• 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 four TSO segments */
54 int sysctl_tcp_limit_output_bytes __read_mostly = 262144;
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 /* By default, RFC2861 behavior.  */
63 int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
64 
65 unsigned int sysctl_tcp_notsent_lowat __read_mostly = UINT_MAX;
66 EXPORT_SYMBOL(sysctl_tcp_notsent_lowat);
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 	NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
88 		      tcp_skb_pcount(skb));
89 }
90 
91 /* SND.NXT, if window was not shrunk.
92  * If window has been shrunk, what should we make? It is not clear at all.
93  * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
94  * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
95  * invalid. OK, let's make this for now:
96  */
tcp_acceptable_seq(const struct sock * sk)97 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
98 {
99 	const struct tcp_sock *tp = tcp_sk(sk);
100 
101 	if (!before(tcp_wnd_end(tp), tp->snd_nxt))
102 		return tp->snd_nxt;
103 	else
104 		return tcp_wnd_end(tp);
105 }
106 
107 /* Calculate mss to advertise in SYN segment.
108  * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
109  *
110  * 1. It is independent of path mtu.
111  * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
112  * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
113  *    attached devices, because some buggy hosts are confused by
114  *    large MSS.
115  * 4. We do not make 3, we advertise MSS, calculated from first
116  *    hop device mtu, but allow to raise it to ip_rt_min_advmss.
117  *    This may be overridden via information stored in routing table.
118  * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
119  *    probably even Jumbo".
120  */
tcp_advertise_mss(struct sock * sk)121 static __u16 tcp_advertise_mss(struct sock *sk)
122 {
123 	struct tcp_sock *tp = tcp_sk(sk);
124 	const struct dst_entry *dst = __sk_dst_get(sk);
125 	int mss = tp->advmss;
126 
127 	if (dst) {
128 		unsigned int metric = dst_metric_advmss(dst);
129 
130 		if (metric < mss) {
131 			mss = metric;
132 			tp->advmss = mss;
133 		}
134 	}
135 
136 	return (__u16)mss;
137 }
138 
139 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
140  * This is the first part of cwnd validation mechanism.
141  */
tcp_cwnd_restart(struct sock * sk,s32 delta)142 void tcp_cwnd_restart(struct sock *sk, s32 delta)
143 {
144 	struct tcp_sock *tp = tcp_sk(sk);
145 	u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
146 	u32 cwnd = tp->snd_cwnd;
147 
148 	tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
149 
150 	tp->snd_ssthresh = tcp_current_ssthresh(sk);
151 	restart_cwnd = min(restart_cwnd, cwnd);
152 
153 	while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
154 		cwnd >>= 1;
155 	tp->snd_cwnd = max(cwnd, restart_cwnd);
156 	tp->snd_cwnd_stamp = tcp_time_stamp;
157 	tp->snd_cwnd_used = 0;
158 }
159 
160 /* Congestion state accounting after a packet has been sent. */
tcp_event_data_sent(struct tcp_sock * tp,struct sock * sk)161 static void tcp_event_data_sent(struct tcp_sock *tp,
162 				struct sock *sk)
163 {
164 	struct inet_connection_sock *icsk = inet_csk(sk);
165 	const u32 now = tcp_time_stamp;
166 
167 	if (tcp_packets_in_flight(tp) == 0)
168 		tcp_ca_event(sk, CA_EVENT_TX_START);
169 
170 	tp->lsndtime = now;
171 
172 	/* If it is a reply for ato after last received
173 	 * packet, enter pingpong mode.
174 	 */
175 	if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
176 		icsk->icsk_ack.pingpong = 1;
177 }
178 
179 /* Account for an ACK we sent. */
tcp_event_ack_sent(struct sock * sk,unsigned int pkts,u32 rcv_nxt)180 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts,
181 				      u32 rcv_nxt)
182 {
183 	struct tcp_sock *tp = tcp_sk(sk);
184 
185 	if (unlikely(rcv_nxt != tp->rcv_nxt))
186 		return;  /* Special ACK sent by DCTCP to reflect ECN */
187 	tcp_dec_quickack_mode(sk, pkts);
188 	inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
189 }
190 
191 
tcp_default_init_rwnd(u32 mss)192 u32 tcp_default_init_rwnd(u32 mss)
193 {
194 	/* Initial receive window should be twice of TCP_INIT_CWND to
195 	 * enable proper sending of new unsent data during fast recovery
196 	 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
197 	 * limit when mss is larger than 1460.
198 	 */
199 	u32 init_rwnd = sysctl_tcp_default_init_rwnd;
200 
201 	if (mss > 1460)
202 		init_rwnd = max((1460 * init_rwnd) / mss, 2U);
203 	return init_rwnd;
204 }
205 
206 /* Determine a window scaling and initial window to offer.
207  * Based on the assumption that the given amount of space
208  * will be offered. Store the results in the tp structure.
209  * NOTE: for smooth operation initial space offering should
210  * be a multiple of mss if possible. We assume here that mss >= 1.
211  * This MUST be enforced by all callers.
212  */
tcp_select_initial_window(int __space,__u32 mss,__u32 * rcv_wnd,__u32 * window_clamp,int wscale_ok,__u8 * rcv_wscale,__u32 init_rcv_wnd)213 void tcp_select_initial_window(int __space, __u32 mss,
214 			       __u32 *rcv_wnd, __u32 *window_clamp,
215 			       int wscale_ok, __u8 *rcv_wscale,
216 			       __u32 init_rcv_wnd)
217 {
218 	unsigned int space = (__space < 0 ? 0 : __space);
219 
220 	/* If no clamp set the clamp to the max possible scaled window */
221 	if (*window_clamp == 0)
222 		(*window_clamp) = (65535 << 14);
223 	space = min(*window_clamp, space);
224 
225 	/* Quantize space offering to a multiple of mss if possible. */
226 	if (space > mss)
227 		space = (space / mss) * mss;
228 
229 	/* NOTE: offering an initial window larger than 32767
230 	 * will break some buggy TCP stacks. If the admin tells us
231 	 * it is likely we could be speaking with such a buggy stack
232 	 * we will truncate our initial window offering to 32K-1
233 	 * unless the remote has sent us a window scaling option,
234 	 * which we interpret as a sign the remote TCP is not
235 	 * misinterpreting the window field as a signed quantity.
236 	 */
237 	if (sysctl_tcp_workaround_signed_windows)
238 		(*rcv_wnd) = min(space, MAX_TCP_WINDOW);
239 	else
240 		(*rcv_wnd) = space;
241 
242 	(*rcv_wscale) = 0;
243 	if (wscale_ok) {
244 		/* Set window scaling on max possible window
245 		 * See RFC1323 for an explanation of the limit to 14
246 		 */
247 		space = max_t(u32, space, sysctl_tcp_rmem[2]);
248 		space = max_t(u32, space, sysctl_rmem_max);
249 		space = min_t(u32, space, *window_clamp);
250 		while (space > 65535 && (*rcv_wscale) < 14) {
251 			space >>= 1;
252 			(*rcv_wscale)++;
253 		}
254 	}
255 
256 	if (mss > (1 << *rcv_wscale)) {
257 		if (!init_rcv_wnd) /* Use default unless specified otherwise */
258 			init_rcv_wnd = tcp_default_init_rwnd(mss);
259 		*rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
260 	}
261 
262 	/* Set the clamp no higher than max representable value */
263 	(*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
264 }
265 EXPORT_SYMBOL(tcp_select_initial_window);
266 
267 /* Chose a new window to advertise, update state in tcp_sock for the
268  * socket, and return result with RFC1323 scaling applied.  The return
269  * value can be stuffed directly into th->window for an outgoing
270  * frame.
271  */
tcp_select_window(struct sock * sk)272 static u16 tcp_select_window(struct sock *sk)
273 {
274 	struct tcp_sock *tp = tcp_sk(sk);
275 	u32 old_win = tp->rcv_wnd;
276 	u32 cur_win = tcp_receive_window(tp);
277 	u32 new_win = __tcp_select_window(sk);
278 
279 	/* Never shrink the offered window */
280 	if (new_win < cur_win) {
281 		/* Danger Will Robinson!
282 		 * Don't update rcv_wup/rcv_wnd here or else
283 		 * we will not be able to advertise a zero
284 		 * window in time.  --DaveM
285 		 *
286 		 * Relax Will Robinson.
287 		 */
288 		if (new_win == 0)
289 			NET_INC_STATS(sock_net(sk),
290 				      LINUX_MIB_TCPWANTZEROWINDOWADV);
291 		new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
292 	}
293 	tp->rcv_wnd = new_win;
294 	tp->rcv_wup = tp->rcv_nxt;
295 
296 	/* Make sure we do not exceed the maximum possible
297 	 * scaled window.
298 	 */
299 	if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
300 		new_win = min(new_win, MAX_TCP_WINDOW);
301 	else
302 		new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
303 
304 	/* RFC1323 scaling applied */
305 	new_win >>= tp->rx_opt.rcv_wscale;
306 
307 	/* If we advertise zero window, disable fast path. */
308 	if (new_win == 0) {
309 		tp->pred_flags = 0;
310 		if (old_win)
311 			NET_INC_STATS(sock_net(sk),
312 				      LINUX_MIB_TCPTOZEROWINDOWADV);
313 	} else if (old_win == 0) {
314 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
315 	}
316 
317 	return new_win;
318 }
319 
320 /* Packet ECN state for a SYN-ACK */
tcp_ecn_send_synack(struct sock * sk,struct sk_buff * skb)321 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
322 {
323 	const struct tcp_sock *tp = tcp_sk(sk);
324 
325 	TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
326 	if (!(tp->ecn_flags & TCP_ECN_OK))
327 		TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
328 	else if (tcp_ca_needs_ecn(sk))
329 		INET_ECN_xmit(sk);
330 }
331 
332 /* Packet ECN state for a SYN.  */
tcp_ecn_send_syn(struct sock * sk,struct sk_buff * skb)333 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
334 {
335 	struct tcp_sock *tp = tcp_sk(sk);
336 	bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
337 		       tcp_ca_needs_ecn(sk);
338 
339 	if (!use_ecn) {
340 		const struct dst_entry *dst = __sk_dst_get(sk);
341 
342 		if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
343 			use_ecn = true;
344 	}
345 
346 	tp->ecn_flags = 0;
347 
348 	if (use_ecn) {
349 		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
350 		tp->ecn_flags = TCP_ECN_OK;
351 		if (tcp_ca_needs_ecn(sk))
352 			INET_ECN_xmit(sk);
353 	}
354 }
355 
tcp_ecn_clear_syn(struct sock * sk,struct sk_buff * skb)356 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
357 {
358 	if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
359 		/* tp->ecn_flags are cleared at a later point in time when
360 		 * SYN ACK is ultimatively being received.
361 		 */
362 		TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
363 }
364 
365 static void
tcp_ecn_make_synack(const struct request_sock * req,struct tcphdr * th)366 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
367 {
368 	if (inet_rsk(req)->ecn_ok)
369 		th->ece = 1;
370 }
371 
372 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
373  * be sent.
374  */
tcp_ecn_send(struct sock * sk,struct sk_buff * skb,int tcp_header_len)375 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
376 				int tcp_header_len)
377 {
378 	struct tcp_sock *tp = tcp_sk(sk);
379 
380 	if (tp->ecn_flags & TCP_ECN_OK) {
381 		/* Not-retransmitted data segment: set ECT and inject CWR. */
382 		if (skb->len != tcp_header_len &&
383 		    !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
384 			INET_ECN_xmit(sk);
385 			if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
386 				tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
387 				tcp_hdr(skb)->cwr = 1;
388 				skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
389 			}
390 		} else if (!tcp_ca_needs_ecn(sk)) {
391 			/* ACK or retransmitted segment: clear ECT|CE */
392 			INET_ECN_dontxmit(sk);
393 		}
394 		if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
395 			tcp_hdr(skb)->ece = 1;
396 	}
397 }
398 
399 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
400  * auto increment end seqno.
