299  * 3) Finally you can add the INP_DONT_SACK_QUEUE to not even
300  *    be awoken if a SACK has arrived. You would do this when
301  *    you were not only running a pacing for output timer
302  *    but a Rack timer as well i.e. you know you are in recovery
303  *    and are in the process (via the timers) of dealing with
304  *    the loss.
305  *
306  * Now a critical thing you must be aware of here is that the
307  * use of the flags has a far greater scope then just your
308  * typical LRO. Why? Well thats because in the normal compressed
309  * LRO case at the end of a driver interupt all packets are going
310  * to get presented to the transport no matter if there is one
311  * or 100. With the MBUF_QUEUE model, this is not true. You will
312  * only be awoken to process the queue of packets when:
313  *     a) The flags discussed above allow it.
314  *          <or>
315  *     b) You exceed a ack or data limit (by default the
316  *        ack limit is infinity (64k acks) and the data
317  *        limit is 64k of new TCP data)
318  *         <or>
319  *     c) The push bit has been set by the peer
320  */
321 
322 static int
ctf_process_inbound_raw(struct tcpcb * tp,struct mbuf * m,int has_pkt)323 ctf_process_inbound_raw(struct tcpcb *tp, struct mbuf *m, int has_pkt)
324 {
325 	/*
326 	 * We are passed a raw change of mbuf packets
327 	 * that arrived in LRO. They are linked via
328 	 * the m_nextpkt link in the pkt-headers.
329 	 *
330 	 * We process each one by:
331 	 * a) saving off the next
332 	 * b) stripping off the ether-header
333 	 * c) formulating the arguments for tfb_do_segment_nounlock()
334 	 * d) calling each mbuf to tfb_do_segment_nounlock()
335 	 *    after adjusting the time to match the arrival time.
336 	 * Note that the LRO code assures no IP options are present.
337 	 *
338 	 * The symantics for calling tfb_do_segment_nounlock() are the
339 	 * following:
340 	 * 1) It returns 0 if all went well and you (the caller) need
341 	 *    to release the lock.
342 	 * 2) If nxt_pkt is set, then the function will surpress calls
343 	 *    to tcp_output() since you are promising to call again
344 	 *    with another packet.
345 	 * 3) If it returns 1, then you must free all the packets being
346 	 *    shipped in, the tcb has been destroyed (or about to be destroyed).
347 	 */
348 	struct mbuf *m_save;
349 	struct tcphdr *th;
350 #ifdef INET6
351 	struct ip6_hdr *ip6 = NULL;	/* Keep compiler happy. */
352 #endif
353 #ifdef INET
354 	struct ip *ip = NULL;		/* Keep compiler happy. */
355 #endif
356 	struct ifnet *ifp;
357 	struct timeval tv;
358 	struct inpcb *inp __diagused;
359 	int32_t retval, nxt_pkt, tlen, off;
360 	int etype = 0;
361 	uint16_t drop_hdrlen;
362 	uint8_t iptos;
363 
364 	inp = tptoinpcb(tp);
365 	INP_WLOCK_ASSERT(inp);
366 	NET_EPOCH_ASSERT();
367 	KASSERT(m != NULL, ("ctf_process_inbound_raw: m == NULL"));
368 	ifp = m_rcvif(m);
369 	KASSERT(ifp != NULL, ("ctf_process_inbound_raw: ifp == NULL"));
370 	CURVNET_SET(ifp->if_vnet);
371 	tcp_get_usecs(&tv);
372 	while (m) {
