• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * INET		An implementation of the TCP/IP protocol suite for the LINUX
3  *		operating system.  INET is implemented using the  BSD Socket
4  *		interface as the means of communication with the user level.
5  *
6  *		Implementation of the Transmission Control Protocol(TCP).
7  *
8  * Authors:	Ross Biro
9  *		Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10  *		Mark Evans, <evansmp@uhura.aston.ac.uk>
11  *		Corey Minyard <wf-rch!minyard@relay.EU.net>
12  *		Florian La Roche, <flla@stud.uni-sb.de>
13  *		Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14  *		Linus Torvalds, <torvalds@cs.helsinki.fi>
15  *		Alan Cox, <gw4pts@gw4pts.ampr.org>
16  *		Matthew Dillon, <dillon@apollo.west.oic.com>
17  *		Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18  *		Jorge Cwik, <jorge@laser.satlink.net>
19  */
20 
21 /*
22  * Changes:	Pedro Roque	:	Retransmit queue handled by TCP.
23  *				:	Fragmentation on mtu decrease
24  *				:	Segment collapse on retransmit
25  *				:	AF independence
26  *
27  *		Linus Torvalds	:	send_delayed_ack
28  *		David S. Miller	:	Charge memory using the right skb
29  *					during syn/ack processing.
30  *		David S. Miller :	Output engine completely rewritten.
31  *		Andrea Arcangeli:	SYNACK carry ts_recent in tsecr.
32  *		Cacophonix Gaul :	draft-minshall-nagle-01
33  *		J Hadi Salim	:	ECN support
34  *
35  */
36 
37 #define pr_fmt(fmt) "TCP: " fmt
38 
39 #include <net/tcp.h>
40 
41 #include <linux/compiler.h>
42 #include <linux/gfp.h>
43 #include <linux/module.h>
44 
45 /* People can turn this off for buggy TCP's found in printers etc. */
46 int sysctl_tcp_retrans_collapse __read_mostly = 1;
47 
48 /* People can turn this on to work with those rare, broken TCPs that
49  * interpret the window field as a signed quantity.
50  */
51 int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
52 
53 /* Default TSQ limit of two TSO segments */
54 int sysctl_tcp_limit_output_bytes __read_mostly = 131072;
55 
56 /* This limits the percentage of the congestion window which we
57  * will allow a single TSO frame to consume.  Building TSO frames
58  * which are too large can cause TCP streams to be bursty.
59  */
60 int sysctl_tcp_tso_win_divisor __read_mostly = 3;
61 
62 int sysctl_tcp_mtu_probing __read_mostly = 0;
63 int sysctl_tcp_base_mss __read_mostly = TCP_BASE_MSS;
64 
65 /* By default, RFC2861 behavior.  */
66 int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
67 
68 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
69 			   int push_one, gfp_t gfp);
70 
71 /* Account for new data that has been sent to the network. */
tcp_event_new_data_sent(struct sock * sk,const struct sk_buff * skb)72 static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb)
73 {
74 	struct inet_connection_sock *icsk = inet_csk(sk);
75 	struct tcp_sock *tp = tcp_sk(sk);
76 	unsigned int prior_packets = tp->packets_out;
77 
78 	tcp_advance_send_head(sk, skb);
79 	tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
80 
81 	tp->packets_out += tcp_skb_pcount(skb);
82 	if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
83 	    icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
84 		tcp_rearm_rto(sk);
85 	}
86 }
87 
88 /* SND.NXT, if window was not shrunk.
89  * If window has been shrunk, what should we make? It is not clear at all.
90  * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
91  * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
92  * invalid. OK, let's make this for now:
93  */
tcp_acceptable_seq(const struct sock * sk)94 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
95 {
96 	const struct tcp_sock *tp = tcp_sk(sk);
97 
98 	if (!before(tcp_wnd_end(tp), tp->snd_nxt))
99 		return tp->snd_nxt;
100 	else
101 		return tcp_wnd_end(tp);
102 }
103 
104 /* Calculate mss to advertise in SYN segment.
105  * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
106  *
107  * 1. It is independent of path mtu.
108  * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
109  * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
110  *    attached devices, because some buggy hosts are confused by
111  *    large MSS.
112  * 4. We do not make 3, we advertise MSS, calculated from first
113  *    hop device mtu, but allow to raise it to ip_rt_min_advmss.
114  *    This may be overridden via information stored in routing table.
115  * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
116  *    probably even Jumbo".
117  */
tcp_advertise_mss(struct sock * sk)118 static __u16 tcp_advertise_mss(struct sock *sk)
119 {
120 	struct tcp_sock *tp = tcp_sk(sk);
121 	const struct dst_entry *dst = __sk_dst_get(sk);
122 	int mss = tp->advmss;
123 
124 	if (dst) {
125 		unsigned int metric = dst_metric_advmss(dst);
126 
127 		if (metric < mss) {
128 			mss = metric;
129 			tp->advmss = mss;
130 		}
131 	}
132 
133 	return (__u16)mss;
134 }
135 
136 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
137  * This is the first part of cwnd validation mechanism. */
tcp_cwnd_restart(struct sock * sk,const struct dst_entry * dst)138 static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst)
139 {
140 	struct tcp_sock *tp = tcp_sk(sk);
141 	s32 delta = tcp_time_stamp - tp->lsndtime;
142 	u32 restart_cwnd = tcp_init_cwnd(tp, dst);
143 	u32 cwnd = tp->snd_cwnd;
144 
145 	tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
146 
147 	tp->snd_ssthresh = tcp_current_ssthresh(sk);
148 	restart_cwnd = min(restart_cwnd, cwnd);
149 
150 	while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
151 		cwnd >>= 1;
152 	tp->snd_cwnd = max(cwnd, restart_cwnd);
153 	tp->snd_cwnd_stamp = tcp_time_stamp;
154 	tp->snd_cwnd_used = 0;
155 }
156 
157 /* Congestion state accounting after a packet has been sent. */
tcp_event_data_sent(struct tcp_sock * tp,struct sock * sk)158 static void tcp_event_data_sent(struct tcp_sock *tp,
159 				struct sock *sk)
160 {
161 	struct inet_connection_sock *icsk = inet_csk(sk);
162 	const u32 now = tcp_time_stamp;
163 
164 	if (sysctl_tcp_slow_start_after_idle &&
165 	    (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
166 		tcp_cwnd_restart(sk, __sk_dst_get(sk));
167 
168 	tp->lsndtime = now;
169 
170 	/* If it is a reply for ato after last received
171 	 * packet, enter pingpong mode.
172 	 */
173 	if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
174 		icsk->icsk_ack.pingpong = 1;
175 }
176 
177 /* Account for an ACK we sent. */
tcp_event_ack_sent(struct sock * sk,unsigned int pkts)178 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
179 {
180 	tcp_dec_quickack_mode(sk, pkts);
181 	inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
182 }
183 
184 /* Determine a window scaling and initial window to offer.
185  * Based on the assumption that the given amount of space
186  * will be offered. Store the results in the tp structure.
187  * NOTE: for smooth operation initial space offering should
188  * be a multiple of mss if possible. We assume here that mss >= 1.
189  * This MUST be enforced by all callers.
190  */
tcp_select_initial_window(int __space,__u32 mss,__u32 * rcv_wnd,__u32 * window_clamp,int wscale_ok,__u8 * rcv_wscale,__u32 init_rcv_wnd)191 void tcp_select_initial_window(int __space, __u32 mss,
192 			       __u32 *rcv_wnd, __u32 *window_clamp,
193 			       int wscale_ok, __u8 *rcv_wscale,
194 			       __u32 init_rcv_wnd)
195 {
196 	unsigned int space = (__space < 0 ? 0 : __space);
197 
198 	/* If no clamp set the clamp to the max possible scaled window */
199 	if (*window_clamp == 0)
200 		(*window_clamp) = (65535 << 14);
201 	space = min(*window_clamp, space);
202 
203 	/* Quantize space offering to a multiple of mss if possible. */
204 	if (space > mss)
205 		space = (space / mss) * mss;
206 
207 	/* NOTE: offering an initial window larger than 32767
208 	 * will break some buggy TCP stacks. If the admin tells us
209 	 * it is likely we could be speaking with such a buggy stack
210 	 * we will truncate our initial window offering to 32K-1
211 	 * unless the remote has sent us a window scaling option,
212 	 * which we interpret as a sign the remote TCP is not
213 	 * misinterpreting the window field as a signed quantity.
214 	 */
215 	if (sysctl_tcp_workaround_signed_windows)
216 		(*rcv_wnd) = min(space, MAX_TCP_WINDOW);
217 	else
218 		(*rcv_wnd) = space;
219 
220 	(*rcv_wscale) = 0;
221 	if (wscale_ok) {
222 		/* Set window scaling on max possible window
223 		 * See RFC1323 for an explanation of the limit to 14
224 		 */
225 		space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
226 		space = min_t(u32, space, *window_clamp);
227 		while (space > 65535 && (*rcv_wscale) < 14) {
228 			space >>= 1;
229 			(*rcv_wscale)++;
230 		}
231 	}
232 
233 	/* Set initial window to a value enough for senders starting with
234 	 * initial congestion window of sysctl_tcp_default_init_rwnd. Place
235 	 * a limit on the initial window when mss is larger than 1460.
236 	 */
237 	if (mss > (1 << *rcv_wscale)) {
238 		int init_cwnd = sysctl_tcp_default_init_rwnd;
239 		if (mss > 1460)
240 			init_cwnd = max_t(u32, (1460 * init_cwnd) / mss, 2);
241 		/* when initializing use the value from init_rcv_wnd
242 		 * rather than the default from above
243 		 */
244 		if (init_rcv_wnd)
245 			*rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
246 		else
247 			*rcv_wnd = min(*rcv_wnd, init_cwnd * mss);
248 	}
249 
250 	/* Set the clamp no higher than max representable value */
251 	(*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
252 }
253 EXPORT_SYMBOL(tcp_select_initial_window);
254 
255 /* Chose a new window to advertise, update state in tcp_sock for the
256  * socket, and return result with RFC1323 scaling applied.  The return
257  * value can be stuffed directly into th->window for an outgoing
258  * frame.
259  */
tcp_select_window(struct sock * sk)260 static u16 tcp_select_window(struct sock *sk)
261 {
262 	struct tcp_sock *tp = tcp_sk(sk);
263 	u32 cur_win = tcp_receive_window(tp);
264 	u32 new_win = __tcp_select_window(sk);
265 
266 	/* Never shrink the offered window */
267 	if (new_win < cur_win) {
268 		/* Danger Will Robinson!
269 		 * Don't update rcv_wup/rcv_wnd here or else
270 		 * we will not be able to advertise a zero
271 		 * window in time.  --DaveM
272 		 *
273 		 * Relax Will Robinson.
274 		 */
275 		new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
276 	}
277 	tp->rcv_wnd = new_win;
278 	tp->rcv_wup = tp->rcv_nxt;
279 
280 	/* Make sure we do not exceed the maximum possible
281 	 * scaled window.
282 	 */
283 	if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
284 		new_win = min(new_win, MAX_TCP_WINDOW);
285 	else
286 		new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
287 
288 	/* RFC1323 scaling applied */
289 	new_win >>= tp->rx_opt.rcv_wscale;
290 
291 	/* If we advertise zero window, disable fast path. */
292 	if (new_win == 0)
293 		tp->pred_flags = 0;
294 
295 	return new_win;
296 }
297 
298 /* Packet ECN state for a SYN-ACK */
TCP_ECN_send_synack(const struct tcp_sock * tp,struct sk_buff * skb)299 static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb)
300 {
301 	TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
302 	if (!(tp->ecn_flags & TCP_ECN_OK))
303 		TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
304 }
305 
306 /* Packet ECN state for a SYN.  */
TCP_ECN_send_syn(struct sock * sk,struct sk_buff * skb)307 static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
308 {
309 	struct tcp_sock *tp = tcp_sk(sk);
310 
311 	tp->ecn_flags = 0;
312 	if (sock_net(sk)->ipv4.sysctl_tcp_ecn == 1) {
313 		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
314 		tp->ecn_flags = TCP_ECN_OK;
315 	}
316 }
317 
318 static __inline__ void
TCP_ECN_make_synack(const struct request_sock * req,struct tcphdr * th)319 TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th)
320 {
321 	if (inet_rsk(req)->ecn_ok)
322 		th->ece = 1;
323 }
324 
325 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
326  * be sent.
327  */
TCP_ECN_send(struct sock * sk,struct sk_buff * skb,int tcp_header_len)328 static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
329 				int tcp_header_len)
330 {
331 	struct tcp_sock *tp = tcp_sk(sk);
332 
333 	if (tp->ecn_flags & TCP_ECN_OK) {
334 		/* Not-retransmitted data segment: set ECT and inject CWR. */
335 		if (skb->len != tcp_header_len &&
336 		    !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
337 			INET_ECN_xmit(sk);
338 			if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
339 				tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
340 				tcp_hdr(skb)->cwr = 1;
341 				skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
342 			}
343 		} else {
344 			/* ACK or retransmitted segment: clear ECT|CE */
345 			INET_ECN_dontxmit(sk);
346 		}
347 		if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
348 			tcp_hdr(skb)->ece = 1;
349 	}
350 }
351 
352 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
353  * auto increment end seqno.
354  */
tcp_init_nondata_skb(struct sk_buff * skb,u32 seq,u8 flags)355 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
356 {
357 	skb->ip_summed = CHECKSUM_PARTIAL;
358 	skb->csum = 0;
359 
360 	TCP_SKB_CB(skb)->tcp_flags = flags;
361 	TCP_SKB_CB(skb)->sacked = 0;
362 
363 	skb_shinfo(skb)->gso_segs = 1;
364 	skb_shinfo(skb)->gso_size = 0;
365 	skb_shinfo(skb)->gso_type = 0;
366 
367 	TCP_SKB_CB(skb)->seq = seq;
368 	if (flags & (TCPHDR_SYN | TCPHDR_FIN))
369 		seq++;
370 	TCP_SKB_CB(skb)->end_seq = seq;
371 }
372 
tcp_urg_mode(const struct tcp_sock * tp)373 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
374 {
375 	return tp->snd_una != tp->snd_up;
376 }
377 
378 #define OPTION_SACK_ADVERTISE	(1 << 0)
379 #define OPTION_TS		(1 << 1)
380 #define OPTION_MD5		(1 << 2)
381 #define OPTION_WSCALE		(1 << 3)
382 #define OPTION_FAST_OPEN_COOKIE	(1 << 8)
383 
384 struct tcp_out_options {
385 	u16 options;		/* bit field of OPTION_* */
386 	u16 mss;		/* 0 to disable */
387 	u8 ws;			/* window scale, 0 to disable */
388 	u8 num_sack_blocks;	/* number of SACK blocks to include */
389 	u8 hash_size;		/* bytes in hash_location */
390 	__u8 *hash_location;	/* temporary pointer, overloaded */
391 	__u32 tsval, tsecr;	/* need to include OPTION_TS */
392 	struct tcp_fastopen_cookie *fastopen_cookie;	/* Fast open cookie */
393 };
394 
395 /* Write previously computed TCP options to the packet.