401  */
tcp_init_nondata_skb(struct sk_buff * skb,u32 seq,u8 flags)402 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
403 {
404 	skb->ip_summed = CHECKSUM_PARTIAL;
405 	skb->csum = 0;
406 
407 	TCP_SKB_CB(skb)->tcp_flags = flags;
408 	TCP_SKB_CB(skb)->sacked = 0;
409 
410 	tcp_skb_pcount_set(skb, 1);
411 
412 	TCP_SKB_CB(skb)->seq = seq;
413 	if (flags & (TCPHDR_SYN | TCPHDR_FIN))
414 		seq++;
415 	TCP_SKB_CB(skb)->end_seq = seq;
416 }
417 
tcp_urg_mode(const struct tcp_sock * tp)418 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
419 {
420 	return tp->snd_una != tp->snd_up;
421 }
422 
423 #define OPTION_SACK_ADVERTISE	(1 << 0)
424 #define OPTION_TS		(1 << 1)
425 #define OPTION_MD5		(1 << 2)
426 #define OPTION_WSCALE		(1 << 3)
427 #define OPTION_FAST_OPEN_COOKIE	(1 << 8)
428 
429 struct tcp_out_options {
430 	u16 options;		/* bit field of OPTION_* */
431 	u16 mss;		/* 0 to disable */
432 	u8 ws;			/* window scale, 0 to disable */
433 	u8 num_sack_blocks;	/* number of SACK blocks to include */
434 	u8 hash_size;		/* bytes in hash_location */
435 	__u8 *hash_location;	/* temporary pointer, overloaded */
436 	__u32 tsval, tsecr;	/* need to include OPTION_TS */
437 	struct tcp_fastopen_cookie *fastopen_cookie;	/* Fast open cookie */
438 };
439 
440 /* Write previously computed TCP options to the packet.
441  *
442  * Beware: Something in the Internet is very sensitive to the ordering of
443  * TCP options, we learned this through the hard way, so be careful here.
444  * Luckily we can at least blame others for their non-compliance but from
445  * inter-operability perspective it seems that we're somewhat stuck with
446  * the ordering which we have been using if we want to keep working with
447  * those broken things (not that it currently hurts anybody as there isn't
448  * particular reason why the ordering would need to be changed).
449  *
450  * At least SACK_PERM as the first option is known to lead to a disaster
451  * (but it may well be that other scenarios fail similarly).
452  */
tcp_options_write(__be32 * ptr,struct tcp_sock * tp,struct tcp_out_options * opts)453 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
454 			      struct tcp_out_options *opts)
455 {
456 	u16 options = opts->options;	/* mungable copy */
457 
458 	if (unlikely(OPTION_MD5 & options)) {
459 		*ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
460 			       (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
461 		/* overload cookie hash location */
462 		opts->hash_location = (__u8 *)ptr;
463 		ptr += 4;
464 	}
465 
466 	if (unlikely(opts->mss)) {
467 		*ptr++ = htonl((TCPOPT_MSS << 24) |
468 			       (TCPOLEN_MSS << 16) |
469 			       opts->mss);
470 	}
471 
472 	if (likely(OPTION_TS & options)) {
473 		if (unlikely(OPTION_SACK_ADVERTISE & options)) {
474 			*ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
475 				       (TCPOLEN_SACK_PERM << 16) |
476 				       (TCPOPT_TIMESTAMP << 8) |
477 				       TCPOLEN_TIMESTAMP);
478 			options &= ~OPTION_SACK_ADVERTISE;
479 		} else {
480 			*ptr++ = htonl((TCPOPT_NOP << 24) |
481 				       (TCPOPT_NOP << 16) |
482 				       (TCPOPT_TIMESTAMP << 8) |
483 				       TCPOLEN_TIMESTAMP);
484 		}
485 		*ptr++ = htonl(opts->tsval);
486 		*ptr++ = htonl(opts->tsecr);
487 	}
488 
489 	if (unlikely(OPTION_SACK_ADVERTISE & options)) {
490 		*ptr++ = htonl((TCPOPT_NOP << 24) |
491 			       (TCPOPT_NOP << 16) |
492 			       (TCPOPT_SACK_PERM << 8) |
493 			       TCPOLEN_SACK_PERM);
494 	}
495 
496 	if (unlikely(OPTION_WSCALE & options)) {
497 		*ptr++ = htonl((TCPOPT_NOP << 24) |
498 			       (TCPOPT_WINDOW << 16) |
499 			       (TCPOLEN_WINDOW << 8) |
500 			       opts->ws);
501 	}
502 
503 	if (unlikely(opts->num_sack_blocks)) {
504 		struct tcp_sack_block *sp = tp->rx_opt.dsack ?
505 			tp->duplicate_sack : tp->selective_acks;
506 		int this_sack;
507 
508 		*ptr++ = htonl((TCPOPT_NOP  << 24) |
509 			       (TCPOPT_NOP  << 16) |
510 			       (TCPOPT_SACK <<  8) |
511 			       (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
512 						     TCPOLEN_SACK_PERBLOCK)));
513 
514 		for (this_sack = 0; this_sack < opts->num_sack_blocks;
515 		     ++this_sack) {
516 			*ptr++ = htonl(sp[this_sack].start_seq);
517 			*ptr++ = htonl(sp[this_sack].end_seq);
518 		}
519 
520 		tp->rx_opt.dsack = 0;
521 	}
522 
523 	if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
524 		struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
525 		u8 *p = (u8 *)ptr;
526 		u32 len; /* Fast Open option length */
527 
528 		if (foc->exp) {
529 			len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
530 			*ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
531 				     TCPOPT_FASTOPEN_MAGIC);
532 			p += TCPOLEN_EXP_FASTOPEN_BASE;
533 		} else {
534 			len = TCPOLEN_FASTOPEN_BASE + foc->len;
535 			*p++ = TCPOPT_FASTOPEN;
536 			*p++ = len;
537 		}
538 
539 		memcpy(p, foc->val, foc->len);
540 		if ((len & 3) == 2) {
541 			p[foc->len] = TCPOPT_NOP;
542 			p[foc->len + 1] = TCPOPT_NOP;
543 		}
544 		ptr += (len + 3) >> 2;
545 	}
546 }
547 
548 /* Compute TCP options for SYN packets. This is not the final
549  * network wire format yet.
550  */
tcp_syn_options(struct sock * sk,struct sk_buff * skb,struct tcp_out_options * opts,struct tcp_md5sig_key ** md5)551 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
552 				struct tcp_out_options *opts,
553 				struct tcp_md5sig_key **md5)
554 {
555 	struct tcp_sock *tp = tcp_sk(sk);
556 	unsigned int remaining = MAX_TCP_OPTION_SPACE;
557 	struct tcp_fastopen_request *fastopen = tp->fastopen_req;
558 
559 #ifdef CONFIG_TCP_MD5SIG
560 	*md5 = tp->af_specific->md5_lookup(sk, sk);
561 	if (*md5) {
562 		opts->options |= OPTION_MD5;
563 		remaining -= TCPOLEN_MD5SIG_ALIGNED;
564 	}
565 #else
566 	*md5 = NULL;
567 #endif
568 
569 	/* We always get an MSS option.  The option bytes which will be seen in
570 	 * normal data packets should timestamps be used, must be in the MSS
571 	 * advertised.  But we subtract them from tp->mss_cache so that
572 	 * calculations in tcp_sendmsg are simpler etc.  So account for this
573 	 * fact here if necessary.  If we don't do this correctly, as a
574 	 * receiver we won't recognize data packets as being full sized when we
575 	 * should, and thus we won't abide by the delayed ACK rules correctly.
576 	 * SACKs don't matter, we never delay an ACK when we have any of those
577 	 * going out.  */
578 	opts->mss = tcp_advertise_mss(sk);
579 	remaining -= TCPOLEN_MSS_ALIGNED;
580 
581 	if (likely(sysctl_tcp_timestamps && !*md5)) {
582 		opts->options |= OPTION_TS;
583 		opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
584 		opts->tsecr = tp->rx_opt.ts_recent;
585 		remaining -= TCPOLEN_TSTAMP_ALIGNED;
586 	}
587 	if (likely(sysctl_tcp_window_scaling)) {
588 		opts->ws = tp->rx_opt.rcv_wscale;
589 		opts->options |= OPTION_WSCALE;
590 		remaining -= TCPOLEN_WSCALE_ALIGNED;
591 	}
592 	if (likely(sysctl_tcp_sack)) {
593 		opts->options |= OPTION_SACK_ADVERTISE;
594 		if (unlikely(!(OPTION_TS & opts->options)))
595 			remaining -= TCPOLEN_SACKPERM_ALIGNED;
596 	}
597 
598 	if (fastopen && fastopen->cookie.len >= 0) {
599 		u32 need = fastopen->cookie.len;
600 
601 		need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
602 					       TCPOLEN_FASTOPEN_BASE;
603 		need = (need + 3) & ~3U;  /* Align to 32 bits */
604 		if (remaining >= need) {
605 			opts->options |= OPTION_FAST_OPEN_COOKIE;
606 			opts->fastopen_cookie = &fastopen->cookie;
607 			remaining -= need;
608 			tp->syn_fastopen = 1;
609 			tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
610 		}
611 	}
612 
613 	return MAX_TCP_OPTION_SPACE - remaining;
614 }
615 
616 /* Set up TCP options for SYN-ACKs. */
tcp_synack_options(struct request_sock * req,unsigned int mss,struct sk_buff * skb,struct tcp_out_options * opts,const struct tcp_md5sig_key * md5,struct tcp_fastopen_cookie * foc)617 static unsigned int tcp_synack_options(struct request_sock *req,
618 				       unsigned int mss, struct sk_buff *skb,
619 				       struct tcp_out_options *opts,
620 				       const struct tcp_md5sig_key *md5,
621 				       struct tcp_fastopen_cookie *foc)
622 {
623 	struct inet_request_sock *ireq = inet_rsk(req);
624 	unsigned int remaining = MAX_TCP_OPTION_SPACE;
625 
626 #ifdef CONFIG_TCP_MD5SIG
627 	if (md5) {
628 		opts->options |= OPTION_MD5;
629 		remaining -= TCPOLEN_MD5SIG_ALIGNED;
630 
631 		/* We can't fit any SACK blocks in a packet with MD5 + TS
632 		 * options. There was discussion about disabling SACK
633 		 * rather than TS in order to fit in better with old,
634 		 * buggy kernels, but that was deemed to be unnecessary.
635 		 */
636 		ireq->tstamp_ok &= !ireq->sack_ok;
637 	}
638 #endif
639 
640 	/* We always send an MSS option. */
641 	opts->mss = mss;
642 	remaining -= TCPOLEN_MSS_ALIGNED;
643 
644 	if (likely(ireq->wscale_ok)) {
645 		opts->ws = ireq->rcv_wscale;
646 		opts->options |= OPTION_WSCALE;
647 		remaining -= TCPOLEN_WSCALE_ALIGNED;
648 	}
649 	if (likely(ireq->tstamp_ok)) {
650 		opts->options |= OPTION_TS;
651 		opts->tsval = tcp_skb_timestamp(skb);
652 		opts->tsecr = req->ts_recent;
653 		remaining -= TCPOLEN_TSTAMP_ALIGNED;
654 	}
655 	if (likely(ireq->sack_ok)) {
656 		opts->options |= OPTION_SACK_ADVERTISE;
657 		if (unlikely(!ireq->tstamp_ok))
658 			remaining -= TCPOLEN_SACKPERM_ALIGNED;
659 	}
660 	if (foc != NULL && foc->len >= 0) {
661 		u32 need = foc->len;
662 
663 		need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
664 				   TCPOLEN_FASTOPEN_BASE;
665 		need = (need + 3) & ~3U;  /* Align to 32 bits */
666 		if (remaining >= need) {
667 			opts->options |= OPTION_FAST_OPEN_COOKIE;
668 			opts->fastopen_cookie = foc;
669 			remaining -= need;
670 		}
671 	}
672 
673 	return MAX_TCP_OPTION_SPACE - remaining;
674 }
675 
676 /* Compute TCP options for ESTABLISHED sockets. This is not the
677  * final wire format yet.
678  */
tcp_established_options(struct sock * sk,struct sk_buff * skb,struct tcp_out_options * opts,struct tcp_md5sig_key ** md5)679 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
680 					struct tcp_out_options *opts,
681 					struct tcp_md5sig_key **md5)
682 {
683 	struct tcp_sock *tp = tcp_sk(sk);
684 	unsigned int size = 0;
685 	unsigned int eff_sacks;
686 
687 	opts->options = 0;
688 
689 #ifdef CONFIG_TCP_MD5SIG
690 	*md5 = tp->af_specific->md5_lookup(sk, sk);
691 	if (unlikely(*md5)) {
692 		opts->options |= OPTION_MD5;
693 		size += TCPOLEN_MD5SIG_ALIGNED;
694 	}
695 #else
696 	*md5 = NULL;
697 #endif
698 
699 	if (likely(tp->rx_opt.tstamp_ok)) {
700 		opts->options |= OPTION_TS;
701 		opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
702 		opts->tsecr = tp->rx_opt.ts_recent;
703 		size += TCPOLEN_TSTAMP_ALIGNED;
704 	}
705 
706 	eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
707 	if (unlikely(eff_sacks)) {
708 		const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
709 		opts->num_sack_blocks =
710 			min_t(unsigned int, eff_sacks,
711 			      (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
712 			      TCPOLEN_SACK_PERBLOCK);
713 		if (likely(opts->num_sack_blocks))
714 			size += TCPOLEN_SACK_BASE_ALIGNED +
715 				opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
716 	}
717 
718 	return size;
719 }
720 
721 
722 /* TCP SMALL QUEUES (TSQ)
723  *
724  * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
725  * to reduce RTT and bufferbloat.