373 		m_save = m->m_nextpkt;
374 		m->m_nextpkt = NULL;
375 		if ((m->m_flags & M_ACKCMP) == 0) {
376 			/* Now lets get the ether header */
377 			etype = ctf_get_enet_type(ifp, m);
378 			if (etype == -1) {
379 				/* Skip this packet it was freed by checksum */
380 				goto skipped_pkt;
381 			}
382 			KASSERT(((etype == ETHERTYPE_IPV6) || (etype == ETHERTYPE_IP)),
383 				("tp:%p m:%p etype:0x%x -- not IP or IPv6", tp, m, etype));
384 			/* Trim off the ethernet header */
385 			switch (etype) {
386 #ifdef INET6
387 			case ETHERTYPE_IPV6:
388 				ip6 = mtod(m, struct ip6_hdr *);
389 				th = (struct tcphdr *)(ip6 + 1);
390 				tlen = ntohs(ip6->ip6_plen);
391 				drop_hdrlen = sizeof(*ip6);
392 				iptos = (ntohl(ip6->ip6_flow) >> 20) & 0xff;
393 				break;
394 #endif
395 #ifdef INET
396 			case ETHERTYPE_IP:
397 				ip = mtod(m, struct ip *);
398 				th = (struct tcphdr *)(ip + 1);
399 				drop_hdrlen = sizeof(*ip);
400 				iptos = ip->ip_tos;
401 				tlen = ntohs(ip->ip_len) - sizeof(struct ip);
402 				break;
403 #endif
404 			} /* end switch */
405 			off = th->th_off << 2;
406 			if (off < sizeof (struct tcphdr) || off > tlen) {
407 				printf("off:%d < hdrlen:%zu || > tlen:%u -- dump\n",
408 				       off,
409 				       sizeof(struct tcphdr),
410 				       tlen);
411 				KMOD_TCPSTAT_INC(tcps_rcvbadoff);
412 				m_freem(m);
413 				goto skipped_pkt;
414 			}
415 			tlen -= off;
416 			drop_hdrlen += off;
417 			/*
418 			 * Now lets setup the timeval to be when we should
419 			 * have been called (if we can).
420 			 */
421 			m->m_pkthdr.lro_nsegs = 1;
422 			/* Now what about next packet? */
423 		} else {
424 			/*
425 			 * This mbuf is an array of acks that have
426 			 * been compressed. We assert the inp has
427 			 * the flag set to enable this!
428 			 */
429 			KASSERT((tp->t_flags2 & TF2_MBUF_ACKCMP),
430 			    ("tp:%p no TF2_MBUF_ACKCMP flags?", tp));
431 			tlen = 0;
432 			drop_hdrlen = 0;
433 			th = NULL;
434 			iptos = 0;
435 		}
436 		tcp_get_usecs(&tv);
437 		if (m_save || has_pkt)
438 			nxt_pkt = 1;
439 		else
440 			nxt_pkt = 0;
441 		if ((m->m_flags & M_ACKCMP) == 0)
442 			KMOD_TCPSTAT_INC(tcps_rcvtotal);
443 		else
444 			KMOD_TCPSTAT_ADD(tcps_rcvtotal, (m->m_len / sizeof(struct tcp_ackent)));
445 		retval = (*tp->t_fb->tfb_do_segment_nounlock)(tp, m, th,
446 		    drop_hdrlen, tlen, iptos, nxt_pkt, &tv);
447 		if (retval) {
448 			/* We lost the lock and tcb probably */
449 			m = m_save;
450 			while(m) {
451 				m_save = m->m_nextpkt;
452 				m->m_nextpkt = NULL;
453 				m_freem(m);
454 				m = m_save;
455 			}
456 			CURVNET_RESTORE();
457 			INP_UNLOCK_ASSERT(inp);
458 			return (retval);
459 		}
460 skipped_pkt:
461 		m = m_save;
462 	}
463 	CURVNET_RESTORE();
464 	return (0);
465 }
466 
467 int
ctf_do_queued_segments(struct tcpcb * tp,int have_pkt)468 ctf_do_queued_segments(struct tcpcb *tp, int have_pkt)
469 {
470 	struct mbuf *m;
471 
472 	/* First lets see if we have old packets */
473 	if ((m = STAILQ_FIRST(&tp->t_inqueue)) != NULL) {
474 		STAILQ_INIT(&tp->t_inqueue);