396  *
397  * Beware: Something in the Internet is very sensitive to the ordering of
398  * TCP options, we learned this through the hard way, so be careful here.
399  * Luckily we can at least blame others for their non-compliance but from
400  * inter-operatibility perspective it seems that we're somewhat stuck with
401  * the ordering which we have been using if we want to keep working with
402  * those broken things (not that it currently hurts anybody as there isn't
403  * particular reason why the ordering would need to be changed).
404  *
405  * At least SACK_PERM as the first option is known to lead to a disaster
406  * (but it may well be that other scenarios fail similarly).
407  */
tcp_options_write(__be32 * ptr,struct tcp_sock * tp,struct tcp_out_options * opts)408 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
409 			      struct tcp_out_options *opts)
410 {
411 	u16 options = opts->options;	/* mungable copy */
412 
413 	if (unlikely(OPTION_MD5 & options)) {
414 		*ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
415 			       (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
416 		/* overload cookie hash location */
417 		opts->hash_location = (__u8 *)ptr;
418 		ptr += 4;
419 	}
420 
421 	if (unlikely(opts->mss)) {
422 		*ptr++ = htonl((TCPOPT_MSS << 24) |
423 			       (TCPOLEN_MSS << 16) |
424 			       opts->mss);
425 	}
426 
427 	if (likely(OPTION_TS & options)) {
428 		if (unlikely(OPTION_SACK_ADVERTISE & options)) {
429 			*ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
430 				       (TCPOLEN_SACK_PERM << 16) |
431 				       (TCPOPT_TIMESTAMP << 8) |
432 				       TCPOLEN_TIMESTAMP);
433 			options &= ~OPTION_SACK_ADVERTISE;
434 		} else {
435 			*ptr++ = htonl((TCPOPT_NOP << 24) |
436 				       (TCPOPT_NOP << 16) |
437 				       (TCPOPT_TIMESTAMP << 8) |
438 				       TCPOLEN_TIMESTAMP);
439 		}
440 		*ptr++ = htonl(opts->tsval);
441 		*ptr++ = htonl(opts->tsecr);
442 	}
443 
444 	if (unlikely(OPTION_SACK_ADVERTISE & options)) {
445 		*ptr++ = htonl((TCPOPT_NOP << 24) |
446 			       (TCPOPT_NOP << 16) |
447 			       (TCPOPT_SACK_PERM << 8) |
448 			       TCPOLEN_SACK_PERM);
449 	}
450 
451 	if (unlikely(OPTION_WSCALE & options)) {
452 		*ptr++ = htonl((TCPOPT_NOP << 24) |
453 			       (TCPOPT_WINDOW << 16) |
454 			       (TCPOLEN_WINDOW << 8) |
455 			       opts->ws);
456 	}
457 
458 	if (unlikely(opts->num_sack_blocks)) {
459 		struct tcp_sack_block *sp = tp->rx_opt.dsack ?
460 			tp->duplicate_sack : tp->selective_acks;
461 		int this_sack;
462 
463 		*ptr++ = htonl((TCPOPT_NOP  << 24) |
464 			       (TCPOPT_NOP  << 16) |
465 			       (TCPOPT_SACK <<  8) |
466 			       (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
467 						     TCPOLEN_SACK_PERBLOCK)));
468 
469 		for (this_sack = 0; this_sack < opts->num_sack_blocks;
470 		     ++this_sack) {
471 			*ptr++ = htonl(sp[this_sack].start_seq);
472 			*ptr++ = htonl(sp[this_sack].end_seq);
473 		}
474 
475 		tp->rx_opt.dsack = 0;
476 	}
477 
478 	if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
479 		struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
480 
481 		*ptr++ = htonl((TCPOPT_EXP << 24) |
482 			       ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) |
483 			       TCPOPT_FASTOPEN_MAGIC);
484 
485 		memcpy(ptr, foc->val, foc->len);
486 		if ((foc->len & 3) == 2) {
487 			u8 *align = ((u8 *)ptr) + foc->len;
488 			align[0] = align[1] = TCPOPT_NOP;
489 		}
490 		ptr += (foc->len + 3) >> 2;
491 	}
492 }
493 
494 /* Compute TCP options for SYN packets. This is not the final
495  * network wire format yet.
496  */
tcp_syn_options(struct sock * sk,struct sk_buff * skb,struct tcp_out_options * opts,struct tcp_md5sig_key ** md5)497 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
498 				struct tcp_out_options *opts,
499 				struct tcp_md5sig_key **md5)
500 {
501 	struct tcp_sock *tp = tcp_sk(sk);
502 	unsigned int remaining = MAX_TCP_OPTION_SPACE;
503 	struct tcp_fastopen_request *fastopen = tp->fastopen_req;
504 
505 #ifdef CONFIG_TCP_MD5SIG
506 	*md5 = tp->af_specific->md5_lookup(sk, sk);
507 	if (*md5) {
508 		opts->options |= OPTION_MD5;
509 		remaining -= TCPOLEN_MD5SIG_ALIGNED;
510 	}
511 #else
512 	*md5 = NULL;
513 #endif
514 
515 	/* We always get an MSS option.  The option bytes which will be seen in
516 	 * normal data packets should timestamps be used, must be in the MSS
517 	 * advertised.  But we subtract them from tp->mss_cache so that
518 	 * calculations in tcp_sendmsg are simpler etc.  So account for this
519 	 * fact here if necessary.  If we don't do this correctly, as a
520 	 * receiver we won't recognize data packets as being full sized when we
521 	 * should, and thus we won't abide by the delayed ACK rules correctly.
522 	 * SACKs don't matter, we never delay an ACK when we have any of those
523 	 * going out.  */
524 	opts->mss = tcp_advertise_mss(sk);
525 	remaining -= TCPOLEN_MSS_ALIGNED;
526 
527 	if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
528 		opts->options |= OPTION_TS;
529 		opts->tsval = TCP_SKB_CB(skb)->when + tp->tsoffset;
530 		opts->tsecr = tp->rx_opt.ts_recent;
531 		remaining -= TCPOLEN_TSTAMP_ALIGNED;
532 	}
533 	if (likely(sysctl_tcp_window_scaling)) {
534 		opts->ws = tp->rx_opt.rcv_wscale;
535 		opts->options |= OPTION_WSCALE;
536 		remaining -= TCPOLEN_WSCALE_ALIGNED;
537 	}
538 	if (likely(sysctl_tcp_sack)) {
539 		opts->options |= OPTION_SACK_ADVERTISE;
540 		if (unlikely(!(OPTION_TS & opts->options)))
541 			remaining -= TCPOLEN_SACKPERM_ALIGNED;
542 	}
543 
544 	if (fastopen && fastopen->cookie.len >= 0) {
545 		u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len;
546 		need = (need + 3) & ~3U;  /* Align to 32 bits */
547 		if (remaining >= need) {
548 			opts->options |= OPTION_FAST_OPEN_COOKIE;
549 			opts->fastopen_cookie = &fastopen->cookie;
550 			remaining -= need;
551 			tp->syn_fastopen = 1;
552 		}
553 	}
554 
555 	return MAX_TCP_OPTION_SPACE - remaining;
556 }
557 
558 /* Set up TCP options for SYN-ACKs. */
tcp_synack_options(struct sock * sk,struct request_sock * req,unsigned int mss,struct sk_buff * skb,struct tcp_out_options * opts,struct tcp_md5sig_key ** md5,struct tcp_fastopen_cookie * foc)559 static unsigned int tcp_synack_options(struct sock *sk,
560 				   struct request_sock *req,
561 				   unsigned int mss, struct sk_buff *skb,
562 				   struct tcp_out_options *opts,
563 				   struct tcp_md5sig_key **md5,
564 				   struct tcp_fastopen_cookie *foc)
565 {
566 	struct inet_request_sock *ireq = inet_rsk(req);
567 	unsigned int remaining = MAX_TCP_OPTION_SPACE;
568 
569 #ifdef CONFIG_TCP_MD5SIG
570 	*md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
571 	if (*md5) {
572 		opts->options |= OPTION_MD5;
573 		remaining -= TCPOLEN_MD5SIG_ALIGNED;
574 
575 		/* We can't fit any SACK blocks in a packet with MD5 + TS
576 		 * options. There was discussion about disabling SACK
577 		 * rather than TS in order to fit in better with old,
578 		 * buggy kernels, but that was deemed to be unnecessary.
579 		 */
580 		ireq->tstamp_ok &= !ireq->sack_ok;
581 	}
582 #else
583 	*md5 = NULL;
584 #endif
585 
586 	/* We always send an MSS option. */
587 	opts->mss = mss;
588 	remaining -= TCPOLEN_MSS_ALIGNED;
589 
590 	if (likely(ireq->wscale_ok)) {
591 		opts->ws = ireq->rcv_wscale;
592 		opts->options |= OPTION_WSCALE;
593 		remaining -= TCPOLEN_WSCALE_ALIGNED;
594 	}
595 	if (likely(ireq->tstamp_ok)) {
596 		opts->options |= OPTION_TS;
597 		opts->tsval = TCP_SKB_CB(skb)->when;
598 		opts->tsecr = req->ts_recent;
599 		remaining -= TCPOLEN_TSTAMP_ALIGNED;
600 	}
601 	if (likely(ireq->sack_ok)) {
602 		opts->options |= OPTION_SACK_ADVERTISE;
603 		if (unlikely(!ireq->tstamp_ok))
604 			remaining -= TCPOLEN_SACKPERM_ALIGNED;
605 	}
606 	if (foc != NULL) {
607 		u32 need = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
608 		need = (need + 3) & ~3U;  /* Align to 32 bits */
609 		if (remaining >= need) {
610 			opts->options |= OPTION_FAST_OPEN_COOKIE;
611 			opts->fastopen_cookie = foc;
612 			remaining -= need;
613 		}
614 	}
615 
616 	return MAX_TCP_OPTION_SPACE - remaining;
617 }
618 
619 /* Compute TCP options for ESTABLISHED sockets. This is not the
620  * final wire format yet.
621  */
tcp_established_options(struct sock * sk,struct sk_buff * skb,struct tcp_out_options * opts,struct tcp_md5sig_key ** md5)622 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
623 					struct tcp_out_options *opts,
624 					struct tcp_md5sig_key **md5)
625 {
626 	struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
627 	struct tcp_sock *tp = tcp_sk(sk);
628 	unsigned int size = 0;
629 	unsigned int eff_sacks;
630 
631 #ifdef CONFIG_TCP_MD5SIG
632 	*md5 = tp->af_specific->md5_lookup(sk, sk);
633 	if (unlikely(*md5)) {
634 		opts->options |= OPTION_MD5;
635 		size += TCPOLEN_MD5SIG_ALIGNED;
636 	}
637 #else
638 	*md5 = NULL;
639 #endif
640 
641 	if (likely(tp->rx_opt.tstamp_ok)) {
642 		opts->options |= OPTION_TS;
643 		opts->tsval = tcb ? tcb->when + tp->tsoffset : 0;
644 		opts->tsecr = tp->rx_opt.ts_recent;
645 		size += TCPOLEN_TSTAMP_ALIGNED;
646 	}
647 
648 	eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
649 	if (unlikely(eff_sacks)) {
650 		const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
651 		opts->num_sack_blocks =
652 			min_t(unsigned int, eff_sacks,
653 			      (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
654 			      TCPOLEN_SACK_PERBLOCK);
655 		size += TCPOLEN_SACK_BASE_ALIGNED +
656 			opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
657 	}
658 
659 	return size;
660 }
661 
662 
663 /* TCP SMALL QUEUES (TSQ)
664  *
665  * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
666  * to reduce RTT and bufferbloat.
667  * We do this using a special skb destructor (tcp_wfree).
668  *
669  * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
670  * needs to be reallocated in a driver.
671  * The invariant being skb->truesize substracted from sk->sk_wmem_alloc
672  *
673  * Since transmit from skb destructor is forbidden, we use a tasklet
674  * to process all sockets that eventually need to send more skbs.
675  * We use one tasklet per cpu, with its own queue of sockets.
676  */
677 struct tsq_tasklet {
678 	struct tasklet_struct	tasklet;
679 	struct list_head	head; /* queue of tcp sockets */
680 };
681 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
682 
tcp_tsq_handler(struct sock * sk)683 static void tcp_tsq_handler(struct sock *sk)
684 {
685 	if ((1 << sk->sk_state) &
686 	    (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
687 	     TCPF_CLOSE_WAIT  | TCPF_LAST_ACK))
688 		tcp_write_xmit(sk, tcp_current_mss(sk), 0, 0, GFP_ATOMIC);
689 }
690 /*
691  * One tasklest per cpu tries to send more skbs.
692  * We run in tasklet context but need to disable irqs when
693  * transfering tsq->head because tcp_wfree() might
694  * interrupt us (non NAPI drivers)
695  */
tcp_tasklet_func(unsigned long data)696 static void tcp_tasklet_func(unsigned long data)
697 {
698 	struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
699 	LIST_HEAD(list);
700 	unsigned long flags;
701 	struct list_head *q, *n;
702 	struct tcp_sock *tp;
703 	struct sock *sk;
704 
705 	local_irq_save(flags);
706 	list_splice_init(&tsq->head, &list);
707 	local_irq_restore(flags);
708 
709 	list_for_each_safe(q, n, &list) {
710 		tp = list_entry(q, struct tcp_sock, tsq_node);
711 		list_del(&tp->tsq_node);
712 
713 		sk = (struct sock *)tp;
714 		bh_lock_sock(sk);
715 
716 		if (!sock_owned_by_user(sk)) {
717 			tcp_tsq_handler(sk);
718 		} else {
719 			/* defer the work to tcp_release_cb() */
720 			set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
721 		}
722 		bh_unlock_sock(sk);
723 
724 		clear_bit(TSQ_QUEUED, &tp->tsq_flags);
725 		sk_free(sk);
726 	}
727 }
728 
729 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) |		\
730 			  (1UL << TCP_WRITE_TIMER_DEFERRED) |	\
731 			  (1UL << TCP_DELACK_TIMER_DEFERRED) |	\
732 			  (1UL << TCP_MTU_REDUCED_DEFERRED))
733 /**
734  * tcp_release_cb - tcp release_sock() callback
735  * @sk: socket
736  *
737  * called from release_sock() to perform protocol dependent
738  * actions before socket release.