726  * We do this using a special skb destructor (tcp_wfree).
727  *
728  * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
729  * needs to be reallocated in a driver.
730  * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
731  *
732  * Since transmit from skb destructor is forbidden, we use a tasklet
733  * to process all sockets that eventually need to send more skbs.
734  * We use one tasklet per cpu, with its own queue of sockets.
735  */
736 struct tsq_tasklet {
737 	struct tasklet_struct	tasklet;
738 	struct list_head	head; /* queue of tcp sockets */
739 };
740 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
741 
tcp_tsq_handler(struct sock * sk)742 static void tcp_tsq_handler(struct sock *sk)
743 {
744 	if ((1 << sk->sk_state) &
745 	    (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
746 	     TCPF_CLOSE_WAIT  | TCPF_LAST_ACK))
747 		tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle,
748 			       0, GFP_ATOMIC);
749 }
750 /*
751  * One tasklet per cpu tries to send more skbs.
752  * We run in tasklet context but need to disable irqs when
753  * transferring tsq->head because tcp_wfree() might
754  * interrupt us (non NAPI drivers)
755  */
tcp_tasklet_func(unsigned long data)756 static void tcp_tasklet_func(unsigned long data)
757 {
758 	struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
759 	LIST_HEAD(list);
760 	unsigned long flags;
761 	struct list_head *q, *n;
762 	struct tcp_sock *tp;
763 	struct sock *sk;
764 
765 	local_irq_save(flags);
766 	list_splice_init(&tsq->head, &list);
767 	local_irq_restore(flags);
768 
769 	list_for_each_safe(q, n, &list) {
770 		tp = list_entry(q, struct tcp_sock, tsq_node);
771 		list_del(&tp->tsq_node);
772 
773 		sk = (struct sock *)tp;
774 		bh_lock_sock(sk);
775 
776 		if (!sock_owned_by_user(sk)) {
777 			tcp_tsq_handler(sk);
778 		} else {
779 			/* defer the work to tcp_release_cb() */
780 			set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
781 		}
782 		bh_unlock_sock(sk);
783 
784 		clear_bit(TSQ_QUEUED, &tp->tsq_flags);
785 		sk_free(sk);
786 	}
787 }
788 
789 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) |		\
790 			  (1UL << TCP_WRITE_TIMER_DEFERRED) |	\
791 			  (1UL << TCP_DELACK_TIMER_DEFERRED) |	\
792 			  (1UL << TCP_MTU_REDUCED_DEFERRED))
793 /**
794  * tcp_release_cb - tcp release_sock() callback
795  * @sk: socket
796  *
797  * called from release_sock() to perform protocol dependent
798  * actions before socket release.
799  */
tcp_release_cb(struct sock * sk)800 void tcp_release_cb(struct sock *sk)
801 {
802 	struct tcp_sock *tp = tcp_sk(sk);
803 	unsigned long flags, nflags;
804 
805 	/* perform an atomic operation only if at least one flag is set */
806 	do {
807 		flags = tp->tsq_flags;
808 		if (!(flags & TCP_DEFERRED_ALL))
809 			return;
810 		nflags = flags & ~TCP_DEFERRED_ALL;
811 	} while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
812 
813 	if (flags & (1UL << TCP_TSQ_DEFERRED))
814 		tcp_tsq_handler(sk);
815 
816 	/* Here begins the tricky part :
817 	 * We are called from release_sock() with :
818 	 * 1) BH disabled
819 	 * 2) sk_lock.slock spinlock held
820 	 * 3) socket owned by us (sk->sk_lock.owned == 1)
821 	 *
822 	 * But following code is meant to be called from BH handlers,
823 	 * so we should keep BH disabled, but early release socket ownership
824 	 */
825 	sock_release_ownership(sk);
826 
827 	if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
828 		tcp_write_timer_handler(sk);
829 		__sock_put(sk);
830 	}
831 	if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
832 		tcp_delack_timer_handler(sk);
833 		__sock_put(sk);
834 	}
835 	if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
836 		inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
837 		__sock_put(sk);
838 	}
839 }
840 EXPORT_SYMBOL(tcp_release_cb);
841 
tcp_tasklet_init(void)842 void __init tcp_tasklet_init(void)
843 {
844 	int i;
845 
846 	for_each_possible_cpu(i) {
847 		struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
848 
849 		INIT_LIST_HEAD(&tsq->head);
850 		tasklet_init(&tsq->tasklet,
851 			     tcp_tasklet_func,
852 			     (unsigned long)tsq);
853 	}
854 }
855 
856 /*
857  * Write buffer destructor automatically called from kfree_skb.
858  * We can't xmit new skbs from this context, as we might already
859  * hold qdisc lock.
860  */
tcp_wfree(struct sk_buff * skb)861 void tcp_wfree(struct sk_buff *skb)
862 {
863 	struct sock *sk = skb->sk;
864 	struct tcp_sock *tp = tcp_sk(sk);
865 	int wmem;
866 
867 	/* Keep one reference on sk_wmem_alloc.
868 	 * Will be released by sk_free() from here or tcp_tasklet_func()
869 	 */
870 	wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc);
871 
872 	/* If this softirq is serviced by ksoftirqd, we are likely under stress.
873 	 * Wait until our queues (qdisc + devices) are drained.
874 	 * This gives :
875 	 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
876 	 * - chance for incoming ACK (processed by another cpu maybe)
877 	 *   to migrate this flow (skb->ooo_okay will be eventually set)
878 	 */
879 	if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
880 		goto out;
881 
882 	if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
883 	    !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
884 		unsigned long flags;
885 		struct tsq_tasklet *tsq;
886 
887 		/* queue this socket to tasklet queue */
888 		local_irq_save(flags);
889 		tsq = this_cpu_ptr(&tsq_tasklet);
890 		list_add(&tp->tsq_node, &tsq->head);
891 		tasklet_schedule(&tsq->tasklet);
892 		local_irq_restore(flags);
893 		return;
894 	}
895 out:
896 	sk_free(sk);
897 }
898 
899 /* This routine actually transmits TCP packets queued in by
900  * tcp_do_sendmsg().  This is used by both the initial
901  * transmission and possible later retransmissions.
902  * All SKB's seen here are completely headerless.  It is our
903  * job to build the TCP header, and pass the packet down to
904  * IP so it can do the same plus pass the packet off to the
905  * device.
906  *
907  * We are working here with either a clone of the original
908  * SKB, or a fresh unique copy made by the retransmit engine.
909  */
__tcp_transmit_skb(struct sock * sk,struct sk_buff * skb,int clone_it,gfp_t gfp_mask,u32 rcv_nxt)910 static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
911 			      int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
912 {
913 	const struct inet_connection_sock *icsk = inet_csk(sk);
914 	struct inet_sock *inet;
915 	struct tcp_sock *tp;
916 	struct tcp_skb_cb *tcb;
917 	struct tcp_out_options opts;
918 	unsigned int tcp_options_size, tcp_header_size;
919 	struct tcp_md5sig_key *md5;
920 	struct tcphdr *th;
921 	int err;
922 
923 	BUG_ON(!skb || !tcp_skb_pcount(skb));
924 
925 	if (clone_it) {
926 		skb_mstamp_get(&skb->skb_mstamp);
927 
928 		if (unlikely(skb_cloned(skb)))
929 			skb = pskb_copy(skb, gfp_mask);
930 		else
931 			skb = skb_clone(skb, gfp_mask);
932 		if (unlikely(!skb))
933 			return -ENOBUFS;
934 	}
935 
936 	inet = inet_sk(sk);
937 	tp = tcp_sk(sk);
938 	tcb = TCP_SKB_CB(skb);
939 	memset(&opts, 0, sizeof(opts));
940 
941 	if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
942 		tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
943 	else
944 		tcp_options_size = tcp_established_options(sk, skb, &opts,
945 							   &md5);
946 	tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
947 
948 	/* if no packet is in qdisc/device queue, then allow XPS to select
949 	 * another queue. We can be called from tcp_tsq_handler()
950 	 * which holds one reference to sk_wmem_alloc.
951 	 *
952 	 * TODO: Ideally, in-flight pure ACK packets should not matter here.
953 	 * One way to get this would be to set skb->truesize = 2 on them.
954 	 */
955 	skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
956 
957 	skb_push(skb, tcp_header_size);
958 	skb_reset_transport_header(skb);
959 
960 	skb_orphan(skb);
961 	skb->sk = sk;
962 	skb->destructor = skb_is_tcp_pure_ack(skb) ? sock_wfree : tcp_wfree;
963 	skb_set_hash_from_sk(skb, sk);
964 	atomic_add(skb->truesize, &sk->sk_wmem_alloc);
965 
966 	/* Build TCP header and checksum it. */
967 	th = tcp_hdr(skb);
968 	th->source		= inet->inet_sport;
969 	th->dest		= inet->inet_dport;
970 	th->seq			= htonl(tcb->seq);
971 	th->ack_seq		= htonl(rcv_nxt);
972 	*(((__be16 *)th) + 6)	= htons(((tcp_header_size >> 2) << 12) |
973 					tcb->tcp_flags);
974 
975 	if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
976 		/* RFC1323: The window in SYN & SYN/ACK segments
977 		 * is never scaled.
978 		 */
979 		th->window	= htons(min(tp->rcv_wnd, 65535U));
980 	} else {
981 		th->window	= htons(tcp_select_window(sk));
982 	}
983 	th->check		= 0;
984 	th->urg_ptr		= 0;
985 
986 	/* The urg_mode check is necessary during a below snd_una win probe */
987 	if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
988 		if (before(tp->snd_up, tcb->seq + 0x10000)) {
989 			th->urg_ptr = htons(tp->snd_up - tcb->seq);
990 			th->urg = 1;
991 		} else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
992 			th->urg_ptr = htons(0xFFFF);
993 			th->urg = 1;
994 		}
995 	}
996 
997 	tcp_options_write((__be32 *)(th + 1), tp, &opts);
998 	skb_shinfo(skb)->gso_type = sk->sk_gso_type;
999 	if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
1000 		tcp_ecn_send(sk, skb, tcp_header_size);
1001 
1002 #ifdef CONFIG_TCP_MD5SIG
1003 	/* Calculate the MD5 hash, as we have all we need now */
1004 	if (md5) {
1005 		sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1006 		tp->af_specific->calc_md5_hash(opts.hash_location,
1007 					       md5, sk, skb);
1008 	}
1009 #endif
1010 
1011 	icsk->icsk_af_ops->send_check(sk, skb);
1012 
1013 	if (likely(tcb->tcp_flags & TCPHDR_ACK))
1014 		tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt);
1015 
1016 	if (skb->len != tcp_header_size)
1017 		tcp_event_data_sent(tp, sk);
1018 
1019 	if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1020 		TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1021 			      tcp_skb_pcount(skb));
1022 
1023 	tp->segs_out += tcp_skb_pcount(skb);
1024 	/* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1025 	skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1026 	skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1027 
1028 	/* Our usage of tstamp should remain private */
1029 	skb->tstamp.tv64 = 0;
1030 
1031 	/* Cleanup our debris for IP stacks */
1032 	memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1033 			       sizeof(struct inet6_skb_parm)));
1034 
1035 	err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1036 
1037 	if (likely(err <= 0))
1038 		return err;
1039 
1040 	tcp_enter_cwr(sk);
1041 
1042 	return net_xmit_eval(err);
1043 }
1044 
tcp_transmit_skb(struct sock * sk,struct sk_buff * skb,int clone_it,gfp_t gfp_mask)1045 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1046 			    gfp_t gfp_mask)
1047 {
1048 	return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1049 				  tcp_sk(sk)->rcv_nxt);
1050 }
1051 
1052 /* This routine just queues the buffer for sending.
1053  *
1054  * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1055  * otherwise socket can stall.
1056  */
tcp_queue_skb(struct sock * sk,struct sk_buff * skb)1057 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1058 {
1059 	struct tcp_sock *tp = tcp_sk(sk);
1060 
1061 	/* Advance write_seq and place onto the write_queue. */
1062 	tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1063 	__skb_header_release(skb);
1064 	tcp_add_write_queue_tail(sk, skb);
1065 	sk->sk_wmem_queued += skb->truesize;
1066 	sk_mem_charge(sk, skb->truesize);
1067 }
1068 
1069 /* Initialize TSO segments for a packet. */
tcp_set_skb_tso_segs(struct sk_buff * skb,unsigned int mss_now)1070 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1071 {
1072 	if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
1073 		/* Avoid the costly divide in the normal
1074 		 * non-TSO case.