475 		if (ctf_process_inbound_raw(tp, m, have_pkt)) {
476 			/* We lost the tcpcb (maybe a RST came in)? */
477 			return(1);
478 		}
479 	}
480 	return (0);
481 }
482 
483 uint32_t
ctf_outstanding(struct tcpcb * tp)484 ctf_outstanding(struct tcpcb *tp)
485 {
486 	uint32_t bytes_out;
487 
488 	bytes_out = tp->snd_max - tp->snd_una;
489 	if (tp->t_state < TCPS_ESTABLISHED)
490 		bytes_out++;
491 	if (tp->t_flags & TF_SENTFIN)
492 		bytes_out++;
493 	return (bytes_out);
494 }
495 
496 uint32_t
ctf_flight_size(struct tcpcb * tp,uint32_t rc_sacked)497 ctf_flight_size(struct tcpcb *tp, uint32_t rc_sacked)
498 {
499 	if (rc_sacked <= ctf_outstanding(tp))
500 		return(ctf_outstanding(tp) - rc_sacked);
501 	else {
502 		return (0);
503 	}
504 }
505 
506 void
ctf_do_dropwithreset(struct mbuf * m,struct tcpcb * tp,struct tcphdr * th,int32_t tlen)507 ctf_do_dropwithreset(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th,
508     int32_t tlen)
509 {
510 	tcp_dropwithreset(m, th, tp, tlen);
511 	if (tp != NULL)
512 		INP_WUNLOCK(tptoinpcb(tp));
513 }
514 
515 void
ctf_ack_war_checks(struct tcpcb * tp)516 ctf_ack_war_checks(struct tcpcb *tp)
517 {
518 	sbintime_t now;
519 
520 	if ((V_tcp_ack_war_time_window > 0) && (V_tcp_ack_war_cnt > 0)) {
521 		now = getsbinuptime();
522 		if (tp->t_challenge_ack_end < now) {
523 			tp->t_challenge_ack_cnt = 0;
524 			tp->t_challenge_ack_end = now +
525 			    V_tcp_ack_war_time_window * SBT_1MS;
526 		}
527 		if (tp->t_challenge_ack_cnt < V_tcp_ack_war_cnt) {
528 			tp->t_challenge_ack_cnt++;
529 			tp->t_flags |= TF_ACKNOW;
530 		} else
531 			tp->t_flags &= ~TF_ACKNOW;
532 	} else
533 		tp->t_flags |= TF_ACKNOW;
534 }
535 
536 /*
537  * ctf_drop_checks returns 1 for you should not proceed. It places
538  * in ret_val what should be returned 1/0 by the caller. The 1 indicates
539  * that the TCB is unlocked and probably dropped. The 0 indicates the
540  * TCB is still valid and locked.
541  */
542 int
ctf_drop_checks(struct tcpopt * to,struct mbuf * m,struct tcphdr * th,struct tcpcb * tp,int32_t * tlenp,int32_t * thf,int32_t * drop_hdrlen,int32_t * ret_val)543 ctf_drop_checks(struct tcpopt *to, struct mbuf *m, struct tcphdr *th,
544 		struct tcpcb *tp, int32_t *tlenp,
545 		int32_t *thf, int32_t *drop_hdrlen, int32_t *ret_val)
546 {
547 	int32_t todrop;
548 	int32_t thflags;
549 	int32_t tlen;
550 
551 	thflags = *thf;
552 	tlen = *tlenp;
553 	todrop = tp->rcv_nxt - th->th_seq;
554 	if (todrop > 0) {
555 		if (thflags & TH_SYN) {
556 			thflags &= ~TH_SYN;
557 			th->th_seq++;
558 			if (th->th_urp > 1)
559 				th->th_urp--;
560 			else
561 				thflags &= ~TH_URG;
562 			todrop--;
563 		}
564 		/*
565 		 * Following if statement from Stevens, vol. 2, p. 960.
566 		 */
567 		if (todrop > tlen
568 		    || (todrop == tlen && (thflags & TH_FIN) == 0)) {
569 			/*
570 			 * Any valid FIN must be to the left of the window.
571 			 * At this point the FIN must be a duplicate or out
572 			 * of sequence; drop it.
573 			 */
574 			thflags &= ~TH_FIN;
575 			/*
576 			 * Send an ACK to resynchronize and drop any data.