739  */
tcp_release_cb(struct sock * sk)740 void tcp_release_cb(struct sock *sk)
741 {
742 	struct tcp_sock *tp = tcp_sk(sk);
743 	unsigned long flags, nflags;
744 
745 	/* perform an atomic operation only if at least one flag is set */
746 	do {
747 		flags = tp->tsq_flags;
748 		if (!(flags & TCP_DEFERRED_ALL))
749 			return;
750 		nflags = flags & ~TCP_DEFERRED_ALL;
751 	} while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
752 
753 	if (flags & (1UL << TCP_TSQ_DEFERRED))
754 		tcp_tsq_handler(sk);
755 
756 	if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
757 		tcp_write_timer_handler(sk);
758 		__sock_put(sk);
759 	}
760 	if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
761 		tcp_delack_timer_handler(sk);
762 		__sock_put(sk);
763 	}
764 	if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
765 		sk->sk_prot->mtu_reduced(sk);
766 		__sock_put(sk);
767 	}
768 }
769 EXPORT_SYMBOL(tcp_release_cb);
770 
tcp_tasklet_init(void)771 void __init tcp_tasklet_init(void)
772 {
773 	int i;
774 
775 	for_each_possible_cpu(i) {
776 		struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
777 
778 		INIT_LIST_HEAD(&tsq->head);
779 		tasklet_init(&tsq->tasklet,
780 			     tcp_tasklet_func,
781 			     (unsigned long)tsq);
782 	}
783 }
784 
785 /*
786  * Write buffer destructor automatically called from kfree_skb.
787  * We cant xmit new skbs from this context, as we might already
788  * hold qdisc lock.
789  */
tcp_wfree(struct sk_buff * skb)790 void tcp_wfree(struct sk_buff *skb)
791 {
792 	struct sock *sk = skb->sk;
793 	struct tcp_sock *tp = tcp_sk(sk);
794 
795 	if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
796 	    !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
797 		unsigned long flags;
798 		struct tsq_tasklet *tsq;
799 
800 		/* Keep a ref on socket.
801 		 * This last ref will be released in tcp_tasklet_func()
802 		 */
803 		atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc);
804 
805 		/* queue this socket to tasklet queue */
806 		local_irq_save(flags);
807 		tsq = &__get_cpu_var(tsq_tasklet);
808 		list_add(&tp->tsq_node, &tsq->head);
809 		tasklet_schedule(&tsq->tasklet);
810 		local_irq_restore(flags);
811 	} else {
812 		sock_wfree(skb);
813 	}
814 }
815 
816 /* This routine actually transmits TCP packets queued in by
817  * tcp_do_sendmsg().  This is used by both the initial
818  * transmission and possible later retransmissions.
819  * All SKB's seen here are completely headerless.  It is our
820  * job to build the TCP header, and pass the packet down to
821  * IP so it can do the same plus pass the packet off to the
822  * device.
823  *
824  * We are working here with either a clone of the original
825  * SKB, or a fresh unique copy made by the retransmit engine.
826  */
tcp_transmit_skb(struct sock * sk,struct sk_buff * skb,int clone_it,gfp_t gfp_mask)827 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
828 			    gfp_t gfp_mask)
829 {
830 	const struct inet_connection_sock *icsk = inet_csk(sk);
831 	struct inet_sock *inet;
832 	struct tcp_sock *tp;
833 	struct tcp_skb_cb *tcb;
834 	struct tcp_out_options opts;
835 	unsigned int tcp_options_size, tcp_header_size;
836 	struct tcp_md5sig_key *md5;
837 	struct tcphdr *th;
838 	int err;
839 
840 	BUG_ON(!skb || !tcp_skb_pcount(skb));
841 
842 	/* If congestion control is doing timestamping, we must
843 	 * take such a timestamp before we potentially clone/copy.
844 	 */
845 	if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
846 		__net_timestamp(skb);
847 
848 	if (likely(clone_it)) {
849 		const struct sk_buff *fclone = skb + 1;
850 
851 		if (unlikely(skb->fclone == SKB_FCLONE_ORIG &&
852 			     fclone->fclone == SKB_FCLONE_CLONE))
853 			NET_INC_STATS_BH(sock_net(sk),
854 					 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
855 
856 		if (unlikely(skb_cloned(skb)))
857 			skb = pskb_copy(skb, gfp_mask);
858 		else
859 			skb = skb_clone(skb, gfp_mask);
860 		if (unlikely(!skb))
861 			return -ENOBUFS;
862 	}
863 
864 	inet = inet_sk(sk);
865 	tp = tcp_sk(sk);
866 	tcb = TCP_SKB_CB(skb);
867 	memset(&opts, 0, sizeof(opts));
868 
869 	if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
870 		tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
871 	else
872 		tcp_options_size = tcp_established_options(sk, skb, &opts,
873 							   &md5);
874 	tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
875 
876 	if (tcp_packets_in_flight(tp) == 0)
877 		tcp_ca_event(sk, CA_EVENT_TX_START);
878 
879 	/* if no packet is in qdisc/device queue, then allow XPS to select
880 	 * another queue.
881 	 */
882 	skb->ooo_okay = sk_wmem_alloc_get(sk) == 0;
883 
884 	skb_push(skb, tcp_header_size);
885 	skb_reset_transport_header(skb);
886 
887 	skb_orphan(skb);
888 	skb->sk = sk;
889 	skb->destructor = (sysctl_tcp_limit_output_bytes > 0) ?
890 			  tcp_wfree : sock_wfree;
891 	atomic_add(skb->truesize, &sk->sk_wmem_alloc);
892 
893 	/* Build TCP header and checksum it. */
894 	th = tcp_hdr(skb);
895 	th->source		= inet->inet_sport;
896 	th->dest		= inet->inet_dport;
897 	th->seq			= htonl(tcb->seq);
898 	th->ack_seq		= htonl(tp->rcv_nxt);
899 	*(((__be16 *)th) + 6)	= htons(((tcp_header_size >> 2) << 12) |
900 					tcb->tcp_flags);
901 
902 	if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
903 		/* RFC1323: The window in SYN & SYN/ACK segments
904 		 * is never scaled.
905 		 */
906 		th->window	= htons(min(tp->rcv_wnd, 65535U));
907 	} else {
908 		th->window	= htons(tcp_select_window(sk));
909 	}
910 	th->check		= 0;
911 	th->urg_ptr		= 0;
912 
913 	/* The urg_mode check is necessary during a below snd_una win probe */
914 	if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
915 		if (before(tp->snd_up, tcb->seq + 0x10000)) {
916 			th->urg_ptr = htons(tp->snd_up - tcb->seq);
917 			th->urg = 1;
918 		} else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
919 			th->urg_ptr = htons(0xFFFF);
920 			th->urg = 1;
921 		}
922 	}
923 
924 	tcp_options_write((__be32 *)(th + 1), tp, &opts);
925 	if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
926 		TCP_ECN_send(sk, skb, tcp_header_size);
927 
928 #ifdef CONFIG_TCP_MD5SIG
929 	/* Calculate the MD5 hash, as we have all we need now */
930 	if (md5) {
931 		sk_nocaps_add(sk, NETIF_F_GSO_MASK);
932 		tp->af_specific->calc_md5_hash(opts.hash_location,
933 					       md5, sk, NULL, skb);
934 	}
935 #endif
936 
937 	icsk->icsk_af_ops->send_check(sk, skb);
938 
939 	if (likely(tcb->tcp_flags & TCPHDR_ACK))
940 		tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
941 
942 	if (skb->len != tcp_header_size)
943 		tcp_event_data_sent(tp, sk);
944 
945 	if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
946 		TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
947 			      tcp_skb_pcount(skb));
948 
949 	err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl);
950 	if (likely(err <= 0))
951 		return err;
952 
953 	tcp_enter_cwr(sk, 1);
954 
955 	return net_xmit_eval(err);
956 }
957 
958 /* This routine just queues the buffer for sending.
959  *
960  * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
961  * otherwise socket can stall.
962  */
tcp_queue_skb(struct sock * sk,struct sk_buff * skb)963 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
964 {
965 	struct tcp_sock *tp = tcp_sk(sk);
966 
967 	/* Advance write_seq and place onto the write_queue. */
968 	tp->write_seq = TCP_SKB_CB(skb)->end_seq;
969 	skb_header_release(skb);
970 	tcp_add_write_queue_tail(sk, skb);
971 	sk->sk_wmem_queued += skb->truesize;
972 	sk_mem_charge(sk, skb->truesize);
973 }
974 
975 /* Initialize TSO segments for a packet. */
tcp_set_skb_tso_segs(const struct sock * sk,struct sk_buff * skb,unsigned int mss_now)976 static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb,
977 				 unsigned int mss_now)
978 {
979 	if (skb->len <= mss_now || !sk_can_gso(sk) ||
980 	    skb->ip_summed == CHECKSUM_NONE) {
981 		/* Avoid the costly divide in the normal
982 		 * non-TSO case.
983 		 */
984 		skb_shinfo(skb)->gso_segs = 1;
985 		skb_shinfo(skb)->gso_size = 0;
986 		skb_shinfo(skb)->gso_type = 0;
987 	} else {
988 		skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
989 		skb_shinfo(skb)->gso_size = mss_now;
990 		skb_shinfo(skb)->gso_type = sk->sk_gso_type;
991 	}
992 }
993 
994 /* When a modification to fackets out becomes necessary, we need to check
995  * skb is counted to fackets_out or not.
996  */
tcp_adjust_fackets_out(struct sock * sk,const struct sk_buff * skb,int decr)997 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
998 				   int decr)
999 {
1000 	struct tcp_sock *tp = tcp_sk(sk);
1001 
1002 	if (!tp->sacked_out || tcp_is_reno(tp))
1003 		return;
1004 
1005 	if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1006 		tp->fackets_out -= decr;
1007 }
1008 
1009 /* Pcount in the middle of the write queue got changed, we need to do various
1010  * tweaks to fix counters
1011  */
tcp_adjust_pcount(struct sock * sk,const struct sk_buff * skb,int decr)1012 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1013 {
1014 	struct tcp_sock *tp = tcp_sk(sk);
1015 
1016 	tp->packets_out -= decr;
1017 
1018 	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1019 		tp->sacked_out -= decr;
1020 	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1021 		tp->retrans_out -= decr;
1022 	if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1023 		tp->lost_out -= decr;
1024 
1025 	/* Reno case is special. Sigh... */
1026 	if (tcp_is_reno(tp) && decr > 0)
1027 		tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1028 
1029 	tcp_adjust_fackets_out(sk, skb, decr);
1030 
1031 	if (tp->lost_skb_hint &&
1032 	    before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1033 	    (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1034 		tp->lost_cnt_hint -= decr;
1035 
1036 	tcp_verify_left_out(tp);
1037 }
1038 
1039 /* Function to create two new TCP segments.  Shrinks the given segment
1040  * to the specified size and appends a new segment with the rest of the
1041  * packet to the list.  This won't be called frequently, I hope.
1042  * Remember, these are still headerless SKBs at this point.
1043  */
tcp_fragment(struct sock * sk,struct sk_buff * skb,u32 len,unsigned int mss_now)1044 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1045 		 unsigned int mss_now)
1046 {
1047 	struct tcp_sock *tp = tcp_sk(sk);
1048 	struct sk_buff *buff;
1049 	int nsize, old_factor;
1050 	int nlen;
1051 	u8 flags;
1052 
1053 	if (WARN_ON(len > skb->len))
1054 		return -EINVAL;
1055 
1056 	nsize = skb_headlen(skb) - len;
1057 	if (nsize < 0)
1058 		nsize = 0;
1059 
1060 	if (skb_cloned(skb) &&
1061 	    skb_is_nonlinear(skb) &&
1062 	    pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1063 		return -ENOMEM;
1064 
1065 	/* Get a new skb... force flag on. */
1066 	buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
1067 	if (buff == NULL)
1068 		return -ENOMEM; /* We'll just try again later. */
1069 
1070 	sk->sk_wmem_queued += buff->truesize;
1071 	sk_mem_charge(sk, buff->truesize);
1072 	nlen = skb->len - len - nsize;
1073 	buff->truesize += nlen;
1074 	skb->truesize -= nlen;
1075 
1076 	/* Correct the sequence numbers. */
1077 	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1078 	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1079 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1080 
1081 	/* PSH and FIN should only be set in the second packet. */
1082 	flags = TCP_SKB_CB(skb)->tcp_flags;
1083 	TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1084 	TCP_SKB_CB(buff)->tcp_flags = flags;
1085 	TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1086 
1087 	if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1088 		/* Copy and checksum data tail into the new buffer. */
1089 		buff->csum = csum_partial_copy_nocheck(skb->data + len,
1090 						       skb_put(buff, nsize),
1091 						       nsize, 0);
1092 
1093 		skb_trim(skb, len);
1094 
1095 		skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1096 	} else {
1097 		skb->ip_summed = CHECKSUM_PARTIAL;
1098 		skb_split(skb, buff, len);
1099 	}
1100 
1101 	buff->ip_summed = skb->ip_summed;
1102 
1103 	/* Looks stupid, but our code really uses when of
1104 	 * skbs, which it never sent before. --ANK
1105 	 */
1106 	TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
1107 	buff->tstamp = skb->tstamp;
1108 
1109 	old_factor = tcp_skb_pcount(skb);
1110 
1111 	/* Fix up tso_factor for both original and new SKB.  */
1112 	tcp_set_skb_tso_segs(sk, skb, mss_now);
1113 	tcp_set_skb_tso_segs(sk, buff, mss_now);
1114 
1115 	/* If this packet has been sent out already, we must
1116 	 * adjust the various packet counters.
1117 	 */
1118 	if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1119 		int diff = old_factor - tcp_skb_pcount(skb) -
1120 			tcp_skb_pcount(buff);
1121 
1122 		if (diff)
1123 			tcp_adjust_pcount(sk, skb, diff);
1124 	}
1125 
1126 	/* Link BUFF into the send queue. */
1127 	skb_header_release(buff);
1128 	tcp_insert_write_queue_after(skb, buff, sk);
1129 
1130 	return 0;
1131 }
1132 
1133 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1134  * eventually). The difference is that pulled data not copied, but
1135  * immediately discarded.