1075 		 */
1076 		tcp_skb_pcount_set(skb, 1);
1077 		TCP_SKB_CB(skb)->tcp_gso_size = 0;
1078 	} else {
1079 		tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1080 		TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1081 	}
1082 }
1083 
1084 /* When a modification to fackets out becomes necessary, we need to check
1085  * skb is counted to fackets_out or not.
1086  */
tcp_adjust_fackets_out(struct sock * sk,const struct sk_buff * skb,int decr)1087 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1088 				   int decr)
1089 {
1090 	struct tcp_sock *tp = tcp_sk(sk);
1091 
1092 	if (!tp->sacked_out || tcp_is_reno(tp))
1093 		return;
1094 
1095 	if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1096 		tp->fackets_out -= decr;
1097 }
1098 
1099 /* Pcount in the middle of the write queue got changed, we need to do various
1100  * tweaks to fix counters
1101  */
tcp_adjust_pcount(struct sock * sk,const struct sk_buff * skb,int decr)1102 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1103 {
1104 	struct tcp_sock *tp = tcp_sk(sk);
1105 
1106 	tp->packets_out -= decr;
1107 
1108 	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1109 		tp->sacked_out -= decr;
1110 	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1111 		tp->retrans_out -= decr;
1112 	if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1113 		tp->lost_out -= decr;
1114 
1115 	/* Reno case is special. Sigh... */
1116 	if (tcp_is_reno(tp) && decr > 0)
1117 		tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1118 
1119 	tcp_adjust_fackets_out(sk, skb, decr);
1120 
1121 	if (tp->lost_skb_hint &&
1122 	    before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1123 	    (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1124 		tp->lost_cnt_hint -= decr;
1125 
1126 	tcp_verify_left_out(tp);
1127 }
1128 
tcp_fragment_tstamp(struct sk_buff * skb,struct sk_buff * skb2)1129 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1130 {
1131 	struct skb_shared_info *shinfo = skb_shinfo(skb);
1132 
1133 	if (unlikely(shinfo->tx_flags & SKBTX_ANY_TSTAMP) &&
1134 	    !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1135 		struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1136 		u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1137 
1138 		shinfo->tx_flags &= ~tsflags;
1139 		shinfo2->tx_flags |= tsflags;
1140 		swap(shinfo->tskey, shinfo2->tskey);
1141 	}
1142 }
1143 
1144 /* Function to create two new TCP segments.  Shrinks the given segment
1145  * to the specified size and appends a new segment with the rest of the
1146  * packet to the list.  This won't be called frequently, I hope.
1147  * Remember, these are still headerless SKBs at this point.
1148  */
tcp_fragment(struct sock * sk,struct sk_buff * skb,u32 len,unsigned int mss_now,gfp_t gfp)1149 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1150 		 unsigned int mss_now, gfp_t gfp)
1151 {
1152 	struct tcp_sock *tp = tcp_sk(sk);
1153 	struct sk_buff *buff;
1154 	int nsize, old_factor;
1155 	long limit;
1156 	int nlen;
1157 	u8 flags;
1158 
1159 	if (WARN_ON(len > skb->len))
1160 		return -EINVAL;
1161 
1162 	nsize = skb_headlen(skb) - len;
1163 	if (nsize < 0)
1164 		nsize = 0;
1165 
1166 	/* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1167 	 * We need some allowance to not penalize applications setting small
1168 	 * SO_SNDBUF values.
1169 	 * Also allow first and last skb in retransmit queue to be split.
1170 	 */
1171 	limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_MAX_SIZE);
1172 	if (unlikely((sk->sk_wmem_queued >> 1) > limit &&
1173 		     skb != tcp_rtx_queue_head(sk) &&
1174 		     skb != tcp_rtx_queue_tail(sk))) {
1175 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
1176 		return -ENOMEM;
1177 	}
1178 
1179 	if (skb_unclone(skb, gfp))
1180 		return -ENOMEM;
1181 
1182 	/* Get a new skb... force flag on. */
1183 	buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1184 	if (!buff)
1185 		return -ENOMEM; /* We'll just try again later. */
1186 
1187 	sk->sk_wmem_queued += buff->truesize;
1188 	sk_mem_charge(sk, buff->truesize);
1189 	nlen = skb->len - len - nsize;
1190 	buff->truesize += nlen;
1191 	skb->truesize -= nlen;
1192 
1193 	/* Correct the sequence numbers. */
1194 	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1195 	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1196 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1197 
1198 	/* PSH and FIN should only be set in the second packet. */
1199 	flags = TCP_SKB_CB(skb)->tcp_flags;
1200 	TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1201 	TCP_SKB_CB(buff)->tcp_flags = flags;
1202 	TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1203 
1204 	if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1205 		/* Copy and checksum data tail into the new buffer. */
1206 		buff->csum = csum_partial_copy_nocheck(skb->data + len,
1207 						       skb_put(buff, nsize),
1208 						       nsize, 0);
1209 
1210 		skb_trim(skb, len);
1211 
1212 		skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1213 	} else {
1214 		skb->ip_summed = CHECKSUM_PARTIAL;
1215 		skb_split(skb, buff, len);
1216 	}
1217 
1218 	buff->ip_summed = skb->ip_summed;
1219 
1220 	buff->tstamp = skb->tstamp;
1221 	tcp_fragment_tstamp(skb, buff);
1222 
1223 	old_factor = tcp_skb_pcount(skb);
1224 
1225 	/* Fix up tso_factor for both original and new SKB.  */
1226 	tcp_set_skb_tso_segs(skb, mss_now);
1227 	tcp_set_skb_tso_segs(buff, mss_now);
1228 
1229 	/* If this packet has been sent out already, we must
1230 	 * adjust the various packet counters.
1231 	 */
1232 	if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1233 		int diff = old_factor - tcp_skb_pcount(skb) -
1234 			tcp_skb_pcount(buff);
1235 
1236 		if (diff)
1237 			tcp_adjust_pcount(sk, skb, diff);
1238 	}
1239 
1240 	/* Link BUFF into the send queue. */
1241 	__skb_header_release(buff);
1242 	tcp_insert_write_queue_after(skb, buff, sk);
1243 
1244 	return 0;
1245 }
1246 
1247 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1248  * eventually). The difference is that pulled data not copied, but
1249  * immediately discarded.
1250  */
__pskb_trim_head(struct sk_buff * skb,int len)1251 static int __pskb_trim_head(struct sk_buff *skb, int len)
1252 {
1253 	struct skb_shared_info *shinfo;
1254 	int i, k, eat;
1255 
1256 	eat = min_t(int, len, skb_headlen(skb));
1257 	if (eat) {
1258 		__skb_pull(skb, eat);
1259 		len -= eat;
1260 		if (!len)
1261 			return 0;
1262 	}
1263 	eat = len;
1264 	k = 0;
1265 	shinfo = skb_shinfo(skb);
1266 	for (i = 0; i < shinfo->nr_frags; i++) {
1267 		int size = skb_frag_size(&shinfo->frags[i]);
1268 
1269 		if (size <= eat) {
1270 			skb_frag_unref(skb, i);
1271 			eat -= size;
1272 		} else {
1273 			shinfo->frags[k] = shinfo->frags[i];
1274 			if (eat) {
1275 				shinfo->frags[k].page_offset += eat;
1276 				skb_frag_size_sub(&shinfo->frags[k], eat);
1277 				eat = 0;
1278 			}
1279 			k++;
1280 		}
1281 	}
1282 	shinfo->nr_frags = k;
1283 
1284 	skb_reset_tail_pointer(skb);
1285 	skb->data_len -= len;
1286 	skb->len = skb->data_len;
1287 	return len;
1288 }
1289 
1290 /* Remove acked data from a packet in the transmit queue. */
tcp_trim_head(struct sock * sk,struct sk_buff * skb,u32 len)1291 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1292 {
1293 	u32 delta_truesize;
1294 
1295 	if (skb_unclone(skb, GFP_ATOMIC))
1296 		return -ENOMEM;
1297 
1298 	delta_truesize = __pskb_trim_head(skb, len);
1299 
1300 	TCP_SKB_CB(skb)->seq += len;
1301 	skb->ip_summed = CHECKSUM_PARTIAL;
1302 
1303 	if (delta_truesize) {
1304 		skb->truesize	   -= delta_truesize;
1305 		sk->sk_wmem_queued -= delta_truesize;
1306 		sk_mem_uncharge(sk, delta_truesize);
1307 		sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1308 	}
1309 
1310 	/* Any change of skb->len requires recalculation of tso factor. */
1311 	if (tcp_skb_pcount(skb) > 1)
1312 		tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1313 
1314 	return 0;
1315 }
1316 
1317 /* Calculate MSS not accounting any TCP options.  */
__tcp_mtu_to_mss(struct sock * sk,int pmtu)1318 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1319 {
1320 	const struct tcp_sock *tp = tcp_sk(sk);
1321 	const struct inet_connection_sock *icsk = inet_csk(sk);
1322 	int mss_now;
1323 
1324 	/* Calculate base mss without TCP options:
1325 	   It is MMS_S - sizeof(tcphdr) of rfc1122
1326 	 */
1327 	mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1328 
1329 	/* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1330 	if (icsk->icsk_af_ops->net_frag_header_len) {
1331 		const struct dst_entry *dst = __sk_dst_get(sk);
1332 
1333 		if (dst && dst_allfrag(dst))
1334 			mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1335 	}
1336 
1337 	/* Clamp it (mss_clamp does not include tcp options) */
1338 	if (mss_now > tp->rx_opt.mss_clamp)
1339 		mss_now = tp->rx_opt.mss_clamp;
1340 
1341 	/* Now subtract optional transport overhead */
1342 	mss_now -= icsk->icsk_ext_hdr_len;
1343 
1344 	/* Then reserve room for full set of TCP options and 8 bytes of data */
1345 	mss_now = max(mss_now, sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss);
1346 	return mss_now;
1347 }
1348 
1349 /* Calculate MSS. Not accounting for SACKs here.  */
tcp_mtu_to_mss(struct sock * sk,int pmtu)1350 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1351 {
1352 	/* Subtract TCP options size, not including SACKs */
1353 	return __tcp_mtu_to_mss(sk, pmtu) -
1354 	       (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1355 }
1356 EXPORT_SYMBOL(tcp_mtu_to_mss);
1357 
1358 /* Inverse of above */
tcp_mss_to_mtu(struct sock * sk,int mss)1359 int tcp_mss_to_mtu(struct sock *sk, int mss)
1360 {
1361 	const struct tcp_sock *tp = tcp_sk(sk);
1362 	const struct inet_connection_sock *icsk = inet_csk(sk);
1363 	int mtu;
1364 
1365 	mtu = mss +
1366 	      tp->tcp_header_len +
1367 	      icsk->icsk_ext_hdr_len +
1368 	      icsk->icsk_af_ops->net_header_len;
1369 
1370 	/* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1371 	if (icsk->icsk_af_ops->net_frag_header_len) {
1372 		const struct dst_entry *dst = __sk_dst_get(sk);
1373 
1374 		if (dst && dst_allfrag(dst))
1375 			mtu += icsk->icsk_af_ops->net_frag_header_len;
1376 	}
1377 	return mtu;
1378 }
1379 
1380 /* MTU probing init per socket */
tcp_mtup_init(struct sock * sk)1381 void tcp_mtup_init(struct sock *sk)
1382 {
1383 	struct tcp_sock *tp = tcp_sk(sk);
1384 	struct inet_connection_sock *icsk = inet_csk(sk);
1385 	struct net *net = sock_net(sk);
1386 
1387 	icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1388 	icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1389 			       icsk->icsk_af_ops->net_header_len;
1390 	icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1391 	icsk->icsk_mtup.probe_size = 0;
1392 	if (icsk->icsk_mtup.enabled)
1393 		icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1394 }
1395 EXPORT_SYMBOL(tcp_mtup_init);
1396 
1397 /* This function synchronize snd mss to current pmtu/exthdr set.
1398 
1399    tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1400    for TCP options, but includes only bare TCP header.
1401 
1402    tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1403    It is minimum of user_mss and mss received with SYN.
1404    It also does not include TCP options.
1405 
1406    inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1407 
1408    tp->mss_cache is current effective sending mss, including
1409    all tcp options except for SACKs. It is evaluated,
1410    taking into account current pmtu, but never exceeds
1411    tp->rx_opt.mss_clamp.