577 			 * But keep on processing for RST or ACK.
578 			 */
579 			ctf_ack_war_checks(tp);
580 			todrop = tlen;
581 			KMOD_TCPSTAT_INC(tcps_rcvduppack);
582 			KMOD_TCPSTAT_ADD(tcps_rcvdupbyte, todrop);
583 		} else {
584 			KMOD_TCPSTAT_INC(tcps_rcvpartduppack);
585 			KMOD_TCPSTAT_ADD(tcps_rcvpartdupbyte, todrop);
586 		}
587 		/*
588 		 * DSACK - add SACK block for dropped range
589 		 */
590 		if ((todrop > 0) && (tp->t_flags & TF_SACK_PERMIT)) {
591 			/*
592 			 * ACK now, as the next in-sequence segment
593 			 * will clear the DSACK block again
594 			 */
595 			ctf_ack_war_checks(tp);
596 			if (tp->t_flags & TF_ACKNOW)
597 				tcp_update_sack_list(tp, th->th_seq,
598 						     th->th_seq + todrop);
599 		}
600 		*drop_hdrlen += todrop;	/* drop from the top afterwards */
601 		th->th_seq += todrop;
602 		tlen -= todrop;
603 		if (th->th_urp > todrop)
604 			th->th_urp -= todrop;
605 		else {
606 			thflags &= ~TH_URG;
607 			th->th_urp = 0;
608 		}
609 	}
610 	/*
611 	 * If segment ends after window, drop trailing data (and PUSH and
612 	 * FIN); if nothing left, just ACK.
613 	 */
614 	todrop = (th->th_seq + tlen) - (tp->rcv_nxt + tp->rcv_wnd);
615 	if (todrop > 0) {
616 		KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
617 		if (todrop >= tlen) {
618 			KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, tlen);
619 			/*
620 			 * If window is closed can only take segments at
621 			 * window edge, and have to drop data and PUSH from
622 			 * incoming segments.  Continue processing, but
623 			 * remember to ack.  Otherwise, drop segment and
624 			 * ack.
625 			 */
626 			if (tp->rcv_wnd == 0 && th->th_seq == tp->rcv_nxt) {
627 				ctf_ack_war_checks(tp);
628 				KMOD_TCPSTAT_INC(tcps_rcvwinprobe);
629 			} else {
630 				ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val);
631 				return (1);
632 			}
633 		} else
634 			KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
635 		m_adj(m, -todrop);
636 		tlen -= todrop;
637 		thflags &= ~(TH_PUSH | TH_FIN);
638 	}
639 	*thf = thflags;
640 	*tlenp = tlen;
641 	return (0);
642 }
643 
644 /*
645  * The value in ret_val informs the caller
646  * if we dropped the tcb (and lock) or not.
647  * 1 = we dropped it, 0 = the TCB is still locked
648  * and valid.
649  */
650 void
ctf_do_dropafterack(struct mbuf * m,struct tcpcb * tp,struct tcphdr * th,int32_t thflags,int32_t tlen,int32_t * ret_val)651 ctf_do_dropafterack(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th, int32_t thflags, int32_t tlen, int32_t *ret_val)
652 {
653 	/*
654 	 * Generate an ACK dropping incoming segment if it occupies sequence
655 	 * space, where the ACK reflects our state.
656 	 *
657 	 * We can now skip the test for the RST flag since all paths to this
658 	 * code happen after packets containing RST have been dropped.
659 	 *
660 	 * In the SYN-RECEIVED state, don't send an ACK unless the segment
661 	 * we received passes the SYN-RECEIVED ACK test. If it fails send a
662 	 * RST.  This breaks the loop in the "LAND" DoS attack, and also
663 	 * prevents an ACK storm between two listening ports that have been
664 	 * sent forged SYN segments, each with the source address of the
665 	 * other.