1136  */
__pskb_trim_head(struct sk_buff * skb,int len)1137 static void __pskb_trim_head(struct sk_buff *skb, int len)
1138 {
1139 	int i, k, eat;
1140 
1141 	eat = min_t(int, len, skb_headlen(skb));
1142 	if (eat) {
1143 		__skb_pull(skb, eat);
1144 		len -= eat;
1145 		if (!len)
1146 			return;
1147 	}
1148 	eat = len;
1149 	k = 0;
1150 	for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1151 		int size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
1152 
1153 		if (size <= eat) {
1154 			skb_frag_unref(skb, i);
1155 			eat -= size;
1156 		} else {
1157 			skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
1158 			if (eat) {
1159 				skb_shinfo(skb)->frags[k].page_offset += eat;
1160 				skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat);
1161 				eat = 0;
1162 			}
1163 			k++;
1164 		}
1165 	}
1166 	skb_shinfo(skb)->nr_frags = k;
1167 
1168 	skb_reset_tail_pointer(skb);
1169 	skb->data_len -= len;
1170 	skb->len = skb->data_len;
1171 }
1172 
1173 /* Remove acked data from a packet in the transmit queue. */
tcp_trim_head(struct sock * sk,struct sk_buff * skb,u32 len)1174 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1175 {
1176 	if (skb_unclone(skb, GFP_ATOMIC))
1177 		return -ENOMEM;
1178 
1179 	__pskb_trim_head(skb, len);
1180 
1181 	TCP_SKB_CB(skb)->seq += len;
1182 	skb->ip_summed = CHECKSUM_PARTIAL;
1183 
1184 	skb->truesize	     -= len;
1185 	sk->sk_wmem_queued   -= len;
1186 	sk_mem_uncharge(sk, len);
1187 	sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1188 
1189 	/* Any change of skb->len requires recalculation of tso factor. */
1190 	if (tcp_skb_pcount(skb) > 1)
1191 		tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb));
1192 
1193 	return 0;
1194 }
1195 
1196 /* Calculate MSS not accounting any TCP options.  */
__tcp_mtu_to_mss(struct sock * sk,int pmtu)1197 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1198 {
1199 	const struct tcp_sock *tp = tcp_sk(sk);
1200 	const struct inet_connection_sock *icsk = inet_csk(sk);
1201 	int mss_now;
1202 
1203 	/* Calculate base mss without TCP options:
1204 	   It is MMS_S - sizeof(tcphdr) of rfc1122
1205 	 */
1206 	mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1207 
1208 	/* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1209 	if (icsk->icsk_af_ops->net_frag_header_len) {
1210 		const struct dst_entry *dst = __sk_dst_get(sk);
1211 
1212 		if (dst && dst_allfrag(dst))
1213 			mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1214 	}
1215 
1216 	/* Clamp it (mss_clamp does not include tcp options) */
1217 	if (mss_now > tp->rx_opt.mss_clamp)
1218 		mss_now = tp->rx_opt.mss_clamp;
1219 
1220 	/* Now subtract optional transport overhead */
1221 	mss_now -= icsk->icsk_ext_hdr_len;
1222 
1223 	/* Then reserve room for full set of TCP options and 8 bytes of data */
1224 	if (mss_now < 48)
1225 		mss_now = 48;
1226 	return mss_now;
1227 }
1228 
1229 /* Calculate MSS. Not accounting for SACKs here.  */
tcp_mtu_to_mss(struct sock * sk,int pmtu)1230 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1231 {
1232 	/* Subtract TCP options size, not including SACKs */
1233 	return __tcp_mtu_to_mss(sk, pmtu) -
1234 	       (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1235 }
1236 
1237 /* Inverse of above */
tcp_mss_to_mtu(struct sock * sk,int mss)1238 int tcp_mss_to_mtu(struct sock *sk, int mss)
1239 {
1240 	const struct tcp_sock *tp = tcp_sk(sk);
1241 	const struct inet_connection_sock *icsk = inet_csk(sk);
1242 	int mtu;
1243 
1244 	mtu = mss +
1245 	      tp->tcp_header_len +
1246 	      icsk->icsk_ext_hdr_len +
1247 	      icsk->icsk_af_ops->net_header_len;
1248 
1249 	/* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1250 	if (icsk->icsk_af_ops->net_frag_header_len) {
1251 		const struct dst_entry *dst = __sk_dst_get(sk);
1252 
1253 		if (dst && dst_allfrag(dst))
1254 			mtu += icsk->icsk_af_ops->net_frag_header_len;
1255 	}
1256 	return mtu;
1257 }
1258 
1259 /* MTU probing init per socket */
tcp_mtup_init(struct sock * sk)1260 void tcp_mtup_init(struct sock *sk)
1261 {
1262 	struct tcp_sock *tp = tcp_sk(sk);
1263 	struct inet_connection_sock *icsk = inet_csk(sk);
1264 
1265 	icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1266 	icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1267 			       icsk->icsk_af_ops->net_header_len;
1268 	icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1269 	icsk->icsk_mtup.probe_size = 0;
1270 }
1271 EXPORT_SYMBOL(tcp_mtup_init);
1272 
1273 /* This function synchronize snd mss to current pmtu/exthdr set.
1274 
1275    tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1276    for TCP options, but includes only bare TCP header.
1277 
1278    tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1279    It is minimum of user_mss and mss received with SYN.
1280    It also does not include TCP options.
1281 
1282    inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1283 
1284    tp->mss_cache is current effective sending mss, including
1285    all tcp options except for SACKs. It is evaluated,
1286    taking into account current pmtu, but never exceeds
1287    tp->rx_opt.mss_clamp.
1288 
1289    NOTE1. rfc1122 clearly states that advertised MSS
1290    DOES NOT include either tcp or ip options.
1291 
1292    NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1293    are READ ONLY outside this function.		--ANK (980731)
1294  */
tcp_sync_mss(struct sock * sk,u32 pmtu)1295 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1296 {
1297 	struct tcp_sock *tp = tcp_sk(sk);
1298 	struct inet_connection_sock *icsk = inet_csk(sk);
1299 	int mss_now;
1300 
1301 	if (icsk->icsk_mtup.search_high > pmtu)
1302 		icsk->icsk_mtup.search_high = pmtu;
1303 
1304 	mss_now = tcp_mtu_to_mss(sk, pmtu);
1305 	mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1306 
1307 	/* And store cached results */
1308 	icsk->icsk_pmtu_cookie = pmtu;
1309 	if (icsk->icsk_mtup.enabled)
1310 		mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1311 	tp->mss_cache = mss_now;
1312 
1313 	return mss_now;
1314 }
1315 EXPORT_SYMBOL(tcp_sync_mss);
1316 
1317 /* Compute the current effective MSS, taking SACKs and IP options,
1318  * and even PMTU discovery events into account.
1319  */
tcp_current_mss(struct sock * sk)1320 unsigned int tcp_current_mss(struct sock *sk)
1321 {
1322 	const struct tcp_sock *tp = tcp_sk(sk);
1323 	const struct dst_entry *dst = __sk_dst_get(sk);
1324 	u32 mss_now;
1325 	unsigned int header_len;
1326 	struct tcp_out_options opts;
1327 	struct tcp_md5sig_key *md5;
1328 
1329 	mss_now = tp->mss_cache;
1330 
1331 	if (dst) {
1332 		u32 mtu = dst_mtu(dst);
1333 		if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1334 			mss_now = tcp_sync_mss(sk, mtu);
1335 	}
1336 
1337 	header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1338 		     sizeof(struct tcphdr);
1339 	/* The mss_cache is sized based on tp->tcp_header_len, which assumes
1340 	 * some common options. If this is an odd packet (because we have SACK
1341 	 * blocks etc) then our calculated header_len will be different, and
1342 	 * we have to adjust mss_now correspondingly */
1343 	if (header_len != tp->tcp_header_len) {
1344 		int delta = (int) header_len - tp->tcp_header_len;
1345 		mss_now -= delta;
1346 	}
1347 
1348 	return mss_now;
1349 }
1350 
1351 /* Congestion window validation. (RFC2861) */
tcp_cwnd_validate(struct sock * sk)1352 static void tcp_cwnd_validate(struct sock *sk)
1353 {
1354 	struct tcp_sock *tp = tcp_sk(sk);
1355 
1356 	if (tp->packets_out >= tp->snd_cwnd) {
1357 		/* Network is feed fully. */
1358 		tp->snd_cwnd_used = 0;
1359 		tp->snd_cwnd_stamp = tcp_time_stamp;
1360 	} else {
1361 		/* Network starves. */
1362 		if (tp->packets_out > tp->snd_cwnd_used)
1363 			tp->snd_cwnd_used = tp->packets_out;
1364 
1365 		if (sysctl_tcp_slow_start_after_idle &&
1366 		    (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1367 			tcp_cwnd_application_limited(sk);
1368 	}
1369 }
1370 
1371 /* Returns the portion of skb which can be sent right away without
1372  * introducing MSS oddities to segment boundaries. In rare cases where
1373  * mss_now != mss_cache, we will request caller to create a small skb
1374  * per input skb which could be mostly avoided here (if desired).
1375  *
1376  * We explicitly want to create a request for splitting write queue tail
1377  * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1378  * thus all the complexity (cwnd_len is always MSS multiple which we
1379  * return whenever allowed by the other factors). Basically we need the
1380  * modulo only when the receiver window alone is the limiting factor or
1381  * when we would be allowed to send the split-due-to-Nagle skb fully.
1382  */
tcp_mss_split_point(const struct sock * sk,const struct sk_buff * skb,unsigned int mss_now,unsigned int max_segs)1383 static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb,
1384 					unsigned int mss_now, unsigned int max_segs)
1385 {
1386 	const struct tcp_sock *tp = tcp_sk(sk);
1387 	u32 needed, window, max_len;
1388 
1389 	window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1390 	max_len = mss_now * max_segs;
1391 
1392 	if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1393 		return max_len;
1394 
1395 	needed = min(skb->len, window);
1396 
1397 	if (max_len <= needed)
1398 		return max_len;
1399 
1400 	return needed - needed % mss_now;
1401 }
1402 
1403 /* Can at least one segment of SKB be sent right now, according to the
1404  * congestion window rules?  If so, return how many segments are allowed.
1405  */
tcp_cwnd_test(const struct tcp_sock * tp,const struct sk_buff * skb)1406 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1407 					 const struct sk_buff *skb)
1408 {
1409 	u32 in_flight, cwnd;
1410 
1411 	/* Don't be strict about the congestion window for the final FIN.  */
1412 	if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1413 	    tcp_skb_pcount(skb) == 1)
1414 		return 1;
1415 
1416 	in_flight = tcp_packets_in_flight(tp);
1417 	cwnd = tp->snd_cwnd;
1418 	if (in_flight < cwnd)
1419 		return (cwnd - in_flight);
1420 
1421 	return 0;
1422 }
1423 
1424 /* Initialize TSO state of a skb.
1425  * This must be invoked the first time we consider transmitting
1426  * SKB onto the wire.
1427  */
tcp_init_tso_segs(const struct sock * sk,struct sk_buff * skb,unsigned int mss_now)1428 static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb,
1429 			     unsigned int mss_now)
1430 {
1431 	int tso_segs = tcp_skb_pcount(skb);
1432 
1433 	if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1434 		tcp_set_skb_tso_segs(sk, skb, mss_now);
1435 		tso_segs = tcp_skb_pcount(skb);
1436 	}
1437 	return tso_segs;
1438 }
1439 
1440 /* Minshall's variant of the Nagle send check. */
tcp_minshall_check(const struct tcp_sock * tp)1441 static inline bool tcp_minshall_check(const struct tcp_sock *tp)
1442 {
1443 	return after(tp->snd_sml, tp->snd_una) &&
1444 		!after(tp->snd_sml, tp->snd_nxt);
1445 }
1446 
1447 /* Return false, if packet can be sent now without violation Nagle's rules:
1448  * 1. It is full sized.
1449  * 2. Or it contains FIN. (already checked by caller)
1450  * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1451  * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1452  *    With Minshall's modification: all sent small packets are ACKed.
1453  */
tcp_nagle_check(const struct tcp_sock * tp,const struct sk_buff * skb,unsigned int mss_now,int nonagle)1454 static inline bool tcp_nagle_check(const struct tcp_sock *tp,
1455 				  const struct sk_buff *skb,
1456 				  unsigned int mss_now, int nonagle)
1457 {
1458 	return skb->len < mss_now &&
1459 		((nonagle & TCP_NAGLE_CORK) ||
1460 		 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1461 }
1462 
1463 /* Return true if the Nagle test allows this packet to be
1464  * sent now.
1465  */
tcp_nagle_test(const struct tcp_sock * tp,const struct sk_buff * skb,unsigned int cur_mss,int nonagle)1466 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1467 				  unsigned int cur_mss, int nonagle)
1468 {
1469 	/* Nagle rule does not apply to frames, which sit in the middle of the
1470 	 * write_queue (they have no chances to get new data).
1471 	 *
1472 	 * This is implemented in the callers, where they modify the 'nonagle'
1473 	 * argument based upon the location of SKB in the send queue.
1474 	 */
1475 	if (nonagle & TCP_NAGLE_PUSH)
1476 		return true;
1477 
1478 	/* Don't use the nagle rule for urgent data (or for the final FIN). */
1479 	if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1480 		return true;
1481 
1482 	if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1483 		return true;
1484 
1485 	return false;
1486 }
1487 
1488 /* Does at least the first segment of SKB fit into the send window? */
tcp_snd_wnd_test(const struct tcp_sock * tp,const struct sk_buff * skb,unsigned int cur_mss)1489 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1490 			     const struct sk_buff *skb,
1491 			     unsigned int cur_mss)
1492 {
1493 	u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1494 
1495 	if (skb->len > cur_mss)
1496 		end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1497 
1498 	return !after(end_seq, tcp_wnd_end(tp));
1499 }
1500 
1501 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1502  * should be put on the wire right now.  If so, it returns the number of
1503  * packets allowed by the congestion window.
1504  */
tcp_snd_test(const struct sock * sk,struct sk_buff * skb,unsigned int cur_mss,int nonagle)1505 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1506 				 unsigned int cur_mss, int nonagle)
1507 {
1508 	const struct tcp_sock *tp = tcp_sk(sk);
1509 	unsigned int cwnd_quota;
1510 
1511 	tcp_init_tso_segs(sk, skb, cur_mss);
1512 
1513 	if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1514 		return 0;
1515 
1516 	cwnd_quota = tcp_cwnd_test(tp, skb);
1517 	if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1518 		cwnd_quota = 0;
1519 
1520 	return cwnd_quota;
1521 }
1522 
1523 /* Test if sending is allowed right now. */
tcp_may_send_now(struct sock * sk)1524 bool tcp_may_send_now(struct sock *sk)
1525 {
1526 	const struct tcp_sock *tp = tcp_sk(sk);
1527 	struct sk_buff *skb = tcp_send_head(sk);
1528 
1529 	return skb &&
1530 		tcp_snd_test(sk, skb, tcp_current_mss(sk),
1531 			     (tcp_skb_is_last(sk, skb) ?
1532 			      tp->nonagle : TCP_NAGLE_PUSH));
1533 }
1534 
1535 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1536  * which is put after SKB on the list.  It is very much like
1537  * tcp_fragment() except that it may make several kinds of assumptions
1538  * in order to speed up the splitting operation.  In particular, we
1539  * know that all the data is in scatter-gather pages, and that the
1540  * packet has never been sent out before (and thus is not cloned).