1412 
1413    NOTE1. rfc1122 clearly states that advertised MSS
1414    DOES NOT include either tcp or ip options.
1415 
1416    NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1417    are READ ONLY outside this function.		--ANK (980731)
1418  */
tcp_sync_mss(struct sock * sk,u32 pmtu)1419 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1420 {
1421 	struct tcp_sock *tp = tcp_sk(sk);
1422 	struct inet_connection_sock *icsk = inet_csk(sk);
1423 	int mss_now;
1424 
1425 	if (icsk->icsk_mtup.search_high > pmtu)
1426 		icsk->icsk_mtup.search_high = pmtu;
1427 
1428 	mss_now = tcp_mtu_to_mss(sk, pmtu);
1429 	mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1430 
1431 	/* And store cached results */
1432 	icsk->icsk_pmtu_cookie = pmtu;
1433 	if (icsk->icsk_mtup.enabled)
1434 		mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1435 	tp->mss_cache = mss_now;
1436 
1437 	return mss_now;
1438 }
1439 EXPORT_SYMBOL(tcp_sync_mss);
1440 
1441 /* Compute the current effective MSS, taking SACKs and IP options,
1442  * and even PMTU discovery events into account.
1443  */
tcp_current_mss(struct sock * sk)1444 unsigned int tcp_current_mss(struct sock *sk)
1445 {
1446 	const struct tcp_sock *tp = tcp_sk(sk);
1447 	const struct dst_entry *dst = __sk_dst_get(sk);
1448 	u32 mss_now;
1449 	unsigned int header_len;
1450 	struct tcp_out_options opts;
1451 	struct tcp_md5sig_key *md5;
1452 
1453 	mss_now = tp->mss_cache;
1454 
1455 	if (dst) {
1456 		u32 mtu = dst_mtu(dst);
1457 		if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1458 			mss_now = tcp_sync_mss(sk, mtu);
1459 	}
1460 
1461 	header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1462 		     sizeof(struct tcphdr);
1463 	/* The mss_cache is sized based on tp->tcp_header_len, which assumes
1464 	 * some common options. If this is an odd packet (because we have SACK
1465 	 * blocks etc) then our calculated header_len will be different, and
1466 	 * we have to adjust mss_now correspondingly */
1467 	if (header_len != tp->tcp_header_len) {
1468 		int delta = (int) header_len - tp->tcp_header_len;
1469 		mss_now -= delta;
1470 	}
1471 
1472 	return mss_now;
1473 }
1474 
1475 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1476  * As additional protections, we do not touch cwnd in retransmission phases,
1477  * and if application hit its sndbuf limit recently.
1478  */
tcp_cwnd_application_limited(struct sock * sk)1479 static void tcp_cwnd_application_limited(struct sock *sk)
1480 {
1481 	struct tcp_sock *tp = tcp_sk(sk);
1482 
1483 	if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1484 	    sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1485 		/* Limited by application or receiver window. */
1486 		u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1487 		u32 win_used = max(tp->snd_cwnd_used, init_win);
1488 		if (win_used < tp->snd_cwnd) {
1489 			tp->snd_ssthresh = tcp_current_ssthresh(sk);
1490 			tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1491 		}
1492 		tp->snd_cwnd_used = 0;
1493 	}
1494 	tp->snd_cwnd_stamp = tcp_time_stamp;
1495 }
1496 
tcp_cwnd_validate(struct sock * sk,bool is_cwnd_limited)1497 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1498 {
1499 	struct tcp_sock *tp = tcp_sk(sk);
1500 
1501 	/* Track the maximum number of outstanding packets in each
1502 	 * window, and remember whether we were cwnd-limited then.
1503 	 */
1504 	if (!before(tp->snd_una, tp->max_packets_seq) ||
1505 	    tp->packets_out > tp->max_packets_out ||
1506 	    is_cwnd_limited) {
1507 		tp->max_packets_out = tp->packets_out;
1508 		tp->max_packets_seq = tp->snd_nxt;
1509 		tp->is_cwnd_limited = is_cwnd_limited;
1510 	}
1511 
1512 	if (tcp_is_cwnd_limited(sk)) {
1513 		/* Network is feed fully. */
1514 		tp->snd_cwnd_used = 0;
1515 		tp->snd_cwnd_stamp = tcp_time_stamp;
1516 	} else {
1517 		/* Network starves. */
1518 		if (tp->packets_out > tp->snd_cwnd_used)
1519 			tp->snd_cwnd_used = tp->packets_out;
1520 
1521 		if (sysctl_tcp_slow_start_after_idle &&
1522 		    (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1523 			tcp_cwnd_application_limited(sk);
1524 	}
1525 }
1526 
1527 /* Minshall's variant of the Nagle send check. */
tcp_minshall_check(const struct tcp_sock * tp)1528 static bool tcp_minshall_check(const struct tcp_sock *tp)
1529 {
1530 	return after(tp->snd_sml, tp->snd_una) &&
1531 		!after(tp->snd_sml, tp->snd_nxt);
1532 }
1533 
1534 /* Update snd_sml if this skb is under mss
1535  * Note that a TSO packet might end with a sub-mss segment
1536  * The test is really :
1537  * if ((skb->len % mss) != 0)
1538  *        tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1539  * But we can avoid doing the divide again given we already have
1540  *  skb_pcount = skb->len / mss_now
1541  */
tcp_minshall_update(struct tcp_sock * tp,unsigned int mss_now,const struct sk_buff * skb)1542 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1543 				const struct sk_buff *skb)
1544 {
1545 	if (skb->len < tcp_skb_pcount(skb) * mss_now)
1546 		tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1547 }
1548 
1549 /* Return false, if packet can be sent now without violation Nagle's rules:
1550  * 1. It is full sized. (provided by caller in %partial bool)
1551  * 2. Or it contains FIN. (already checked by caller)
1552  * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1553  * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1554  *    With Minshall's modification: all sent small packets are ACKed.
1555  */
tcp_nagle_check(bool partial,const struct tcp_sock * tp,int nonagle)1556 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1557 			    int nonagle)
1558 {
1559 	return partial &&
1560 		((nonagle & TCP_NAGLE_CORK) ||
1561 		 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1562 }
1563 
1564 /* Return how many segs we'd like on a TSO packet,
1565  * to send one TSO packet per ms
1566  */
tcp_tso_autosize(const struct sock * sk,unsigned int mss_now)1567 static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now)
1568 {
1569 	u32 bytes, segs;
1570 
1571 	bytes = min(sk->sk_pacing_rate >> 10,
1572 		    sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1573 
1574 	/* Goal is to send at least one packet per ms,
1575 	 * not one big TSO packet every 100 ms.
1576 	 * This preserves ACK clocking and is consistent
1577 	 * with tcp_tso_should_defer() heuristic.
1578 	 */
1579 	segs = max_t(u32, bytes / mss_now, sysctl_tcp_min_tso_segs);
1580 
1581 	return min_t(u32, segs, sk->sk_gso_max_segs);
1582 }
1583 
1584 /* Returns the portion of skb which can be sent right away */
tcp_mss_split_point(const struct sock * sk,const struct sk_buff * skb,unsigned int mss_now,unsigned int max_segs,int nonagle)1585 static unsigned int tcp_mss_split_point(const struct sock *sk,
1586 					const struct sk_buff *skb,
1587 					unsigned int mss_now,
1588 					unsigned int max_segs,
1589 					int nonagle)
1590 {
1591 	const struct tcp_sock *tp = tcp_sk(sk);
1592 	u32 partial, needed, window, max_len;
1593 
1594 	window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1595 	max_len = mss_now * max_segs;
1596 
1597 	if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1598 		return max_len;
1599 
1600 	needed = min(skb->len, window);
1601 
1602 	if (max_len <= needed)
1603 		return max_len;
1604 
1605 	partial = needed % mss_now;
1606 	/* If last segment is not a full MSS, check if Nagle rules allow us
1607 	 * to include this last segment in this skb.
1608 	 * Otherwise, we'll split the skb at last MSS boundary
1609 	 */
1610 	if (tcp_nagle_check(partial != 0, tp, nonagle))
1611 		return needed - partial;
1612 
1613 	return needed;
1614 }
1615 
1616 /* Can at least one segment of SKB be sent right now, according to the
1617  * congestion window rules?  If so, return how many segments are allowed.
1618  */
tcp_cwnd_test(const struct tcp_sock * tp,const struct sk_buff * skb)1619 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1620 					 const struct sk_buff *skb)
1621 {
1622 	u32 in_flight, cwnd, halfcwnd;
1623 
1624 	/* Don't be strict about the congestion window for the final FIN.  */
1625 	if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1626 	    tcp_skb_pcount(skb) == 1)
1627 		return 1;
1628 
1629 	in_flight = tcp_packets_in_flight(tp);
1630 	cwnd = tp->snd_cwnd;
1631 	if (in_flight >= cwnd)
1632 		return 0;
1633 
1634 	/* For better scheduling, ensure we have at least
1635 	 * 2 GSO packets in flight.
1636 	 */
1637 	halfcwnd = max(cwnd >> 1, 1U);
1638 	return min(halfcwnd, cwnd - in_flight);
1639 }
1640 
1641 /* Initialize TSO state of a skb.
1642  * This must be invoked the first time we consider transmitting
1643  * SKB onto the wire.
1644  */
tcp_init_tso_segs(struct sk_buff * skb,unsigned int mss_now)1645 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1646 {
1647 	int tso_segs = tcp_skb_pcount(skb);
1648 
1649 	if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1650 		tcp_set_skb_tso_segs(skb, mss_now);
1651 		tso_segs = tcp_skb_pcount(skb);
1652 	}
1653 	return tso_segs;
1654 }
1655 
1656 
1657 /* Return true if the Nagle test allows this packet to be
1658  * sent now.
1659  */
tcp_nagle_test(const struct tcp_sock * tp,const struct sk_buff * skb,unsigned int cur_mss,int nonagle)1660 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1661 				  unsigned int cur_mss, int nonagle)
1662 {
1663 	/* Nagle rule does not apply to frames, which sit in the middle of the
1664 	 * write_queue (they have no chances to get new data).
1665 	 *
1666 	 * This is implemented in the callers, where they modify the 'nonagle'
1667 	 * argument based upon the location of SKB in the send queue.
1668 	 */
1669 	if (nonagle & TCP_NAGLE_PUSH)
1670 		return true;
1671 
1672 	/* Don't use the nagle rule for urgent data (or for the final FIN). */
1673 	if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1674 		return true;
1675 
1676 	if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1677 		return true;
1678 
1679 	return false;
1680 }
1681 
1682 /* 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)1683 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1684 			     const struct sk_buff *skb,
1685 			     unsigned int cur_mss)
1686 {
1687 	u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1688 
1689 	if (skb->len > cur_mss)
1690 		end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1691 
1692 	return !after(end_seq, tcp_wnd_end(tp));
1693 }
1694 
1695 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1696  * should be put on the wire right now.  If so, it returns the number of
1697  * packets allowed by the congestion window.
1698  */
tcp_snd_test(const struct sock * sk,struct sk_buff * skb,unsigned int cur_mss,int nonagle)1699 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1700 				 unsigned int cur_mss, int nonagle)
1701 {
1702 	const struct tcp_sock *tp = tcp_sk(sk);
1703 	unsigned int cwnd_quota;
1704 
1705 	tcp_init_tso_segs(skb, cur_mss);
1706 
1707 	if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1708 		return 0;
1709 
1710 	cwnd_quota = tcp_cwnd_test(tp, skb);
1711 	if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1712 		cwnd_quota = 0;
1713 
1714 	return cwnd_quota;
1715 }
1716 
1717 /* Test if sending is allowed right now. */
tcp_may_send_now(struct sock * sk)1718 bool tcp_may_send_now(struct sock *sk)
1719 {
1720 	const struct tcp_sock *tp = tcp_sk(sk);
1721 	struct sk_buff *skb = tcp_send_head(sk);
1722 
1723 	return skb &&
1724 		tcp_snd_test(sk, skb, tcp_current_mss(sk),
1725 			     (tcp_skb_is_last(sk, skb) ?
1726 			      tp->nonagle : TCP_NAGLE_PUSH));
1727 }
1728 
1729 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1730  * which is put after SKB on the list.  It is very much like
1731  * tcp_fragment() except that it may make several kinds of assumptions
1732  * in order to speed up the splitting operation.  In particular, we
1733  * know that all the data is in scatter-gather pages, and that the
1734  * packet has never been sent out before (and thus is not cloned).