666 	 */
667 	if (tp->t_state == TCPS_SYN_RECEIVED && (thflags & TH_ACK) &&
668 	    (SEQ_GT(tp->snd_una, th->th_ack) ||
669 	    SEQ_GT(th->th_ack, tp->snd_max))) {
670 		*ret_val = 1;
671 		ctf_do_dropwithreset(m, tp, th, tlen);
672 		return;
673 	} else
674 		*ret_val = 0;
675 	ctf_ack_war_checks(tp);
676 	if (m)
677 		m_freem(m);
678 }
679 
680 void
ctf_do_drop(struct mbuf * m,struct tcpcb * tp)681 ctf_do_drop(struct mbuf *m, struct tcpcb *tp)
682 {
683 
684 	/*
685 	 * Drop space held by incoming segment and return.
686 	 */
687 	if (tp != NULL)
688 		INP_WUNLOCK(tptoinpcb(tp));
689 	if (m)
690 		m_freem(m);
691 }
692 
693 int
ctf_process_rst(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp)694 ctf_process_rst(struct mbuf *m, struct tcphdr *th, struct socket *so,
695 		struct tcpcb *tp)
696 {
697 	/*
698 	 * RFC5961 Section 3.2
699 	 *
700 	 * - RST drops connection only if SEG.SEQ == RCV.NXT. - If RST is in
701 	 * window, we send challenge ACK.
702 	 *
703 	 * Note: to take into account delayed ACKs, we should test against
704 	 * last_ack_sent instead of rcv_nxt. Note 2: we handle special case
705 	 * of closed window, not covered by the RFC.
706 	 */
707 	int dropped = 0;
708 
709 	if ((SEQ_GEQ(th->th_seq, tp->last_ack_sent) &&
710 	    SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) ||
711 	    (tp->rcv_wnd == 0 && tp->last_ack_sent == th->th_seq)) {
712 		KASSERT(tp->t_state != TCPS_SYN_SENT,
713 		    ("%s: TH_RST for TCPS_SYN_SENT th %p tp %p",
714 		    __func__, th, tp));
715 
716 		if (V_tcp_insecure_rst ||
717 		    (tp->last_ack_sent == th->th_seq) ||
718 		    (tp->rcv_nxt == th->th_seq)) {
719 			KMOD_TCPSTAT_INC(tcps_drops);
720 			/* Drop the connection. */
721 			switch (tp->t_state) {
722 			case TCPS_SYN_RECEIVED:
723 				so->so_error = ECONNREFUSED;
724 				goto close;
725 			case TCPS_ESTABLISHED:
726 			case TCPS_FIN_WAIT_1:
727 			case TCPS_FIN_WAIT_2:
728 			case TCPS_CLOSE_WAIT:
729 			case TCPS_CLOSING:
730 			case TCPS_LAST_ACK:
731 				so->so_error = ECONNRESET;
732 		close:
733 				tcp_state_change(tp, TCPS_CLOSED);
734 				/* FALLTHROUGH */
735 			default:
736 				tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_RST);
737 				tp = tcp_close(tp);
738 			}
739 			dropped = 1;
740 			ctf_do_drop(m, tp);
741 		} else {
742 			KMOD_TCPSTAT_INC(tcps_badrst);
743 			tcp_send_challenge_ack(tp, th, m);
744 		}
745 	} else {
746 		m_freem(m);
747 	}
748 	return (dropped);
749 }
750 
751 /*
752  * The value in ret_val informs the caller
753  * if we dropped the tcb (and lock) or not.
754  * 1 = we dropped it, 0 = the TCB is still locked
755  * and valid.