1541  */
tso_fragment(struct sock * sk,struct sk_buff * skb,unsigned int len,unsigned int mss_now,gfp_t gfp)1542 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1543 			unsigned int mss_now, gfp_t gfp)
1544 {
1545 	struct sk_buff *buff;
1546 	int nlen = skb->len - len;
1547 	u8 flags;
1548 
1549 	/* All of a TSO frame must be composed of paged data.  */
1550 	if (skb->len != skb->data_len)
1551 		return tcp_fragment(sk, skb, len, mss_now);
1552 
1553 	buff = sk_stream_alloc_skb(sk, 0, gfp);
1554 	if (unlikely(buff == NULL))
1555 		return -ENOMEM;
1556 
1557 	sk->sk_wmem_queued += buff->truesize;
1558 	sk_mem_charge(sk, buff->truesize);
1559 	buff->truesize += nlen;
1560 	skb->truesize -= nlen;
1561 
1562 	/* Correct the sequence numbers. */
1563 	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1564 	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1565 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1566 
1567 	/* PSH and FIN should only be set in the second packet. */
1568 	flags = TCP_SKB_CB(skb)->tcp_flags;
1569 	TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1570 	TCP_SKB_CB(buff)->tcp_flags = flags;
1571 
1572 	/* This packet was never sent out yet, so no SACK bits. */
1573 	TCP_SKB_CB(buff)->sacked = 0;
1574 
1575 	buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1576 	skb_split(skb, buff, len);
1577 
1578 	/* Fix up tso_factor for both original and new SKB.  */
1579 	tcp_set_skb_tso_segs(sk, skb, mss_now);
1580 	tcp_set_skb_tso_segs(sk, buff, mss_now);
1581 
1582 	/* Link BUFF into the send queue. */
1583 	skb_header_release(buff);
1584 	tcp_insert_write_queue_after(skb, buff, sk);
1585 
1586 	return 0;
1587 }
1588 
1589 /* Try to defer sending, if possible, in order to minimize the amount
1590  * of TSO splitting we do.  View it as a kind of TSO Nagle test.
1591  *
1592  * This algorithm is from John Heffner.
1593  */
tcp_tso_should_defer(struct sock * sk,struct sk_buff * skb)1594 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1595 {
1596 	struct tcp_sock *tp = tcp_sk(sk);
1597 	const struct inet_connection_sock *icsk = inet_csk(sk);
1598 	u32 send_win, cong_win, limit, in_flight;
1599 	int win_divisor;
1600 
1601 	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1602 		goto send_now;
1603 
1604 	if (icsk->icsk_ca_state != TCP_CA_Open)
1605 		goto send_now;
1606 
1607 	/* Defer for less than two clock ticks. */
1608 	if (tp->tso_deferred &&
1609 	    (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1610 		goto send_now;
1611 
1612 	in_flight = tcp_packets_in_flight(tp);
1613 
1614 	BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1615 
1616 	send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1617 
1618 	/* From in_flight test above, we know that cwnd > in_flight.  */
1619 	cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1620 
1621 	limit = min(send_win, cong_win);
1622 
1623 	/* If a full-sized TSO skb can be sent, do it. */
1624 	if (limit >= min_t(unsigned int, sk->sk_gso_max_size,
1625 			   sk->sk_gso_max_segs * tp->mss_cache))
1626 		goto send_now;
1627 
1628 	/* Middle in queue won't get any more data, full sendable already? */
1629 	if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1630 		goto send_now;
1631 
1632 	win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1633 	if (win_divisor) {
1634 		u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1635 
1636 		/* If at least some fraction of a window is available,
1637 		 * just use it.
1638 		 */
1639 		chunk /= win_divisor;
1640 		if (limit >= chunk)
1641 			goto send_now;
1642 	} else {
1643 		/* Different approach, try not to defer past a single
1644 		 * ACK.  Receiver should ACK every other full sized
1645 		 * frame, so if we have space for more than 3 frames
1646 		 * then send now.
1647 		 */
1648 		if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1649 			goto send_now;
1650 	}
1651 
1652 	/* Ok, it looks like it is advisable to defer.
1653 	 * Do not rearm the timer if already set to not break TCP ACK clocking.
1654 	 */
1655 	if (!tp->tso_deferred)
1656 		tp->tso_deferred = 1 | (jiffies << 1);
1657 
1658 	return true;
1659 
1660 send_now:
1661 	tp->tso_deferred = 0;
1662 	return false;
1663 }
1664 
1665 /* Create a new MTU probe if we are ready.
1666  * MTU probe is regularly attempting to increase the path MTU by
1667  * deliberately sending larger packets.  This discovers routing
1668  * changes resulting in larger path MTUs.
1669  *
1670  * Returns 0 if we should wait to probe (no cwnd available),
1671  *         1 if a probe was sent,
1672  *         -1 otherwise
1673  */
tcp_mtu_probe(struct sock * sk)1674 static int tcp_mtu_probe(struct sock *sk)
1675 {
1676 	struct tcp_sock *tp = tcp_sk(sk);
1677 	struct inet_connection_sock *icsk = inet_csk(sk);
1678 	struct sk_buff *skb, *nskb, *next;
1679 	int len;
1680 	int probe_size;
1681 	int size_needed;
1682 	int copy;
1683 	int mss_now;
1684 
1685 	/* Not currently probing/verifying,
1686 	 * not in recovery,
1687 	 * have enough cwnd, and
1688 	 * not SACKing (the variable headers throw things off) */
1689 	if (!icsk->icsk_mtup.enabled ||
1690 	    icsk->icsk_mtup.probe_size ||
1691 	    inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1692 	    tp->snd_cwnd < 11 ||
1693 	    tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1694 		return -1;
1695 
1696 	/* Very simple search strategy: just double the MSS. */
1697 	mss_now = tcp_current_mss(sk);
1698 	probe_size = 2 * tp->mss_cache;
1699 	size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1700 	if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1701 		/* TODO: set timer for probe_converge_event */
1702 		return -1;
1703 	}
1704 
1705 	/* Have enough data in the send queue to probe? */
1706 	if (tp->write_seq - tp->snd_nxt < size_needed)
1707 		return -1;
1708 
1709 	if (tp->snd_wnd < size_needed)
1710 		return -1;
1711 	if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1712 		return 0;
1713 
1714 	/* Do we need to wait to drain cwnd? With none in flight, don't stall */
1715 	if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1716 		if (!tcp_packets_in_flight(tp))
1717 			return -1;
1718 		else
1719 			return 0;
1720 	}
1721 
1722 	/* We're allowed to probe.  Build it now. */
1723 	if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1724 		return -1;
1725 	sk->sk_wmem_queued += nskb->truesize;
1726 	sk_mem_charge(sk, nskb->truesize);
1727 
1728 	skb = tcp_send_head(sk);
1729 
1730 	TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1731 	TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1732 	TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1733 	TCP_SKB_CB(nskb)->sacked = 0;
1734 	nskb->csum = 0;
1735 	nskb->ip_summed = skb->ip_summed;
1736 
1737 	tcp_insert_write_queue_before(nskb, skb, sk);
1738 
1739 	len = 0;
1740 	tcp_for_write_queue_from_safe(skb, next, sk) {
1741 		copy = min_t(int, skb->len, probe_size - len);
1742 		if (nskb->ip_summed)
1743 			skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1744 		else
1745 			nskb->csum = skb_copy_and_csum_bits(skb, 0,
1746 							    skb_put(nskb, copy),
1747 							    copy, nskb->csum);
1748 
1749 		if (skb->len <= copy) {
1750 			/* We've eaten all the data from this skb.
1751 			 * Throw it away. */
1752 			TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1753 			tcp_unlink_write_queue(skb, sk);
1754 			sk_wmem_free_skb(sk, skb);
1755 		} else {
1756 			TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
1757 						   ~(TCPHDR_FIN|TCPHDR_PSH);
1758 			if (!skb_shinfo(skb)->nr_frags) {
1759 				skb_pull(skb, copy);
1760 				if (skb->ip_summed != CHECKSUM_PARTIAL)
1761 					skb->csum = csum_partial(skb->data,
1762 								 skb->len, 0);
1763 			} else {
1764 				__pskb_trim_head(skb, copy);
1765 				tcp_set_skb_tso_segs(sk, skb, mss_now);
1766 			}
1767 			TCP_SKB_CB(skb)->seq += copy;
1768 		}
1769 
1770 		len += copy;
1771 
1772 		if (len >= probe_size)
1773 			break;
1774 	}
1775 	tcp_init_tso_segs(sk, nskb, nskb->len);
1776 
1777 	/* We're ready to send.  If this fails, the probe will
1778 	 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1779 	TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1780 	if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1781 		/* Decrement cwnd here because we are sending
1782 		 * effectively two packets. */
1783 		tp->snd_cwnd--;
1784 		tcp_event_new_data_sent(sk, nskb);
1785 
1786 		icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1787 		tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1788 		tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1789 
1790 		return 1;
1791 	}
1792 
1793 	return -1;
1794 }
1795 
1796 /* This routine writes packets to the network.  It advances the
1797  * send_head.  This happens as incoming acks open up the remote
1798  * window for us.
1799  *
1800  * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1801  * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1802  * account rare use of URG, this is not a big flaw.
1803  *
1804  * Send at most one packet when push_one > 0. Temporarily ignore
1805  * cwnd limit to force at most one packet out when push_one == 2.
1806 
1807  * Returns true, if no segments are in flight and we have queued segments,
1808  * but cannot send anything now because of SWS or another problem.
1809  */
tcp_write_xmit(struct sock * sk,unsigned int mss_now,int nonagle,int push_one,gfp_t gfp)1810 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1811 			   int push_one, gfp_t gfp)
1812 {
1813 	struct tcp_sock *tp = tcp_sk(sk);
1814 	struct sk_buff *skb;
1815 	unsigned int tso_segs, sent_pkts;
1816 	int cwnd_quota;
1817 	int result;
1818 
1819 	sent_pkts = 0;
1820 
1821 	if (!push_one) {
1822 		/* Do MTU probing. */
1823 		result = tcp_mtu_probe(sk);
1824 		if (!result) {
1825 			return false;
1826 		} else if (result > 0) {
1827 			sent_pkts = 1;
1828 		}
1829 	}
1830 
1831 	while ((skb = tcp_send_head(sk))) {
1832 		unsigned int limit;
1833 
1834 
1835 		tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1836 		BUG_ON(!tso_segs);
1837 
1838 		if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE)
1839 			goto repair; /* Skip network transmission */
1840 
1841 		cwnd_quota = tcp_cwnd_test(tp, skb);
1842 		if (!cwnd_quota) {
1843 			if (push_one == 2)
1844 				/* Force out a loss probe pkt. */
1845 				cwnd_quota = 1;
1846 			else
1847 				break;
1848 		}
1849 
1850 		if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1851 			break;
1852 
1853 		if (tso_segs == 1) {
1854 			if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1855 						     (tcp_skb_is_last(sk, skb) ?
1856 						      nonagle : TCP_NAGLE_PUSH))))
1857 				break;
1858 		} else {
1859 			if (!push_one && tcp_tso_should_defer(sk, skb))
1860 				break;
1861 		}
1862 
1863 		/* TSQ : sk_wmem_alloc accounts skb truesize,
1864 		 * including skb overhead. But thats OK.
1865 		 */
1866 		if (atomic_read(&sk->sk_wmem_alloc) >= sysctl_tcp_limit_output_bytes) {
1867 			set_bit(TSQ_THROTTLED, &tp->tsq_flags);
1868 			break;
1869 		}
1870 		limit = mss_now;
1871 		if (tso_segs > 1 && !tcp_urg_mode(tp))
1872 			limit = tcp_mss_split_point(sk, skb, mss_now,
1873 						    min_t(unsigned int,
1874 							  cwnd_quota,
1875 							  sk->sk_gso_max_segs));
1876 
1877 		if (skb->len > limit &&
1878 		    unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
1879 			break;
1880 
1881 		TCP_SKB_CB(skb)->when = tcp_time_stamp;
1882 
1883 		if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
1884 			break;
1885 
1886 repair:
1887 		/* Advance the send_head.  This one is sent out.
1888 		 * This call will increment packets_out.
1889 		 */
1890 		tcp_event_new_data_sent(sk, skb);
1891 
1892 		tcp_minshall_update(tp, mss_now, skb);
1893 		sent_pkts += tcp_skb_pcount(skb);
1894 
1895 		if (push_one)
1896 			break;
1897 	}
1898 
1899 	if (likely(sent_pkts)) {
1900 		if (tcp_in_cwnd_reduction(sk))
1901 			tp->prr_out += sent_pkts;
1902 
1903 		/* Send one loss probe per tail loss episode. */
1904 		if (push_one != 2)
1905 			tcp_schedule_loss_probe(sk);
1906 		tcp_cwnd_validate(sk);
1907 		return false;
1908 	}
1909 	return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk));
1910 }
1911 
tcp_schedule_loss_probe(struct sock * sk)1912 bool tcp_schedule_loss_probe(struct sock *sk)
1913 {
1914 	struct inet_connection_sock *icsk = inet_csk(sk);
1915 	struct tcp_sock *tp = tcp_sk(sk);
1916 	u32 timeout, tlp_time_stamp, rto_time_stamp;
1917 	u32 rtt = tp->srtt >> 3;
1918 
1919 	if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
1920 		return false;
1921 	/* No consecutive loss probes. */
1922 	if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
1923 		tcp_rearm_rto(sk);
1924 		return false;
1925 	}
1926 	/* Don't do any loss probe on a Fast Open connection before 3WHS
1927 	 * finishes.
1928 	 */
1929 	if (sk->sk_state == TCP_SYN_RECV)
1930 		return false;
1931 
1932 	/* TLP is only scheduled when next timer event is RTO. */
1933 	if (icsk->icsk_pending != ICSK_TIME_RETRANS)
1934 		return false;
1935 
1936 	/* Schedule a loss probe in 2*RTT for SACK capable connections
1937 	 * in Open state, that are either limited by cwnd or application.
1938 	 */
1939 	if (sysctl_tcp_early_retrans < 3 || !rtt || !tp->packets_out ||
1940 	    !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
1941 		return false;
1942 
1943 	if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
1944 	     tcp_send_head(sk))
1945 		return false;
1946 
1947 	/* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
1948 	 * for delayed ack when there's one outstanding packet.
1949 	 */
1950 	timeout = rtt << 1;
1951 	if (tp->packets_out == 1)
1952 		timeout = max_t(u32, timeout,
1953 				(rtt + (rtt >> 1) + TCP_DELACK_MAX));
1954 	timeout = max_t(u32, timeout, msecs_to_jiffies(10));
1955 
1956 	/* If RTO is shorter, just schedule TLP in its place. */
1957 	tlp_time_stamp = tcp_time_stamp + timeout;
1958 	rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
1959 	if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
1960 		s32 delta = rto_time_stamp - tcp_time_stamp;
1961 		if (delta > 0)
1962 			timeout = delta;
1963 	}
1964 
1965 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
1966 				  TCP_RTO_MAX);
1967 	return true;
1968 }
1969 
1970 /* When probe timeout (PTO) fires, send a new segment if one exists, else
1971  * retransmit the last segment.