1735  */
tso_fragment(struct sock * sk,struct sk_buff * skb,unsigned int len,unsigned int mss_now,gfp_t gfp)1736 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1737 			unsigned int mss_now, gfp_t gfp)
1738 {
1739 	struct sk_buff *buff;
1740 	int nlen = skb->len - len;
1741 	u8 flags;
1742 
1743 	/* All of a TSO frame must be composed of paged data.  */
1744 	if (skb->len != skb->data_len)
1745 		return tcp_fragment(sk, skb, len, mss_now, gfp);
1746 
1747 	buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1748 	if (unlikely(!buff))
1749 		return -ENOMEM;
1750 
1751 	sk->sk_wmem_queued += buff->truesize;
1752 	sk_mem_charge(sk, buff->truesize);
1753 	buff->truesize += nlen;
1754 	skb->truesize -= nlen;
1755 
1756 	/* Correct the sequence numbers. */
1757 	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1758 	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1759 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1760 
1761 	/* PSH and FIN should only be set in the second packet. */
1762 	flags = TCP_SKB_CB(skb)->tcp_flags;
1763 	TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1764 	TCP_SKB_CB(buff)->tcp_flags = flags;
1765 
1766 	/* This packet was never sent out yet, so no SACK bits. */
1767 	TCP_SKB_CB(buff)->sacked = 0;
1768 
1769 	buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1770 	skb_split(skb, buff, len);
1771 	tcp_fragment_tstamp(skb, buff);
1772 
1773 	/* Fix up tso_factor for both original and new SKB.  */
1774 	tcp_set_skb_tso_segs(skb, mss_now);
1775 	tcp_set_skb_tso_segs(buff, mss_now);
1776 
1777 	/* Link BUFF into the send queue. */
1778 	__skb_header_release(buff);
1779 	tcp_insert_write_queue_after(skb, buff, sk);
1780 
1781 	return 0;
1782 }
1783 
1784 /* Try to defer sending, if possible, in order to minimize the amount
1785  * of TSO splitting we do.  View it as a kind of TSO Nagle test.
1786  *
1787  * This algorithm is from John Heffner.
1788  */
tcp_tso_should_defer(struct sock * sk,struct sk_buff * skb,bool * is_cwnd_limited,u32 max_segs)1789 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1790 				 bool *is_cwnd_limited, u32 max_segs)
1791 {
1792 	const struct inet_connection_sock *icsk = inet_csk(sk);
1793 	u32 age, send_win, cong_win, limit, in_flight;
1794 	struct tcp_sock *tp = tcp_sk(sk);
1795 	struct skb_mstamp now;
1796 	struct sk_buff *head;
1797 	int win_divisor;
1798 
1799 	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1800 		goto send_now;
1801 
1802 	if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1803 		goto send_now;
1804 
1805 	/* Avoid bursty behavior by allowing defer
1806 	 * only if the last write was recent.
1807 	 */
1808 	if ((s32)(tcp_time_stamp - tp->lsndtime) > 0)
1809 		goto send_now;
1810 
1811 	in_flight = tcp_packets_in_flight(tp);
1812 
1813 	BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1814 
1815 	send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1816 
1817 	/* From in_flight test above, we know that cwnd > in_flight.  */
1818 	cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1819 
1820 	limit = min(send_win, cong_win);
1821 
1822 	/* If a full-sized TSO skb can be sent, do it. */
1823 	if (limit >= max_segs * tp->mss_cache)
1824 		goto send_now;
1825 
1826 	/* Middle in queue won't get any more data, full sendable already? */
1827 	if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1828 		goto send_now;
1829 
1830 	win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1831 	if (win_divisor) {
1832 		u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1833 
1834 		/* If at least some fraction of a window is available,
1835 		 * just use it.
1836 		 */
1837 		chunk /= win_divisor;
1838 		if (limit >= chunk)
1839 			goto send_now;
1840 	} else {
1841 		/* Different approach, try not to defer past a single
1842 		 * ACK.  Receiver should ACK every other full sized
1843 		 * frame, so if we have space for more than 3 frames
1844 		 * then send now.
1845 		 */
1846 		if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1847 			goto send_now;
1848 	}
1849 
1850 	head = tcp_write_queue_head(sk);
1851 	skb_mstamp_get(&now);
1852 	age = skb_mstamp_us_delta(&now, &head->skb_mstamp);
1853 	/* If next ACK is likely to come too late (half srtt), do not defer */
1854 	if (age < (tp->srtt_us >> 4))
1855 		goto send_now;
1856 
1857 	/* Ok, it looks like it is advisable to defer. */
1858 
1859 	if (cong_win < send_win && cong_win <= skb->len)
1860 		*is_cwnd_limited = true;
1861 
1862 	return true;
1863 
1864 send_now:
1865 	return false;
1866 }
1867 
tcp_mtu_check_reprobe(struct sock * sk)1868 static inline void tcp_mtu_check_reprobe(struct sock *sk)
1869 {
1870 	struct inet_connection_sock *icsk = inet_csk(sk);
1871 	struct tcp_sock *tp = tcp_sk(sk);
1872 	struct net *net = sock_net(sk);
1873 	u32 interval;
1874 	s32 delta;
1875 
1876 	interval = net->ipv4.sysctl_tcp_probe_interval;
1877 	delta = tcp_time_stamp - icsk->icsk_mtup.probe_timestamp;
1878 	if (unlikely(delta >= interval * HZ)) {
1879 		int mss = tcp_current_mss(sk);
1880 
1881 		/* Update current search range */
1882 		icsk->icsk_mtup.probe_size = 0;
1883 		icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
1884 			sizeof(struct tcphdr) +
1885 			icsk->icsk_af_ops->net_header_len;
1886 		icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
1887 
1888 		/* Update probe time stamp */
1889 		icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1890 	}
1891 }
1892 
1893 /* Create a new MTU probe if we are ready.
1894  * MTU probe is regularly attempting to increase the path MTU by
1895  * deliberately sending larger packets.  This discovers routing
1896  * changes resulting in larger path MTUs.
1897  *
1898  * Returns 0 if we should wait to probe (no cwnd available),
1899  *         1 if a probe was sent,
1900  *         -1 otherwise
1901  */
tcp_mtu_probe(struct sock * sk)1902 static int tcp_mtu_probe(struct sock *sk)
1903 {
1904 	struct tcp_sock *tp = tcp_sk(sk);
1905 	struct inet_connection_sock *icsk = inet_csk(sk);
1906 	struct sk_buff *skb, *nskb, *next;
1907 	struct net *net = sock_net(sk);
1908 	int len;
1909 	int probe_size;
1910 	int size_needed;
1911 	int copy;
1912 	int mss_now;
1913 	int interval;
1914 
1915 	/* Not currently probing/verifying,
1916 	 * not in recovery,
1917 	 * have enough cwnd, and
1918 	 * not SACKing (the variable headers throw things off) */
1919 	if (!icsk->icsk_mtup.enabled ||
1920 	    icsk->icsk_mtup.probe_size ||
1921 	    inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1922 	    tp->snd_cwnd < 11 ||
1923 	    tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1924 		return -1;
1925 
1926 	/* Use binary search for probe_size between tcp_mss_base,
1927 	 * and current mss_clamp. if (search_high - search_low)
1928 	 * smaller than a threshold, backoff from probing.
1929 	 */
1930 	mss_now = tcp_current_mss(sk);
1931 	probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
1932 				    icsk->icsk_mtup.search_low) >> 1);
1933 	size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1934 	interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
1935 	/* When misfortune happens, we are reprobing actively,
1936 	 * and then reprobe timer has expired. We stick with current
1937 	 * probing process by not resetting search range to its orignal.
1938 	 */
1939 	if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
1940 		interval < net->ipv4.sysctl_tcp_probe_threshold) {
1941 		/* Check whether enough time has elaplased for
1942 		 * another round of probing.
1943 		 */
1944 		tcp_mtu_check_reprobe(sk);
1945 		return -1;
1946 	}
1947 
1948 	/* Have enough data in the send queue to probe? */
1949 	if (tp->write_seq - tp->snd_nxt < size_needed)
1950 		return -1;
1951 
1952 	if (tp->snd_wnd < size_needed)
1953 		return -1;
1954 	if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1955 		return 0;
1956 
1957 	/* Do we need to wait to drain cwnd? With none in flight, don't stall */
1958 	if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1959 		if (!tcp_packets_in_flight(tp))
1960 			return -1;
1961 		else
1962 			return 0;
1963 	}
1964 
1965 	/* We're allowed to probe.  Build it now. */
1966 	nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
1967 	if (!nskb)
1968 		return -1;
1969 	sk->sk_wmem_queued += nskb->truesize;
1970 	sk_mem_charge(sk, nskb->truesize);
1971 
1972 	skb = tcp_send_head(sk);
1973 
1974 	TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1975 	TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1976 	TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1977 	TCP_SKB_CB(nskb)->sacked = 0;
1978 	nskb->csum = 0;
1979 	nskb->ip_summed = skb->ip_summed;
1980 
1981 	tcp_insert_write_queue_before(nskb, skb, sk);
1982 	tcp_highest_sack_replace(sk, skb, nskb);
1983 
1984 	len = 0;
1985 	tcp_for_write_queue_from_safe(skb, next, sk) {
1986 		copy = min_t(int, skb->len, probe_size - len);
1987 		if (nskb->ip_summed) {
1988 			skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1989 		} else {
1990 			__wsum csum = skb_copy_and_csum_bits(skb, 0,
1991 							     skb_put(nskb, copy),
1992 							     copy, 0);
1993 			nskb->csum = csum_block_add(nskb->csum, csum, len);
1994 		}
1995 
1996 		if (skb->len <= copy) {
1997 			/* We've eaten all the data from this skb.
1998 			 * Throw it away. */
1999 			TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2000 			tcp_unlink_write_queue(skb, sk);
2001 			sk_wmem_free_skb(sk, skb);
2002 		} else {
2003 			TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2004 						   ~(TCPHDR_FIN|TCPHDR_PSH);
2005 			if (!skb_shinfo(skb)->nr_frags) {
2006 				skb_pull(skb, copy);
2007 				if (skb->ip_summed != CHECKSUM_PARTIAL)
2008 					skb->csum = csum_partial(skb->data,
2009 								 skb->len, 0);
2010 			} else {
2011 				__pskb_trim_head(skb, copy);
2012 				tcp_set_skb_tso_segs(skb, mss_now);
2013 			}
2014 			TCP_SKB_CB(skb)->seq += copy;
2015 		}
2016 
2017 		len += copy;
2018 
2019 		if (len >= probe_size)
2020 			break;
2021 	}
2022 	tcp_init_tso_segs(nskb, nskb->len);
2023 
2024 	/* We're ready to send.  If this fails, the probe will
2025 	 * be resegmented into mss-sized pieces by tcp_write_xmit().
2026 	 */
2027 	if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2028 		/* Decrement cwnd here because we are sending
2029 		 * effectively two packets. */
2030 		tp->snd_cwnd--;
2031 		tcp_event_new_data_sent(sk, nskb);
2032 
2033 		icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2034 		tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2035 		tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2036 
2037 		return 1;
2038 	}
2039 
2040 	return -1;
2041 }
2042 
2043 /* This routine writes packets to the network.  It advances the
2044  * send_head.  This happens as incoming acks open up the remote
2045  * window for us.
2046  *
2047  * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2048  * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2049  * account rare use of URG, this is not a big flaw.
2050  *
2051  * Send at most one packet when push_one > 0. Temporarily ignore
2052  * cwnd limit to force at most one packet out when push_one == 2.
2053 
2054  * Returns true, if no segments are in flight and we have queued segments,
2055  * but cannot send anything now because of SWS or another problem.
2056  */
tcp_write_xmit(struct sock * sk,unsigned int mss_now,int nonagle,int push_one,gfp_t gfp)2057 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2058 			   int push_one, gfp_t gfp)
2059 {
2060 	struct tcp_sock *tp = tcp_sk(sk);
2061 	struct sk_buff *skb;
2062 	unsigned int tso_segs, sent_pkts;
2063 	int cwnd_quota;
2064 	int result;
2065 	bool is_cwnd_limited = false;
2066 	u32 max_segs;
2067 
2068 	sent_pkts = 0;
2069 
2070 	if (!push_one) {
2071 		/* Do MTU probing. */
2072 		result = tcp_mtu_probe(sk);
2073 		if (!result) {
2074 			return false;
2075 		} else if (result > 0) {
2076 			sent_pkts = 1;
2077 		}
2078 	}
2079 
2080 	max_segs = tcp_tso_autosize(sk, mss_now);
2081 	while ((skb = tcp_send_head(sk))) {
2082 		unsigned int limit;
2083 
2084 		tso_segs = tcp_init_tso_segs(skb, mss_now);
2085 		BUG_ON(!tso_segs);
2086 
2087 		if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2088 			/* "skb_mstamp" is used as a start point for the retransmit timer */
2089 			skb_mstamp_get(&skb->skb_mstamp);
2090 			goto repair; /* Skip network transmission */
2091 		}
2092 
2093 		cwnd_quota = tcp_cwnd_test(tp, skb);
2094 		if (!cwnd_quota) {
2095 			if (push_one == 2)
2096 				/* Force out a loss probe pkt. */
2097 				cwnd_quota = 1;
2098 			else
2099 				break;
2100 		}
2101 
2102 		if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
2103 			break;
2104 
2105 		if (tso_segs == 1) {
2106 			if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2107 						     (tcp_skb_is_last(sk, skb) ?