756  */
757 void
ctf_challenge_ack(struct mbuf * m,struct tcphdr * th,struct tcpcb * tp,uint8_t iptos,int32_t * ret_val)758 ctf_challenge_ack(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, uint8_t iptos, int32_t * ret_val)
759 {
760 
761 	NET_EPOCH_ASSERT();
762 
763 	KMOD_TCPSTAT_INC(tcps_badsyn);
764 	if (V_tcp_insecure_syn &&
765 	    SEQ_GEQ(th->th_seq, tp->last_ack_sent) &&
766 	    SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) {
767 		tp = tcp_drop(tp, ECONNRESET);
768 		*ret_val = 1;
769 		ctf_do_drop(m, tp);
770 	} else {
771 		tcp_ecn_input_syn_sent(tp, tcp_get_flags(th), iptos);
772 		/* Send challenge ACK. */
773 		tcp_respond(tp, mtod(m, void *), th, m, tp->rcv_nxt,
774 		    tp->snd_nxt, TH_ACK);
775 		tp->last_ack_sent = tp->rcv_nxt;
776 		m = NULL;
777 		*ret_val = 0;
778 		ctf_do_drop(m, NULL);
779 	}
780 }
781 
782 /*
783  * ctf_ts_check returns 1 for you should not proceed, the state
784  * machine should return. It places in ret_val what should
785  * be returned 1/0 by the caller (hpts_do_segment). The 1 indicates
786  * that the TCB is unlocked and probably dropped. The 0 indicates the
787  * TCB is still valid and locked.
788  */
789 int
ctf_ts_check(struct mbuf * m,struct tcphdr * th,struct tcpcb * tp,int32_t tlen,int32_t thflags,int32_t * ret_val)790 ctf_ts_check(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp,
791     int32_t tlen, int32_t thflags, int32_t * ret_val)
792 {
793 
794 	if (tcp_ts_getticks() - tp->ts_recent_age > TCP_PAWS_IDLE) {
795 		/*
796 		 * Invalidate ts_recent.  If this segment updates ts_recent,
797 		 * the age will be reset later and ts_recent will get a
798 		 * valid value.  If it does not, setting ts_recent to zero
799 		 * will at least satisfy the requirement that zero be placed
800 		 * in the timestamp echo reply when ts_recent isn't valid.
801 		 * The age isn't reset until we get a valid ts_recent
802 		 * because we don't want out-of-order segments to be dropped
803 		 * when ts_recent is old.
804 		 */
805 		tp->ts_recent = 0;
806 	} else {
807 		KMOD_TCPSTAT_INC(tcps_rcvduppack);
808 		KMOD_TCPSTAT_ADD(tcps_rcvdupbyte, tlen);
809 		KMOD_TCPSTAT_INC(tcps_pawsdrop);
810 		*ret_val = 0;
811 		if (tlen) {
812 			ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val);
813 		} else {
814 			ctf_do_drop(m, NULL);
815 		}
816 		return (1);
817 	}
818 	return (0);
819 }
820 
821 int
ctf_ts_check_ac(struct tcpcb * tp,int32_t thflags)822 ctf_ts_check_ac(struct tcpcb *tp, int32_t thflags)
823 {
824 
825 	if (tcp_ts_getticks() - tp->ts_recent_age > TCP_PAWS_IDLE) {
826 		/*
827 		 * Invalidate ts_recent.  If this segment updates ts_recent,
828 		 * the age will be reset later and ts_recent will get a
829 		 * valid value.  If it does not, setting ts_recent to zero
830 		 * will at least satisfy the requirement that zero be placed
831 		 * in the timestamp echo reply when ts_recent isn't valid.
832 		 * The age isn't reset until we get a valid ts_recent
833 		 * because we don't want out-of-order segments to be dropped
834 		 * when ts_recent is old.
835 		 */
836 		tp->ts_recent = 0;
837 	} else {
838 		KMOD_TCPSTAT_INC(tcps_rcvduppack);
839 		KMOD_TCPSTAT_INC(tcps_pawsdrop);
840 		return (1);
841 	}
842 	return (0);
843 }
844 
845 
846 
847 void
ctf_calc_rwin(struct socket * so,struct tcpcb * tp)848 ctf_calc_rwin(struct socket *so, struct tcpcb *tp)
849 {
850 	int32_t win;
851 
852 	/*
853 	 * Calculate amount of space in receive window, and then do TCP
854 	 * input processing. Receive window is amount of space in rcv queue,
855 	 * but not less than advertised window.