1972  */
tcp_send_loss_probe(struct sock * sk)1973 void tcp_send_loss_probe(struct sock *sk)
1974 {
1975 	struct tcp_sock *tp = tcp_sk(sk);
1976 	struct sk_buff *skb;
1977 	int pcount;
1978 	int mss = tcp_current_mss(sk);
1979 	int err = -1;
1980 
1981 	if (tcp_send_head(sk) != NULL) {
1982 		err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
1983 		goto rearm_timer;
1984 	}
1985 
1986 	/* At most one outstanding TLP retransmission. */
1987 	if (tp->tlp_high_seq)
1988 		goto rearm_timer;
1989 
1990 	/* Retransmit last segment. */
1991 	skb = tcp_write_queue_tail(sk);
1992 	if (WARN_ON(!skb))
1993 		goto rearm_timer;
1994 
1995 	pcount = tcp_skb_pcount(skb);
1996 	if (WARN_ON(!pcount))
1997 		goto rearm_timer;
1998 
1999 	if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2000 		if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss)))
2001 			goto rearm_timer;
2002 		skb = tcp_write_queue_tail(sk);
2003 	}
2004 
2005 	if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2006 		goto rearm_timer;
2007 
2008 	/* Probe with zero data doesn't trigger fast recovery. */
2009 	if (skb->len > 0)
2010 		err = __tcp_retransmit_skb(sk, skb);
2011 
2012 	/* Record snd_nxt for loss detection. */
2013 	if (likely(!err))
2014 		tp->tlp_high_seq = tp->snd_nxt;
2015 
2016 rearm_timer:
2017 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2018 				  inet_csk(sk)->icsk_rto,
2019 				  TCP_RTO_MAX);
2020 
2021 	if (likely(!err))
2022 		NET_INC_STATS_BH(sock_net(sk),
2023 				 LINUX_MIB_TCPLOSSPROBES);
2024 	return;
2025 }
2026 
2027 /* Push out any pending frames which were held back due to
2028  * TCP_CORK or attempt at coalescing tiny packets.
2029  * The socket must be locked by the caller.
2030  */
__tcp_push_pending_frames(struct sock * sk,unsigned int cur_mss,int nonagle)2031 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2032 			       int nonagle)
2033 {
2034 	/* If we are closed, the bytes will have to remain here.
2035 	 * In time closedown will finish, we empty the write queue and
2036 	 * all will be happy.
2037 	 */
2038 	if (unlikely(sk->sk_state == TCP_CLOSE))
2039 		return;
2040 
2041 	if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2042 			   sk_gfp_atomic(sk, GFP_ATOMIC)))
2043 		tcp_check_probe_timer(sk);
2044 }
2045 
2046 /* Send _single_ skb sitting at the send head. This function requires
2047  * true push pending frames to setup probe timer etc.
2048  */
tcp_push_one(struct sock * sk,unsigned int mss_now)2049 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2050 {
2051 	struct sk_buff *skb = tcp_send_head(sk);
2052 
2053 	BUG_ON(!skb || skb->len < mss_now);
2054 
2055 	tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2056 }
2057 
2058 /* This function returns the amount that we can raise the
2059  * usable window based on the following constraints
2060  *
2061  * 1. The window can never be shrunk once it is offered (RFC 793)
2062  * 2. We limit memory per socket
2063  *
2064  * RFC 1122:
2065  * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2066  *  RECV.NEXT + RCV.WIN fixed until:
2067  *  RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2068  *
2069  * i.e. don't raise the right edge of the window until you can raise
2070  * it at least MSS bytes.
2071  *
2072  * Unfortunately, the recommended algorithm breaks header prediction,
2073  * since header prediction assumes th->window stays fixed.
2074  *
2075  * Strictly speaking, keeping th->window fixed violates the receiver
2076  * side SWS prevention criteria. The problem is that under this rule
2077  * a stream of single byte packets will cause the right side of the
2078  * window to always advance by a single byte.
2079  *
2080  * Of course, if the sender implements sender side SWS prevention
2081  * then this will not be a problem.
2082  *
2083  * BSD seems to make the following compromise:
2084  *
2085  *	If the free space is less than the 1/4 of the maximum
2086  *	space available and the free space is less than 1/2 mss,
2087  *	then set the window to 0.
2088  *	[ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2089  *	Otherwise, just prevent the window from shrinking
2090  *	and from being larger than the largest representable value.
2091  *
2092  * This prevents incremental opening of the window in the regime
2093  * where TCP is limited by the speed of the reader side taking
2094  * data out of the TCP receive queue. It does nothing about
2095  * those cases where the window is constrained on the sender side
2096  * because the pipeline is full.
2097  *
2098  * BSD also seems to "accidentally" limit itself to windows that are a
2099  * multiple of MSS, at least until the free space gets quite small.
2100  * This would appear to be a side effect of the mbuf implementation.
2101  * Combining these two algorithms results in the observed behavior
2102  * of having a fixed window size at almost all times.
2103  *
2104  * Below we obtain similar behavior by forcing the offered window to
2105  * a multiple of the mss when it is feasible to do so.
2106  *
2107  * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2108  * Regular options like TIMESTAMP are taken into account.
2109  */
__tcp_select_window(struct sock * sk)2110 u32 __tcp_select_window(struct sock *sk)
2111 {
2112 	struct inet_connection_sock *icsk = inet_csk(sk);
2113 	struct tcp_sock *tp = tcp_sk(sk);
2114 	/* MSS for the peer's data.  Previous versions used mss_clamp
2115 	 * here.  I don't know if the value based on our guesses
2116 	 * of peer's MSS is better for the performance.  It's more correct
2117 	 * but may be worse for the performance because of rcv_mss
2118 	 * fluctuations.  --SAW  1998/11/1
2119 	 */
2120 	int mss = icsk->icsk_ack.rcv_mss;
2121 	int free_space = tcp_space(sk);
2122 	int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
2123 	int window;
2124 
2125 	if (mss > full_space)
2126 		mss = full_space;
2127 
2128 	if (free_space < (full_space >> 1)) {
2129 		icsk->icsk_ack.quick = 0;
2130 
2131 		if (sk_under_memory_pressure(sk))
2132 			tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2133 					       4U * tp->advmss);
2134 
2135 		if (free_space < mss)
2136 			return 0;
2137 	}
2138 
2139 	if (free_space > tp->rcv_ssthresh)
2140 		free_space = tp->rcv_ssthresh;
2141 
2142 	/* Don't do rounding if we are using window scaling, since the
2143 	 * scaled window will not line up with the MSS boundary anyway.
2144 	 */
2145 	window = tp->rcv_wnd;
2146 	if (tp->rx_opt.rcv_wscale) {
2147 		window = free_space;
2148 
2149 		/* Advertise enough space so that it won't get scaled away.
2150 		 * Import case: prevent zero window announcement if
2151 		 * 1<<rcv_wscale > mss.
2152 		 */
2153 		if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2154 			window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2155 				  << tp->rx_opt.rcv_wscale);
2156 	} else {
2157 		/* Get the largest window that is a nice multiple of mss.
2158 		 * Window clamp already applied above.
2159 		 * If our current window offering is within 1 mss of the
2160 		 * free space we just keep it. This prevents the divide
2161 		 * and multiply from happening most of the time.
2162 		 * We also don't do any window rounding when the free space
2163 		 * is too small.
2164 		 */
2165 		if (window <= free_space - mss || window > free_space)
2166 			window = (free_space / mss) * mss;
2167 		else if (mss == full_space &&
2168 			 free_space > window + (full_space >> 1))
2169 			window = free_space;
2170 	}
2171 
2172 	return window;
2173 }
2174 
2175 /* Collapses two adjacent SKB's during retransmission. */
tcp_collapse_retrans(struct sock * sk,struct sk_buff * skb)2176 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2177 {
2178 	struct tcp_sock *tp = tcp_sk(sk);
2179 	struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2180 	int skb_size, next_skb_size;
2181 
2182 	skb_size = skb->len;
2183 	next_skb_size = next_skb->len;
2184 
2185 	BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2186 
2187 	tcp_highest_sack_combine(sk, next_skb, skb);
2188 
2189 	tcp_unlink_write_queue(next_skb, sk);
2190 
2191 	skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2192 				  next_skb_size);
2193 
2194 	if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2195 		skb->ip_summed = CHECKSUM_PARTIAL;
2196 
2197 	if (skb->ip_summed != CHECKSUM_PARTIAL)
2198 		skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2199 
2200 	/* Update sequence range on original skb. */
2201 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2202 
2203 	/* Merge over control information. This moves PSH/FIN etc. over */
2204 	TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2205 
2206 	/* All done, get rid of second SKB and account for it so
2207 	 * packet counting does not break.
2208 	 */
2209 	TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2210 
2211 	/* changed transmit queue under us so clear hints */
2212 	tcp_clear_retrans_hints_partial(tp);
2213 	if (next_skb == tp->retransmit_skb_hint)
2214 		tp->retransmit_skb_hint = skb;
2215 
2216 	tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2217 
2218 	sk_wmem_free_skb(sk, next_skb);
2219 }
2220 
2221 /* Check if coalescing SKBs is legal. */
tcp_can_collapse(const struct sock * sk,const struct sk_buff * skb)2222 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2223 {
2224 	if (tcp_skb_pcount(skb) > 1)
2225 		return false;
2226 	/* TODO: SACK collapsing could be used to remove this condition */
2227 	if (skb_shinfo(skb)->nr_frags != 0)
2228 		return false;
2229 	if (skb_cloned(skb))
2230 		return false;
2231 	if (skb == tcp_send_head(sk))
2232 		return false;
2233 	/* Some heurestics for collapsing over SACK'd could be invented */
2234 	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2235 		return false;
2236 
2237 	return true;
2238 }
2239 
2240 /* Collapse packets in the retransmit queue to make to create
2241  * less packets on the wire. This is only done on retransmission.
2242  */
tcp_retrans_try_collapse(struct sock * sk,struct sk_buff * to,int space)2243 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2244 				     int space)
2245 {
2246 	struct tcp_sock *tp = tcp_sk(sk);
2247 	struct sk_buff *skb = to, *tmp;
2248 	bool first = true;
2249 
2250 	if (!sysctl_tcp_retrans_collapse)
2251 		return;
2252 	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2253 		return;
2254 
2255 	tcp_for_write_queue_from_safe(skb, tmp, sk) {
2256 		if (!tcp_can_collapse(sk, skb))
2257 			break;
2258 
2259 		space -= skb->len;
2260 
2261 		if (first) {
2262 			first = false;
2263 			continue;
2264 		}
2265 
2266 		if (space < 0)
2267 			break;
2268 		/* Punt if not enough space exists in the first SKB for
2269 		 * the data in the second
2270 		 */
2271 		if (skb->len > skb_availroom(to))
2272 			break;
2273 
2274 		if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2275 			break;
2276 
2277 		tcp_collapse_retrans(sk, to);
2278 	}
2279 }
2280 
2281 /* This retransmits one SKB.  Policy decisions and retransmit queue
2282  * state updates are done by the caller.  Returns non-zero if an
2283  * error occurred which prevented the send.
2284  */
__tcp_retransmit_skb(struct sock * sk,struct sk_buff * skb)2285 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2286 {
2287 	struct tcp_sock *tp = tcp_sk(sk);
2288 	struct inet_connection_sock *icsk = inet_csk(sk);
2289 	unsigned int cur_mss;
2290 
2291 	/* Inconslusive MTU probe */
2292 	if (icsk->icsk_mtup.probe_size) {
2293 		icsk->icsk_mtup.probe_size = 0;
2294 	}
2295 
2296 	/* Do not sent more than we queued. 1/4 is reserved for possible
2297 	 * copying overhead: fragmentation, tunneling, mangling etc.
2298 	 */
2299 	if (atomic_read(&sk->sk_wmem_alloc) >
2300 	    min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2301 		return -EAGAIN;
2302 
2303 	if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2304 		if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2305 			BUG();
2306 		if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2307 			return -ENOMEM;
2308 	}
2309 
2310 	if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2311 		return -EHOSTUNREACH; /* Routing failure or similar. */
2312 
2313 	cur_mss = tcp_current_mss(sk);
2314 
2315 	/* If receiver has shrunk his window, and skb is out of
2316 	 * new window, do not retransmit it. The exception is the
2317 	 * case, when window is shrunk to zero. In this case
2318 	 * our retransmit serves as a zero window probe.
2319 	 */
2320 	if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2321 	    TCP_SKB_CB(skb)->seq != tp->snd_una)
2322 		return -EAGAIN;
2323 
2324 	if (skb->len > cur_mss) {
2325 		if (tcp_fragment(sk, skb, cur_mss, cur_mss))
2326 			return -ENOMEM; /* We'll try again later. */
2327 	} else {
2328 		int oldpcount = tcp_skb_pcount(skb);
2329 
2330 		if (unlikely(oldpcount > 1)) {
2331 			tcp_init_tso_segs(sk, skb, cur_mss);
2332 			tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2333 		}
2334 	}
2335 
2336 	tcp_retrans_try_collapse(sk, skb, cur_mss);
2337 
2338 	/* Some Solaris stacks overoptimize and ignore the FIN on a
2339 	 * retransmit when old data is attached.  So strip it off
2340 	 * since it is cheap to do so and saves bytes on the network.
2341 	 */
2342 	if (skb->len > 0 &&
2343 	    (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
2344 	    tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
2345 		if (!pskb_trim(skb, 0)) {
2346 			/* Reuse, even though it does some unnecessary work */
2347 			tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
2348 					     TCP_SKB_CB(skb)->tcp_flags);
2349 			skb->ip_summed = CHECKSUM_NONE;
2350 		}
2351 	}
2352 
2353 	/* Make a copy, if the first transmission SKB clone we made
2354 	 * is still in somebody's hands, else make a clone.
2355 	 */
2356 	TCP_SKB_CB(skb)->when = tcp_time_stamp;
2357 
2358 	/* make sure skb->data is aligned on arches that require it
2359 	 * and check if ack-trimming & collapsing extended the headroom
2360 	 * beyond what csum_start can cover.
2361 	 */
2362 	if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2363 		     skb_headroom(skb) >= 0xFFFF)) {
2364 		struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
2365 						   GFP_ATOMIC);
2366 		return nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2367 			      -ENOBUFS;
2368 	} else {
2369 		return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2370 	}
2371 }
2372 
tcp_retransmit_skb(struct sock * sk,struct sk_buff * skb)2373 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2374 {
2375 	struct tcp_sock *tp = tcp_sk(sk);
2376 	int err = __tcp_retransmit_skb(sk, skb);
2377 
2378 	if (err == 0) {
2379 		/* Update global TCP statistics. */
2380 		TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2381 
2382 		tp->total_retrans++;
2383 
2384 #if FASTRETRANS_DEBUG > 0
2385 		if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2386 			net_dbg_ratelimited("retrans_out leaked\n");
2387 		}
2388 #endif
2389 		if (!tp->retrans_out)
2390 			tp->lost_retrans_low = tp->snd_nxt;
2391 		TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2392 		tp->retrans_out += tcp_skb_pcount(skb);
2393 
2394 		/* Save stamp of the first retransmit. */
2395 		if (!tp->retrans_stamp)
2396 			tp->retrans_stamp = TCP_SKB_CB(skb)->when;
2397 
2398 		tp->undo_retrans += tcp_skb_pcount(skb);
2399 
2400 		/* snd_nxt is stored to detect loss of retransmitted segment,
2401 		 * see tcp_input.c tcp_sacktag_write_queue().