2108 						      nonagle : TCP_NAGLE_PUSH))))
2109 				break;
2110 		} else {
2111 			if (!push_one &&
2112 			    tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2113 						 max_segs))
2114 				break;
2115 		}
2116 
2117 		limit = mss_now;
2118 		if (tso_segs > 1 && !tcp_urg_mode(tp))
2119 			limit = tcp_mss_split_point(sk, skb, mss_now,
2120 						    min_t(unsigned int,
2121 							  cwnd_quota,
2122 							  max_segs),
2123 						    nonagle);
2124 
2125 		if (skb->len > limit &&
2126 		    unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2127 			break;
2128 
2129 		/* TCP Small Queues :
2130 		 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2131 		 * This allows for :
2132 		 *  - better RTT estimation and ACK scheduling
2133 		 *  - faster recovery
2134 		 *  - high rates
2135 		 * Alas, some drivers / subsystems require a fair amount
2136 		 * of queued bytes to ensure line rate.
2137 		 * One example is wifi aggregation (802.11 AMPDU)
2138 		 */
2139 		limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10);
2140 		limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes);
2141 
2142 		if (atomic_read(&sk->sk_wmem_alloc) > limit) {
2143 			set_bit(TSQ_THROTTLED, &tp->tsq_flags);
2144 			/* It is possible TX completion already happened
2145 			 * before we set TSQ_THROTTLED, so we must
2146 			 * test again the condition.
2147 			 */
2148 			smp_mb__after_atomic();
2149 			if (atomic_read(&sk->sk_wmem_alloc) > limit)
2150 				break;
2151 		}
2152 
2153 		/* Argh, we hit an empty skb(), presumably a thread
2154 		 * is sleeping in sendmsg()/sk_stream_wait_memory().
2155 		 * We do not want to send a pure-ack packet and have
2156 		 * a strange looking rtx queue with empty packet(s).
2157 		 */
2158 		if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq)
2159 			break;
2160 
2161 		if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2162 			break;
2163 
2164 repair:
2165 		/* Advance the send_head.  This one is sent out.
2166 		 * This call will increment packets_out.
2167 		 */
2168 		tcp_event_new_data_sent(sk, skb);
2169 
2170 		tcp_minshall_update(tp, mss_now, skb);
2171 		sent_pkts += tcp_skb_pcount(skb);
2172 
2173 		if (push_one)
2174 			break;
2175 	}
2176 
2177 	is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2178 	if (likely(sent_pkts || is_cwnd_limited))
2179 		tcp_cwnd_validate(sk, is_cwnd_limited);
2180 
2181 	if (likely(sent_pkts)) {
2182 		if (tcp_in_cwnd_reduction(sk))
2183 			tp->prr_out += sent_pkts;
2184 
2185 		/* Send one loss probe per tail loss episode. */
2186 		if (push_one != 2)
2187 			tcp_schedule_loss_probe(sk);
2188 		return false;
2189 	}
2190 	return !tp->packets_out && tcp_send_head(sk);
2191 }
2192 
tcp_schedule_loss_probe(struct sock * sk)2193 bool tcp_schedule_loss_probe(struct sock *sk)
2194 {
2195 	struct inet_connection_sock *icsk = inet_csk(sk);
2196 	struct tcp_sock *tp = tcp_sk(sk);
2197 	u32 timeout, tlp_time_stamp, rto_time_stamp;
2198 	u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3);
2199 
2200 	if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
2201 		return false;
2202 	/* No consecutive loss probes. */
2203 	if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
2204 		tcp_rearm_rto(sk);
2205 		return false;
2206 	}
2207 	/* Don't do any loss probe on a Fast Open connection before 3WHS
2208 	 * finishes.
2209 	 */
2210 	if (tp->fastopen_rsk)
2211 		return false;
2212 
2213 	/* TLP is only scheduled when next timer event is RTO. */
2214 	if (icsk->icsk_pending != ICSK_TIME_RETRANS)
2215 		return false;
2216 
2217 	/* Schedule a loss probe in 2*RTT for SACK capable connections
2218 	 * in Open state, that are either limited by cwnd or application.
2219 	 */
2220 	if (sysctl_tcp_early_retrans < 3 || !tp->packets_out ||
2221 	    !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2222 		return false;
2223 
2224 	if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2225 	     tcp_send_head(sk))
2226 		return false;
2227 
2228 	/* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2229 	 * for delayed ack when there's one outstanding packet. If no RTT
2230 	 * sample is available then probe after TCP_TIMEOUT_INIT.
2231 	 */
2232 	timeout = rtt << 1 ? : TCP_TIMEOUT_INIT;
2233 	if (tp->packets_out == 1)
2234 		timeout = max_t(u32, timeout,
2235 				(rtt + (rtt >> 1) + TCP_DELACK_MAX));
2236 	timeout = max_t(u32, timeout, msecs_to_jiffies(10));
2237 
2238 	/* If RTO is shorter, just schedule TLP in its place. */
2239 	tlp_time_stamp = tcp_time_stamp + timeout;
2240 	rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
2241 	if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
2242 		s32 delta = rto_time_stamp - tcp_time_stamp;
2243 		if (delta > 0)
2244 			timeout = delta;
2245 	}
2246 
2247 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2248 				  TCP_RTO_MAX);
2249 	return true;
2250 }
2251 
2252 /* Thanks to skb fast clones, we can detect if a prior transmit of
2253  * a packet is still in a qdisc or driver queue.
2254  * In this case, there is very little point doing a retransmit !
2255  * Note: This is called from BH context only.
2256  */
skb_still_in_host_queue(const struct sock * sk,const struct sk_buff * skb)2257 static bool skb_still_in_host_queue(const struct sock *sk,
2258 				    const struct sk_buff *skb)
2259 {
2260 	if (unlikely(skb_fclone_busy(sk, skb))) {
2261 		NET_INC_STATS_BH(sock_net(sk),
2262 				 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2263 		return true;
2264 	}
2265 	return false;
2266 }
2267 
2268 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2269  * retransmit the last segment.
2270  */
tcp_send_loss_probe(struct sock * sk)2271 void tcp_send_loss_probe(struct sock *sk)
2272 {
2273 	struct tcp_sock *tp = tcp_sk(sk);
2274 	struct sk_buff *skb;
2275 	int pcount;
2276 	int mss = tcp_current_mss(sk);
2277 
2278 	/* At most one outstanding TLP */
2279 	if (tp->tlp_high_seq)
2280 		goto rearm_timer;
2281 
2282 	tp->tlp_retrans = 0;
2283 	skb = tcp_send_head(sk);
2284 	if (skb) {
2285 		if (tcp_snd_wnd_test(tp, skb, mss)) {
2286 			pcount = tp->packets_out;
2287 			tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2288 			if (tp->packets_out > pcount)
2289 				goto probe_sent;
2290 			goto rearm_timer;
2291 		}
2292 		skb = tcp_write_queue_prev(sk, skb);
2293 	} else {
2294 		skb = tcp_write_queue_tail(sk);
2295 	}
2296 
2297 	if (unlikely(!skb)) {
2298 		WARN_ONCE(tp->packets_out,
2299 			  "invalid inflight: %u state %u cwnd %u mss %d\n",
2300 			  tp->packets_out, sk->sk_state, tp->snd_cwnd, mss);
2301 		inet_csk(sk)->icsk_pending = 0;
2302 		return;
2303 	}
2304 
2305 	if (skb_still_in_host_queue(sk, skb))
2306 		goto rearm_timer;
2307 
2308 	pcount = tcp_skb_pcount(skb);
2309 	if (WARN_ON(!pcount))
2310 		goto rearm_timer;
2311 
2312 	if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2313 		if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss,
2314 					  GFP_ATOMIC)))
2315 			goto rearm_timer;
2316 		skb = tcp_write_queue_next(sk, skb);
2317 	}
2318 
2319 	if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2320 		goto rearm_timer;
2321 
2322 	if (__tcp_retransmit_skb(sk, skb))
2323 		goto rearm_timer;
2324 
2325 	tp->tlp_retrans = 1;
2326 
2327 probe_sent:
2328 	/* Record snd_nxt for loss detection. */
2329 	tp->tlp_high_seq = tp->snd_nxt;
2330 
2331 	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2332 	/* Reset s.t. tcp_rearm_rto will restart timer from now */
2333 	inet_csk(sk)->icsk_pending = 0;
2334 rearm_timer:
2335 	tcp_rearm_rto(sk);
2336 }
2337 
2338 /* Push out any pending frames which were held back due to
2339  * TCP_CORK or attempt at coalescing tiny packets.
2340  * The socket must be locked by the caller.
2341  */
__tcp_push_pending_frames(struct sock * sk,unsigned int cur_mss,int nonagle)2342 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2343 			       int nonagle)
2344 {
2345 	/* If we are closed, the bytes will have to remain here.
2346 	 * In time closedown will finish, we empty the write queue and
2347 	 * all will be happy.
2348 	 */
2349 	if (unlikely(sk->sk_state == TCP_CLOSE))
2350 		return;
2351 
2352 	if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2353 			   sk_gfp_atomic(sk, GFP_ATOMIC)))
2354 		tcp_check_probe_timer(sk);
2355 }
2356 
2357 /* Send _single_ skb sitting at the send head. This function requires
2358  * true push pending frames to setup probe timer etc.
2359  */
tcp_push_one(struct sock * sk,unsigned int mss_now)2360 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2361 {
2362 	struct sk_buff *skb = tcp_send_head(sk);
2363 
2364 	BUG_ON(!skb || skb->len < mss_now);
2365 
2366 	tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2367 }
2368 
2369 /* This function returns the amount that we can raise the
2370  * usable window based on the following constraints
2371  *
2372  * 1. The window can never be shrunk once it is offered (RFC 793)
2373  * 2. We limit memory per socket
2374  *
2375  * RFC 1122:
2376  * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2377  *  RECV.NEXT + RCV.WIN fixed until:
2378  *  RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2379  *
2380  * i.e. don't raise the right edge of the window until you can raise
2381  * it at least MSS bytes.
2382  *
2383  * Unfortunately, the recommended algorithm breaks header prediction,
2384  * since header prediction assumes th->window stays fixed.
2385  *
2386  * Strictly speaking, keeping th->window fixed violates the receiver
2387  * side SWS prevention criteria. The problem is that under this rule
2388  * a stream of single byte packets will cause the right side of the
2389  * window to always advance by a single byte.
2390  *
2391  * Of course, if the sender implements sender side SWS prevention
2392  * then this will not be a problem.
2393  *
2394  * BSD seems to make the following compromise:
2395  *
2396  *	If the free space is less than the 1/4 of the maximum
2397  *	space available and the free space is less than 1/2 mss,
2398  *	then set the window to 0.
2399  *	[ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2400  *	Otherwise, just prevent the window from shrinking
2401  *	and from being larger than the largest representable value.
2402  *
2403  * This prevents incremental opening of the window in the regime
2404  * where TCP is limited by the speed of the reader side taking
2405  * data out of the TCP receive queue. It does nothing about
2406  * those cases where the window is constrained on the sender side
2407  * because the pipeline is full.
2408  *
2409  * BSD also seems to "accidentally" limit itself to windows that are a
2410  * multiple of MSS, at least until the free space gets quite small.
2411  * This would appear to be a side effect of the mbuf implementation.
2412  * Combining these two algorithms results in the observed behavior
2413  * of having a fixed window size at almost all times.
2414  *
2415  * Below we obtain similar behavior by forcing the offered window to
2416  * a multiple of the mss when it is feasible to do so.
2417  *
2418  * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2419  * Regular options like TIMESTAMP are taken into account.