856 	 */
857 	win = sbspace(&so->so_rcv);
858 	if (win < 0)
859 		win = 0;
860 	tp->rcv_wnd = imax(win, (int)(tp->rcv_adv - tp->rcv_nxt));
861 }
862 
863 void
ctf_do_dropwithreset_conn(struct mbuf * m,struct tcpcb * tp,struct tcphdr * th,int32_t tlen)864 ctf_do_dropwithreset_conn(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th,
865     int32_t tlen)
866 {
867 
868 	tcp_dropwithreset(m, th, tp, tlen);
869 	tp = tcp_drop(tp, ETIMEDOUT);
870 	if (tp)
871 		INP_WUNLOCK(tptoinpcb(tp));
872 }
873 
874 uint32_t
ctf_fixed_maxseg(struct tcpcb * tp)875 ctf_fixed_maxseg(struct tcpcb *tp)
876 {
877 	return (tcp_fixed_maxseg(tp));
878 }
879 
880 void
ctf_log_sack_filter(struct tcpcb * tp,int num_sack_blks,struct sackblk * sack_blocks)881 ctf_log_sack_filter(struct tcpcb *tp, int num_sack_blks, struct sackblk *sack_blocks)
882 {
883 	if (tcp_bblogging_on(tp)) {
884 		union tcp_log_stackspecific log;
885 		struct timeval tv;
886 
887 		memset(&log, 0, sizeof(log));
888 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
889 		log.u_bbr.flex8 = num_sack_blks;
890 		if (num_sack_blks > 0) {
891 			log.u_bbr.flex1 = sack_blocks[0].start;
892 			log.u_bbr.flex2 = sack_blocks[0].end;
893 		}
894 		if (num_sack_blks > 1) {
895 			log.u_bbr.flex3 = sack_blocks[1].start;
896 			log.u_bbr.flex4 = sack_blocks[1].end;
897 		}
898 		if (num_sack_blks > 2) {
899 			log.u_bbr.flex5 = sack_blocks[2].start;
900 			log.u_bbr.flex6 = sack_blocks[2].end;
901 		}
902 		if (num_sack_blks > 3) {
903 			log.u_bbr.applimited = sack_blocks[3].start;
904 			log.u_bbr.pkts_out = sack_blocks[3].end;
905 		}
906 		TCP_LOG_EVENTP(tp, NULL,
907 		    &tptosocket(tp)->so_rcv,
908 		    &tptosocket(tp)->so_snd,
909 		    TCP_SACK_FILTER_RES, 0,
910 		    0, &log, false, &tv);
911 	}
912 }
913 
914 uint32_t
ctf_decay_count(uint32_t count,uint32_t decay)915 ctf_decay_count(uint32_t count, uint32_t decay)
916 {
917 	/*
918 	 * Given a count, decay it by a set percentage. The
919 	 * percentage is in thousands i.e. 100% = 1000,
920 	 * 19.3% = 193.
921 	 */
922 	uint64_t perc_count, decay_per;
923 	uint32_t decayed_count;
924 	if (decay > 1000) {
925 		/* We don't raise it */
926 		return (count);
927 	}
928 	perc_count = count;
929 	decay_per = decay;
930 	perc_count *= decay_per;
931 	perc_count /= 1000;
932 	/*
933 	 * So now perc_count holds the
934 	 * count decay value.
935 	 */
936 	decayed_count = count - (uint32_t)perc_count;
937 	return(decayed_count);
938 }
939 
940 int32_t
ctf_progress_timeout_check(struct tcpcb * tp,bool log)941 ctf_progress_timeout_check(struct tcpcb *tp, bool log)
942 {
943 	if (tp->t_maxunacktime && tp->t_acktime && TSTMP_GT(ticks, tp->t_acktime)) {
944 		if ((ticks - tp->t_acktime) >= tp->t_maxunacktime) {
945 			/*
946 			 * There is an assumption that the caller
947 			 * will drop the connection so we will
948 			 * increment the counters here.
949 			 */
950 			if (log)
951 				tcp_log_end_status(tp, TCP_EI_STATUS_PROGRESS);
952 #ifdef NETFLIX_STATS
953 			KMOD_TCPSTAT_INC(tcps_progdrops);
954 #endif
955 			return (1);
956 		}
957 	}
958 	return (0);
959 }
960