2402 		 */
2403 		TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2404 	}
2405 	return err;
2406 }
2407 
2408 /* Check if we forward retransmits are possible in the current
2409  * window/congestion state.
2410  */
tcp_can_forward_retransmit(struct sock * sk)2411 static bool tcp_can_forward_retransmit(struct sock *sk)
2412 {
2413 	const struct inet_connection_sock *icsk = inet_csk(sk);
2414 	const struct tcp_sock *tp = tcp_sk(sk);
2415 
2416 	/* Forward retransmissions are possible only during Recovery. */
2417 	if (icsk->icsk_ca_state != TCP_CA_Recovery)
2418 		return false;
2419 
2420 	/* No forward retransmissions in Reno are possible. */
2421 	if (tcp_is_reno(tp))
2422 		return false;
2423 
2424 	/* Yeah, we have to make difficult choice between forward transmission
2425 	 * and retransmission... Both ways have their merits...
2426 	 *
2427 	 * For now we do not retransmit anything, while we have some new
2428 	 * segments to send. In the other cases, follow rule 3 for
2429 	 * NextSeg() specified in RFC3517.
2430 	 */
2431 
2432 	if (tcp_may_send_now(sk))
2433 		return false;
2434 
2435 	return true;
2436 }
2437 
2438 /* This gets called after a retransmit timeout, and the initially
2439  * retransmitted data is acknowledged.  It tries to continue
2440  * resending the rest of the retransmit queue, until either
2441  * we've sent it all or the congestion window limit is reached.
2442  * If doing SACK, the first ACK which comes back for a timeout
2443  * based retransmit packet might feed us FACK information again.
2444  * If so, we use it to avoid unnecessarily retransmissions.
2445  */
tcp_xmit_retransmit_queue(struct sock * sk)2446 void tcp_xmit_retransmit_queue(struct sock *sk)
2447 {
2448 	const struct inet_connection_sock *icsk = inet_csk(sk);
2449 	struct tcp_sock *tp = tcp_sk(sk);
2450 	struct sk_buff *skb;
2451 	struct sk_buff *hole = NULL;
2452 	u32 last_lost;
2453 	int mib_idx;
2454 	int fwd_rexmitting = 0;
2455 
2456 	if (!tp->packets_out)
2457 		return;
2458 
2459 	if (!tp->lost_out)
2460 		tp->retransmit_high = tp->snd_una;
2461 
2462 	if (tp->retransmit_skb_hint) {
2463 		skb = tp->retransmit_skb_hint;
2464 		last_lost = TCP_SKB_CB(skb)->end_seq;
2465 		if (after(last_lost, tp->retransmit_high))
2466 			last_lost = tp->retransmit_high;
2467 	} else {
2468 		skb = tcp_write_queue_head(sk);
2469 		last_lost = tp->snd_una;
2470 	}
2471 
2472 	tcp_for_write_queue_from(skb, sk) {
2473 		__u8 sacked = TCP_SKB_CB(skb)->sacked;
2474 
2475 		if (skb == tcp_send_head(sk))
2476 			break;
2477 		/* we could do better than to assign each time */
2478 		if (hole == NULL)
2479 			tp->retransmit_skb_hint = skb;
2480 
2481 		/* Assume this retransmit will generate
2482 		 * only one packet for congestion window
2483 		 * calculation purposes.  This works because
2484 		 * tcp_retransmit_skb() will chop up the
2485 		 * packet to be MSS sized and all the
2486 		 * packet counting works out.
2487 		 */
2488 		if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2489 			return;
2490 
2491 		if (fwd_rexmitting) {
2492 begin_fwd:
2493 			if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2494 				break;
2495 			mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2496 
2497 		} else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2498 			tp->retransmit_high = last_lost;
2499 			if (!tcp_can_forward_retransmit(sk))
2500 				break;
2501 			/* Backtrack if necessary to non-L'ed skb */
2502 			if (hole != NULL) {
2503 				skb = hole;
2504 				hole = NULL;
2505 			}
2506 			fwd_rexmitting = 1;
2507 			goto begin_fwd;
2508 
2509 		} else if (!(sacked & TCPCB_LOST)) {
2510 			if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2511 				hole = skb;
2512 			continue;
2513 
2514 		} else {
2515 			last_lost = TCP_SKB_CB(skb)->end_seq;
2516 			if (icsk->icsk_ca_state != TCP_CA_Loss)
2517 				mib_idx = LINUX_MIB_TCPFASTRETRANS;
2518 			else
2519 				mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2520 		}
2521 
2522 		if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2523 			continue;
2524 
2525 		if (tcp_retransmit_skb(sk, skb)) {
2526 			NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2527 			return;
2528 		}
2529 		NET_INC_STATS_BH(sock_net(sk), mib_idx);
2530 
2531 		if (tcp_in_cwnd_reduction(sk))
2532 			tp->prr_out += tcp_skb_pcount(skb);
2533 
2534 		if (skb == tcp_write_queue_head(sk))
2535 			inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2536 						  inet_csk(sk)->icsk_rto,
2537 						  TCP_RTO_MAX);
2538 	}
2539 }
2540 
2541 /* Send a fin.  The caller locks the socket for us.  This cannot be
2542  * allowed to fail queueing a FIN frame under any circumstances.
2543  */
tcp_send_fin(struct sock * sk)2544 void tcp_send_fin(struct sock *sk)
2545 {
2546 	struct tcp_sock *tp = tcp_sk(sk);
2547 	struct sk_buff *skb = tcp_write_queue_tail(sk);
2548 	int mss_now;
2549 
2550 	/* Optimization, tack on the FIN if we have a queue of
2551 	 * unsent frames.  But be careful about outgoing SACKS
2552 	 * and IP options.
2553 	 */
2554 	mss_now = tcp_current_mss(sk);
2555 
2556 	if (tcp_send_head(sk) != NULL) {
2557 		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN;
2558 		TCP_SKB_CB(skb)->end_seq++;
2559 		tp->write_seq++;
2560 	} else {
2561 		/* Socket is locked, keep trying until memory is available. */
2562 		for (;;) {
2563 			skb = alloc_skb_fclone(MAX_TCP_HEADER,
2564 					       sk->sk_allocation);
2565 			if (skb)
2566 				break;
2567 			yield();
2568 		}
2569 
2570 		/* Reserve space for headers and prepare control bits. */
2571 		skb_reserve(skb, MAX_TCP_HEADER);
2572 		/* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2573 		tcp_init_nondata_skb(skb, tp->write_seq,
2574 				     TCPHDR_ACK | TCPHDR_FIN);
2575 		tcp_queue_skb(sk, skb);
2576 	}
2577 	__tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2578 }
2579 
2580 /* We get here when a process closes a file descriptor (either due to
2581  * an explicit close() or as a byproduct of exit()'ing) and there
2582  * was unread data in the receive queue.  This behavior is recommended
2583  * by RFC 2525, section 2.17.  -DaveM
2584  */
tcp_send_active_reset(struct sock * sk,gfp_t priority)2585 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2586 {
2587 	struct sk_buff *skb;
2588 
2589 	/* NOTE: No TCP options attached and we never retransmit this. */
2590 	skb = alloc_skb(MAX_TCP_HEADER, priority);
2591 	if (!skb) {
2592 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2593 		return;
2594 	}
2595 
2596 	/* Reserve space for headers and prepare control bits. */
2597 	skb_reserve(skb, MAX_TCP_HEADER);
2598 	tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2599 			     TCPHDR_ACK | TCPHDR_RST);
2600 	/* Send it off. */
2601 	TCP_SKB_CB(skb)->when = tcp_time_stamp;
2602 	if (tcp_transmit_skb(sk, skb, 0, priority))
2603 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2604 
2605 	TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2606 }
2607 
2608 /* Send a crossed SYN-ACK during socket establishment.
2609  * WARNING: This routine must only be called when we have already sent
2610  * a SYN packet that crossed the incoming SYN that caused this routine
2611  * to get called. If this assumption fails then the initial rcv_wnd
2612  * and rcv_wscale values will not be correct.
2613  */
tcp_send_synack(struct sock * sk)2614 int tcp_send_synack(struct sock *sk)
2615 {
2616 	struct sk_buff *skb;
2617 
2618 	skb = tcp_write_queue_head(sk);
2619 	if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2620 		pr_debug("%s: wrong queue state\n", __func__);
2621 		return -EFAULT;
2622 	}
2623 	if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2624 		if (skb_cloned(skb)) {
2625 			struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2626 			if (nskb == NULL)
2627 				return -ENOMEM;
2628 			tcp_unlink_write_queue(skb, sk);
2629 			skb_header_release(nskb);
2630 			__tcp_add_write_queue_head(sk, nskb);
2631 			sk_wmem_free_skb(sk, skb);
2632 			sk->sk_wmem_queued += nskb->truesize;
2633 			sk_mem_charge(sk, nskb->truesize);
2634 			skb = nskb;
2635 		}
2636 
2637 		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2638 		TCP_ECN_send_synack(tcp_sk(sk), skb);
2639 	}
2640 	TCP_SKB_CB(skb)->when = tcp_time_stamp;
2641 	return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2642 }
2643 
2644 /**
2645  * tcp_make_synack - Prepare a SYN-ACK.
2646  * sk: listener socket
2647  * dst: dst entry attached to the SYNACK
2648  * req: request_sock pointer
2649  *
2650  * Allocate one skb and build a SYNACK packet.
2651  * @dst is consumed : Caller should not use it again.
2652  */
tcp_make_synack(struct sock * sk,struct dst_entry * dst,struct request_sock * req,struct tcp_fastopen_cookie * foc)2653 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2654 				struct request_sock *req,
2655 				struct tcp_fastopen_cookie *foc)
2656 {
2657 	struct tcp_out_options opts;
2658 	struct inet_request_sock *ireq = inet_rsk(req);
2659 	struct tcp_sock *tp = tcp_sk(sk);
2660 	struct tcphdr *th;
2661 	struct sk_buff *skb;
2662 	struct tcp_md5sig_key *md5;
2663 	int tcp_header_size;
2664 	int mss;
2665 
2666 	skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
2667 	if (unlikely(!skb)) {
2668 		dst_release(dst);
2669 		return NULL;
2670 	}
2671 	/* Reserve space for headers. */
2672 	skb_reserve(skb, MAX_TCP_HEADER);
2673 
2674 	skb_dst_set(skb, dst);
2675 	security_skb_owned_by(skb, sk);
2676 
2677 	mss = dst_metric_advmss(dst);
2678 	if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2679 		mss = tp->rx_opt.user_mss;
2680 
2681 	if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2682 		__u8 rcv_wscale;
2683 		/* Set this up on the first call only */
2684 		req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2685 
2686 		/* limit the window selection if the user enforce a smaller rx buffer */
2687 		if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2688 		    (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0))
2689 			req->window_clamp = tcp_full_space(sk);
2690 
2691 		/* tcp_full_space because it is guaranteed to be the first packet */
2692 		tcp_select_initial_window(tcp_full_space(sk),
2693 			mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2694 			&req->rcv_wnd,
2695 			&req->window_clamp,
2696 			ireq->wscale_ok,
2697 			&rcv_wscale,
2698 			dst_metric(dst, RTAX_INITRWND));
2699 		ireq->rcv_wscale = rcv_wscale;
2700 	}
2701 
2702 	memset(&opts, 0, sizeof(opts));
2703 #ifdef CONFIG_SYN_COOKIES
2704 	if (unlikely(req->cookie_ts))
2705 		TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2706 	else
2707 #endif
2708 	TCP_SKB_CB(skb)->when = tcp_time_stamp;
2709 	tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, &md5,
2710 					     foc) + sizeof(*th);
2711 
2712 	skb_push(skb, tcp_header_size);
2713 	skb_reset_transport_header(skb);
2714 
2715 	th = tcp_hdr(skb);
2716 	memset(th, 0, sizeof(struct tcphdr));
2717 	th->syn = 1;
2718 	th->ack = 1;
2719 	TCP_ECN_make_synack(req, th);
2720 	th->source = ireq->loc_port;
2721 	th->dest = ireq->rmt_port;
2722 	/* Setting of flags are superfluous here for callers (and ECE is
2723 	 * not even correctly set)
2724 	 */
2725 	tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2726 			     TCPHDR_SYN | TCPHDR_ACK);
2727 
2728 	th->seq = htonl(TCP_SKB_CB(skb)->seq);
2729 	/* XXX data is queued and acked as is. No buffer/window check */
2730 	th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
2731 
2732 	/* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2733 	th->window = htons(min(req->rcv_wnd, 65535U));
2734 	tcp_options_write((__be32 *)(th + 1), tp, &opts);
2735 	th->doff = (tcp_header_size >> 2);
2736 	TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb));
2737 
2738 #ifdef CONFIG_TCP_MD5SIG
2739 	/* Okay, we have all we need - do the md5 hash if needed */
2740 	if (md5) {
2741 		tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
2742 					       md5, NULL, req, skb);
2743 	}
2744 #endif
2745 
2746 	return skb;
2747 }
2748 EXPORT_SYMBOL(tcp_make_synack);
2749 
2750 /* Do all connect socket setups that can be done AF independent. */
tcp_connect_init(struct sock * sk)2751 void tcp_connect_init(struct sock *sk)
2752 {
2753 	const struct dst_entry *dst = __sk_dst_get(sk);
2754 	struct tcp_sock *tp = tcp_sk(sk);
2755 	__u8 rcv_wscale;
2756 
2757 	/* We'll fix this up when we get a response from the other end.
2758 	 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2759 	 */
2760 	tp->tcp_header_len = sizeof(struct tcphdr) +
2761 		(sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2762 
2763 #ifdef CONFIG_TCP_MD5SIG
2764 	if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2765 		tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2766 #endif
2767 
2768 	/* If user gave his TCP_MAXSEG, record it to clamp */
2769 	if (tp->rx_opt.user_mss)
2770 		tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2771 	tp->max_window = 0;
2772 	tcp_mtup_init(sk);
2773 	tcp_sync_mss(sk, dst_mtu(dst));
2774 
2775 	if (!tp->window_clamp)
2776 		tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2777 	tp->advmss = dst_metric_advmss(dst);
2778 	if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2779 		tp->advmss = tp->rx_opt.user_mss;
2780 
2781 	tcp_initialize_rcv_mss(sk);
2782 
2783 	/* limit the window selection if the user enforce a smaller rx buffer */
2784 	if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2785 	    (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
2786 		tp->window_clamp = tcp_full_space(sk);
2787 
2788 	tcp_select_initial_window(tcp_full_space(sk),
2789 				  tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2790 				  &tp->rcv_wnd,
2791 				  &tp->window_clamp,
2792 				  sysctl_tcp_window_scaling,
2793 				  &rcv_wscale,
2794 				  dst_metric(dst, RTAX_INITRWND));
2795 
2796 	tp->rx_opt.rcv_wscale = rcv_wscale;
2797 	tp->rcv_ssthresh = tp->rcv_wnd;
2798 
2799 	sk->sk_err = 0;
2800 	sock_reset_flag(sk, SOCK_DONE);
2801 	tp->snd_wnd = 0;
2802 	tcp_init_wl(tp, 0);
2803 	tp->snd_una = tp->write_seq;
2804 	tp->snd_sml = tp->write_seq;
2805 	tp->snd_up = tp->write_seq;
2806 	tp->snd_nxt = tp->write_seq;
2807 
2808 	if (likely(!tp->repair))
2809 		tp->rcv_nxt = 0;
2810 	tp->rcv_wup = tp->rcv_nxt;
2811 	tp->copied_seq = tp->rcv_nxt;
2812 
2813 	inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2814 	inet_csk(sk)->icsk_retransmits = 0;
2815 	tcp_clear_retrans(tp);
2816 }
2817 
tcp_connect_queue_skb(struct sock * sk,struct sk_buff * skb)2818 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
2819 {
2820 	struct tcp_sock *tp = tcp_sk(sk);
2821 	struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
2822 
2823 	tcb->end_seq += skb->len;
2824 	skb_header_release(skb);
2825 	__tcp_add_write_queue_tail(sk, skb);
2826 	sk->sk_wmem_queued += skb->truesize;
2827 	sk_mem_charge(sk, skb->truesize);
2828 	tp->write_seq = tcb->end_seq;
2829 	tp->packets_out += tcp_skb_pcount(skb);
2830 }
2831 
2832 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
2833  * queue a data-only packet after the regular SYN, such that regular SYNs
2834  * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
2835  * only the SYN sequence, the data are retransmitted in the first ACK.
2836  * If cookie is not cached or other error occurs, falls back to send a
2837  * regular SYN with Fast Open cookie request option.
2838  */
tcp_send_syn_data(struct sock * sk,struct sk_buff * syn)2839 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
2840 {
2841 	struct tcp_sock *tp = tcp_sk(sk);
2842 	struct tcp_fastopen_request *fo = tp->fastopen_req;
2843 	int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen;
2844 	struct sk_buff *syn_data = NULL, *data;
2845 	unsigned long last_syn_loss = 0;
2846 
2847 	tp->rx_opt.mss_clamp = tp->advmss;  /* If MSS is not cached */
2848 	tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
2849 			       &syn_loss, &last_syn_loss);
2850 	/* Recurring FO SYN losses: revert to regular handshake temporarily */
2851 	if (syn_loss > 1 &&
2852 	    time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
2853 		fo->cookie.len = -1;
2854 		goto fallback;
2855 	}
2856 
2857 	if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
2858 		fo->cookie.len = -1;
2859 	else if (fo->cookie.len <= 0)
2860 		goto fallback;
2861 
2862 	/* MSS for SYN-data is based on cached MSS and bounded by PMTU and
2863 	 * user-MSS. Reserve maximum option space for middleboxes that add
2864 	 * private TCP options. The cost is reduced data space in SYN :(
2865 	 */
2866 	if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
2867 		tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2868 	space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
2869 		MAX_TCP_OPTION_SPACE;
2870 
2871 	syn_data = skb_copy_expand(syn, skb_headroom(syn), space,
2872 				   sk->sk_allocation);
2873 	if (syn_data == NULL)
2874 		goto fallback;
2875 
2876 	for (i = 0; i < iovlen && syn_data->len < space; ++i) {
2877 		struct iovec *iov = &fo->data->msg_iov[i];
2878 		unsigned char __user *from = iov->iov_base;
2879 		int len = iov->iov_len;
2880 
2881 		if (syn_data->len + len > space)
2882 			len = space - syn_data->len;
2883 		else if (i + 1 == iovlen)
2884 			/* No more data pending in inet_wait_for_connect() */
2885 			fo->data = NULL;
2886 
2887 		if (skb_add_data(syn_data, from, len))
2888 			goto fallback;
2889 	}
2890 
2891 	/* Queue a data-only packet after the regular SYN for retransmission */
2892 	data = pskb_copy(syn_data, sk->sk_allocation);
2893 	if (data == NULL)
2894 		goto fallback;
2895 	TCP_SKB_CB(data)->seq++;
2896 	TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN;
2897 	TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH);
2898 	tcp_connect_queue_skb(sk, data);
2899 	fo->copied = data->len;
2900 
2901 	if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) {
2902 		tp->syn_data = (fo->copied > 0);
2903 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
2904 		goto done;
2905 	}
2906 	syn_data = NULL;
2907 
2908 fallback:
2909 	/* Send a regular SYN with Fast Open cookie request option */
2910 	if (fo->cookie.len > 0)
2911 		fo->cookie.len = 0;
2912 	err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
2913 	if (err)
2914 		tp->syn_fastopen = 0;
2915 	kfree_skb(syn_data);
2916 done:
2917 	fo->cookie.len = -1;  /* Exclude Fast Open option for SYN retries */
2918 	return err;
2919 }
2920 
2921 /* Build a SYN and send it off. */
tcp_connect(struct sock * sk)2922 int tcp_connect(struct sock *sk)
2923 {
2924 	struct tcp_sock *tp = tcp_sk(sk);
2925 	struct sk_buff *buff;
2926 	int err;
2927 
2928 	tcp_connect_init(sk);
2929 
2930 	if (unlikely(tp->repair)) {
2931 		tcp_finish_connect(sk, NULL);
2932 		return 0;
2933 	}
2934 
2935 	buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2936 	if (unlikely(buff == NULL))
2937 		return -ENOBUFS;
2938 
2939 	/* Reserve space for headers. */
2940 	skb_reserve(buff, MAX_TCP_HEADER);
2941 
2942 	tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
2943 	tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp;
2944 	tcp_connect_queue_skb(sk, buff);
2945 	TCP_ECN_send_syn(sk, buff);
2946 
2947 	/* Send off SYN; include data in Fast Open. */
2948 	err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
2949 	      tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
2950 	if (err == -ECONNREFUSED)
2951 		return err;
2952 
2953 	/* We change tp->snd_nxt after the tcp_transmit_skb() call
2954 	 * in order to make this packet get counted in tcpOutSegs.
2955 	 */
2956 	tp->snd_nxt = tp->write_seq;
2957 	tp->pushed_seq = tp->write_seq;
2958 	TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
2959 
2960 	/* Timer for repeating the SYN until an answer. */
2961 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2962 				  inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2963 	return 0;
2964 }
2965 EXPORT_SYMBOL(tcp_connect);
2966 
2967 /* Send out a delayed ack, the caller does the policy checking
2968  * to see if we should even be here.  See tcp_input.c:tcp_ack_snd_check()
2969  * for details.
2970  */
tcp_send_delayed_ack(struct sock * sk)2971 void tcp_send_delayed_ack(struct sock *sk)
2972 {
2973 	struct inet_connection_sock *icsk = inet_csk(sk);
2974 	int ato = icsk->icsk_ack.ato;
2975 	unsigned long timeout;
2976 
2977 	if (ato > TCP_DELACK_MIN) {
2978 		const struct tcp_sock *tp = tcp_sk(sk);
2979 		int max_ato = HZ / 2;
2980 
2981 		if (icsk->icsk_ack.pingpong ||
2982 		    (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
2983 			max_ato = TCP_DELACK_MAX;
2984 
2985 		/* Slow path, intersegment interval is "high". */
2986 
2987 		/* If some rtt estimate is known, use it to bound delayed ack.
2988 		 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2989 		 * directly.
2990 		 */
2991 		if (tp->srtt) {
2992 			int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
2993 
2994 			if (rtt < max_ato)
2995 				max_ato = rtt;
2996 		}
2997 
2998 		ato = min(ato, max_ato);
2999 	}
3000 
3001 	/* Stay within the limit we were given */
3002 	timeout = jiffies + ato;
3003 
3004 	/* Use new timeout only if there wasn't a older one earlier. */
3005 	if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3006 		/* If delack timer was blocked or is about to expire,
3007 		 * send ACK now.
3008 		 */
3009 		if (icsk->icsk_ack.blocked ||
3010 		    time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3011 			tcp_send_ack(sk);
3012 			return;
3013 		}
3014 
3015 		if (!time_before(timeout, icsk->icsk_ack.timeout))
3016 			timeout = icsk->icsk_ack.timeout;
3017 	}
3018 	icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3019 	icsk->icsk_ack.timeout = timeout;
3020 	sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3021 }
3022 
3023 /* This routine sends an ack and also updates the window. */
tcp_send_ack(struct sock * sk)3024 void tcp_send_ack(struct sock *sk)
3025 {
3026 	struct sk_buff *buff;
3027 
3028 	/* If we have been reset, we may not send again. */
3029 	if (sk->sk_state == TCP_CLOSE)
3030 		return;
3031 
3032 	/* We are not putting this on the write queue, so
3033 	 * tcp_transmit_skb() will set the ownership to this
3034 	 * sock.
3035 	 */
3036 	buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3037 	if (buff == NULL) {
3038 		inet_csk_schedule_ack(sk);
3039 		inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3040 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3041 					  TCP_DELACK_MAX, TCP_RTO_MAX);
3042 		return;
3043 	}
3044 
3045 	/* Reserve space for headers and prepare control bits. */
3046 	skb_reserve(buff, MAX_TCP_HEADER);
3047 	tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3048 
3049 	/* Send it off, this clears delayed acks for us. */
3050 	TCP_SKB_CB(buff)->when = tcp_time_stamp;
3051 	tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC));
3052 }
3053 
3054 /* This routine sends a packet with an out of date sequence
3055  * number. It assumes the other end will try to ack it.
3056  *
3057  * Question: what should we make while urgent mode?
3058  * 4.4BSD forces sending single byte of data. We cannot send
3059  * out of window data, because we have SND.NXT==SND.MAX...
3060  *
3061  * Current solution: to send TWO zero-length segments in urgent mode:
3062  * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3063  * out-of-date with SND.UNA-1 to probe window.
3064  */
tcp_xmit_probe_skb(struct sock * sk,int urgent)3065 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
3066 {
3067 	struct tcp_sock *tp = tcp_sk(sk);
3068 	struct sk_buff *skb;
3069 
3070 	/* We don't queue it, tcp_transmit_skb() sets ownership. */
3071 	skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3072 	if (skb == NULL)
3073 		return -1;
3074 
3075 	/* Reserve space for headers and set control bits. */
3076 	skb_reserve(skb, MAX_TCP_HEADER);
3077 	/* Use a previous sequence.  This should cause the other
3078 	 * end to send an ack.  Don't queue or clone SKB, just
3079 	 * send it.
3080 	 */
3081 	tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3082 	TCP_SKB_CB(skb)->when = tcp_time_stamp;
3083 	return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
3084 }
3085 
tcp_send_window_probe(struct sock * sk)3086 void tcp_send_window_probe(struct sock *sk)
3087 {
3088 	if (sk->sk_state == TCP_ESTABLISHED) {
3089 		tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3090 		tcp_sk(sk)->snd_nxt = tcp_sk(sk)->write_seq;
3091 		tcp_xmit_probe_skb(sk, 0);
3092 	}
3093 }
3094 
3095 /* Initiate keepalive or window probe from timer. */
tcp_write_wakeup(struct sock * sk)3096 int tcp_write_wakeup(struct sock *sk)
3097 {
3098 	struct tcp_sock *tp = tcp_sk(sk);
3099 	struct sk_buff *skb;
3100 
3101 	if (sk->sk_state == TCP_CLOSE)
3102 		return -1;
3103 
3104 	if ((skb = tcp_send_head(sk)) != NULL &&
3105 	    before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3106 		int err;
3107 		unsigned int mss = tcp_current_mss(sk);
3108 		unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3109 
3110 		if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3111 			tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3112 
3113 		/* We are probing the opening of a window
3114 		 * but the window size is != 0
3115 		 * must have been a result SWS avoidance ( sender )
3116 		 */
3117 		if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3118 		    skb->len > mss) {
3119 			seg_size = min(seg_size, mss);
3120 			TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3121 			if (tcp_fragment(sk, skb, seg_size, mss))
3122 				return -1;
3123 		} else if (!tcp_skb_pcount(skb))
3124 			tcp_set_skb_tso_segs(sk, skb, mss);
3125 
3126 		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3127 		TCP_SKB_CB(skb)->when = tcp_time_stamp;
3128 		err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3129 		if (!err)
3130 			tcp_event_new_data_sent(sk, skb);
3131 		return err;
3132 	} else {
3133 		if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3134 			tcp_xmit_probe_skb(sk, 1);
3135 		return tcp_xmit_probe_skb(sk, 0);
3136 	}
3137 }
3138 
3139 /* A window probe timeout has occurred.  If window is not closed send
3140  * a partial packet else a zero probe.
3141  */
tcp_send_probe0(struct sock * sk)3142 void tcp_send_probe0(struct sock *sk)
3143 {
3144 	struct inet_connection_sock *icsk = inet_csk(sk);
3145 	struct tcp_sock *tp = tcp_sk(sk);
3146 	int err;
3147 
3148 	err = tcp_write_wakeup(sk);
3149 
3150 	if (tp->packets_out || !tcp_send_head(sk)) {
3151 		/* Cancel probe timer, if it is not required. */
3152 		icsk->icsk_probes_out = 0;
3153 		icsk->icsk_backoff = 0;
3154 		return;
3155 	}
3156 
3157 	if (err <= 0) {
3158 		if (icsk->icsk_backoff < sysctl_tcp_retries2)
3159 			icsk->icsk_backoff++;
3160 		icsk->icsk_probes_out++;
3161 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3162 					  min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3163 					  TCP_RTO_MAX);
3164 	} else {
3165 		/* If packet was not sent due to local congestion,
3166 		 * do not backoff and do not remember icsk_probes_out.
3167 		 * Let local senders to fight for local resources.
3168 		 *
3169 		 * Use accumulated backoff yet.
3170 		 */
3171 		if (!icsk->icsk_probes_out)
3172 			icsk->icsk_probes_out = 1;
3173 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3174 					  min(icsk->icsk_rto << icsk->icsk_backoff,
3175 					      TCP_RESOURCE_PROBE_INTERVAL),
3176 					  TCP_RTO_MAX);
3177 	}
3178 }
3179