2420  */
__tcp_select_window(struct sock * sk)2421 u32 __tcp_select_window(struct sock *sk)
2422 {
2423 	struct inet_connection_sock *icsk = inet_csk(sk);
2424 	struct tcp_sock *tp = tcp_sk(sk);
2425 	/* MSS for the peer's data.  Previous versions used mss_clamp
2426 	 * here.  I don't know if the value based on our guesses
2427 	 * of peer's MSS is better for the performance.  It's more correct
2428 	 * but may be worse for the performance because of rcv_mss
2429 	 * fluctuations.  --SAW  1998/11/1
2430 	 */
2431 	int mss = icsk->icsk_ack.rcv_mss;
2432 	int free_space = tcp_space(sk);
2433 	int allowed_space = tcp_full_space(sk);
2434 	int full_space = min_t(int, tp->window_clamp, allowed_space);
2435 	int window;
2436 
2437 	if (unlikely(mss > full_space)) {
2438 		mss = full_space;
2439 		if (mss <= 0)
2440 			return 0;
2441 	}
2442 	if (free_space < (full_space >> 1)) {
2443 		icsk->icsk_ack.quick = 0;
2444 
2445 		if (tcp_under_memory_pressure(sk))
2446 			tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2447 					       4U * tp->advmss);
2448 
2449 		/* free_space might become our new window, make sure we don't
2450 		 * increase it due to wscale.
2451 		 */
2452 		free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2453 
2454 		/* if free space is less than mss estimate, or is below 1/16th
2455 		 * of the maximum allowed, try to move to zero-window, else
2456 		 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2457 		 * new incoming data is dropped due to memory limits.
2458 		 * With large window, mss test triggers way too late in order
2459 		 * to announce zero window in time before rmem limit kicks in.
2460 		 */
2461 		if (free_space < (allowed_space >> 4) || free_space < mss)
2462 			return 0;
2463 	}
2464 
2465 	if (free_space > tp->rcv_ssthresh)
2466 		free_space = tp->rcv_ssthresh;
2467 
2468 	/* Don't do rounding if we are using window scaling, since the
2469 	 * scaled window will not line up with the MSS boundary anyway.
2470 	 */
2471 	window = tp->rcv_wnd;
2472 	if (tp->rx_opt.rcv_wscale) {
2473 		window = free_space;
2474 
2475 		/* Advertise enough space so that it won't get scaled away.
2476 		 * Import case: prevent zero window announcement if
2477 		 * 1<<rcv_wscale > mss.
2478 		 */
2479 		if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2480 			window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2481 				  << tp->rx_opt.rcv_wscale);
2482 	} else {
2483 		/* Get the largest window that is a nice multiple of mss.
2484 		 * Window clamp already applied above.
2485 		 * If our current window offering is within 1 mss of the
2486 		 * free space we just keep it. This prevents the divide
2487 		 * and multiply from happening most of the time.
2488 		 * We also don't do any window rounding when the free space
2489 		 * is too small.
2490 		 */
2491 		if (window <= free_space - mss || window > free_space)
2492 			window = (free_space / mss) * mss;
2493 		else if (mss == full_space &&
2494 			 free_space > window + (full_space >> 1))
2495 			window = free_space;
2496 	}
2497 
2498 	return window;
2499 }
2500 
2501 /* Collapses two adjacent SKB's during retransmission. */
tcp_collapse_retrans(struct sock * sk,struct sk_buff * skb)2502 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2503 {
2504 	struct tcp_sock *tp = tcp_sk(sk);
2505 	struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2506 	int skb_size, next_skb_size;
2507 
2508 	skb_size = skb->len;
2509 	next_skb_size = next_skb->len;
2510 
2511 	BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2512 
2513 	tcp_highest_sack_replace(sk, next_skb, skb);
2514 
2515 	tcp_unlink_write_queue(next_skb, sk);
2516 
2517 	skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2518 				  next_skb_size);
2519 
2520 	if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2521 		skb->ip_summed = CHECKSUM_PARTIAL;
2522 
2523 	if (skb->ip_summed != CHECKSUM_PARTIAL)
2524 		skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2525 
2526 	/* Update sequence range on original skb. */
2527 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2528 
2529 	/* Merge over control information. This moves PSH/FIN etc. over */
2530 	TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2531 
2532 	/* All done, get rid of second SKB and account for it so
2533 	 * packet counting does not break.
2534 	 */
2535 	TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2536 
2537 	/* changed transmit queue under us so clear hints */
2538 	tcp_clear_retrans_hints_partial(tp);
2539 	if (next_skb == tp->retransmit_skb_hint)
2540 		tp->retransmit_skb_hint = skb;
2541 
2542 	tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2543 
2544 	sk_wmem_free_skb(sk, next_skb);
2545 }
2546 
2547 /* Check if coalescing SKBs is legal. */
tcp_can_collapse(const struct sock * sk,const struct sk_buff * skb)2548 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2549 {
2550 	if (tcp_skb_pcount(skb) > 1)
2551 		return false;
2552 	/* TODO: SACK collapsing could be used to remove this condition */
2553 	if (skb_shinfo(skb)->nr_frags != 0)
2554 		return false;
2555 	if (skb_cloned(skb))
2556 		return false;
2557 	if (skb == tcp_send_head(sk))
2558 		return false;
2559 	/* Some heurestics for collapsing over SACK'd could be invented */
2560 	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2561 		return false;
2562 
2563 	return true;
2564 }
2565 
2566 /* Collapse packets in the retransmit queue to make to create
2567  * less packets on the wire. This is only done on retransmission.
2568  */
tcp_retrans_try_collapse(struct sock * sk,struct sk_buff * to,int space)2569 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2570 				     int space)
2571 {
2572 	struct tcp_sock *tp = tcp_sk(sk);
2573 	struct sk_buff *skb = to, *tmp;
2574 	bool first = true;
2575 
2576 	if (!sysctl_tcp_retrans_collapse)
2577 		return;
2578 	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2579 		return;
2580 
2581 	tcp_for_write_queue_from_safe(skb, tmp, sk) {
2582 		if (!tcp_can_collapse(sk, skb))
2583 			break;
2584 
2585 		space -= skb->len;
2586 
2587 		if (first) {
2588 			first = false;
2589 			continue;
2590 		}
2591 
2592 		if (space < 0)
2593 			break;
2594 		/* Punt if not enough space exists in the first SKB for
2595 		 * the data in the second
2596 		 */
2597 		if (skb->len > skb_availroom(to))
2598 			break;
2599 
2600 		if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2601 			break;
2602 
2603 		tcp_collapse_retrans(sk, to);
2604 	}
2605 }
2606 
2607 /* This retransmits one SKB.  Policy decisions and retransmit queue
2608  * state updates are done by the caller.  Returns non-zero if an
2609  * error occurred which prevented the send.
2610  */
__tcp_retransmit_skb(struct sock * sk,struct sk_buff * skb)2611 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2612 {
2613 	struct tcp_sock *tp = tcp_sk(sk);
2614 	struct inet_connection_sock *icsk = inet_csk(sk);
2615 	unsigned int cur_mss;
2616 	int err;
2617 
2618 	/* Inconslusive MTU probe */
2619 	if (icsk->icsk_mtup.probe_size) {
2620 		icsk->icsk_mtup.probe_size = 0;
2621 	}
2622 
2623 	/* Do not sent more than we queued. 1/4 is reserved for possible
2624 	 * copying overhead: fragmentation, tunneling, mangling etc.
2625 	 */
2626 	if (atomic_read(&sk->sk_wmem_alloc) >
2627 	    min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2628 		  sk->sk_sndbuf))
2629 		return -EAGAIN;
2630 
2631 	if (skb_still_in_host_queue(sk, skb))
2632 		return -EBUSY;
2633 
2634 	if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2635 		if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
2636 			WARN_ON_ONCE(1);
2637 			return -EINVAL;
2638 		}
2639 		if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2640 			return -ENOMEM;
2641 	}
2642 
2643 	if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2644 		return -EHOSTUNREACH; /* Routing failure or similar. */
2645 
2646 	cur_mss = tcp_current_mss(sk);
2647 
2648 	/* If receiver has shrunk his window, and skb is out of
2649 	 * new window, do not retransmit it. The exception is the
2650 	 * case, when window is shrunk to zero. In this case
2651 	 * our retransmit serves as a zero window probe.
2652 	 */
2653 	if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2654 	    TCP_SKB_CB(skb)->seq != tp->snd_una)
2655 		return -EAGAIN;
2656 
2657 	if (skb->len > cur_mss) {
2658 		if (tcp_fragment(sk, skb, cur_mss, cur_mss, GFP_ATOMIC))
2659 			return -ENOMEM; /* We'll try again later. */
2660 	} else {
2661 		int oldpcount = tcp_skb_pcount(skb);
2662 
2663 		if (unlikely(oldpcount > 1)) {
2664 			if (skb_unclone(skb, GFP_ATOMIC))
2665 				return -ENOMEM;
2666 			tcp_init_tso_segs(skb, cur_mss);
2667 			tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2668 		}
2669 	}
2670 
2671 	/* RFC3168, section 6.1.1.1. ECN fallback */
2672 	if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2673 		tcp_ecn_clear_syn(sk, skb);
2674 
2675 	tcp_retrans_try_collapse(sk, skb, cur_mss);
2676 
2677 	/* Make a copy, if the first transmission SKB clone we made
2678 	 * is still in somebody's hands, else make a clone.
2679 	 */
2680 
2681 	/* make sure skb->data is aligned on arches that require it
2682 	 * and check if ack-trimming & collapsing extended the headroom
2683 	 * beyond what csum_start can cover.
2684 	 */
2685 	if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2686 		     skb_headroom(skb) >= 0xFFFF)) {
2687 		struct sk_buff *nskb;
2688 
2689 		skb_mstamp_get(&skb->skb_mstamp);
2690 		nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
2691 		err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2692 			     -ENOBUFS;
2693 	} else {
2694 		err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2695 	}
2696 
2697 	if (likely(!err)) {
2698 		TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2699 		/* Update global TCP statistics. */
2700 		TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2701 		if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2702 			NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2703 		tp->total_retrans++;
2704 	}
2705 	return err;
2706 }
2707 
tcp_retransmit_skb(struct sock * sk,struct sk_buff * skb)2708 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2709 {
2710 	struct tcp_sock *tp = tcp_sk(sk);
2711 	int err = __tcp_retransmit_skb(sk, skb);
2712 
2713 	if (err == 0) {
2714 #if FASTRETRANS_DEBUG > 0
2715 		if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2716 			net_dbg_ratelimited("retrans_out leaked\n");
2717 		}
2718 #endif
2719 		TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2720 		tp->retrans_out += tcp_skb_pcount(skb);
2721 
2722 		/* Save stamp of the first retransmit. */
2723 		if (!tp->retrans_stamp)
2724 			tp->retrans_stamp = tcp_skb_timestamp(skb);
2725 
2726 	} else if (err != -EBUSY) {
2727 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2728 	}
2729 
2730 	if (tp->undo_retrans < 0)
2731 		tp->undo_retrans = 0;
2732 	tp->undo_retrans += tcp_skb_pcount(skb);
2733 	return err;
2734 }
2735 
2736 /* Check if we forward retransmits are possible in the current
2737  * window/congestion state.
2738  */
tcp_can_forward_retransmit(struct sock * sk)2739 static bool tcp_can_forward_retransmit(struct sock *sk)
2740 {
2741 	const struct inet_connection_sock *icsk = inet_csk(sk);
2742 	const struct tcp_sock *tp = tcp_sk(sk);
2743 
2744 	/* Forward retransmissions are possible only during Recovery. */
2745 	if (icsk->icsk_ca_state != TCP_CA_Recovery)
2746 		return false;
2747 
2748 	/* No forward retransmissions in Reno are possible. */
2749 	if (tcp_is_reno(tp))
2750 		return false;
2751 
2752 	/* Yeah, we have to make difficult choice between forward transmission
2753 	 * and retransmission... Both ways have their merits...
2754 	 *
2755 	 * For now we do not retransmit anything, while we have some new
2756 	 * segments to send. In the other cases, follow rule 3 for
2757 	 * NextSeg() specified in RFC3517.
2758 	 */
2759 
2760 	if (tcp_may_send_now(sk))
2761 		return false;
2762 
2763 	return true;
2764 }
2765 
2766 /* This gets called after a retransmit timeout, and the initially
2767  * retransmitted data is acknowledged.  It tries to continue
2768  * resending the rest of the retransmit queue, until either
2769  * we've sent it all or the congestion window limit is reached.
2770  * If doing SACK, the first ACK which comes back for a timeout
2771  * based retransmit packet might feed us FACK information again.
2772  * If so, we use it to avoid unnecessarily retransmissions.
2773  */
tcp_xmit_retransmit_queue(struct sock * sk)2774 void tcp_xmit_retransmit_queue(struct sock *sk)
2775 {
2776 	const struct inet_connection_sock *icsk = inet_csk(sk);
2777 	struct tcp_sock *tp = tcp_sk(sk);
2778 	struct sk_buff *skb;
2779 	struct sk_buff *hole = NULL;
2780 	u32 last_lost;
2781 	int mib_idx;
2782 	int fwd_rexmitting = 0;
2783 
2784 	if (!tp->packets_out)
2785 		return;
2786 
2787 	if (!tp->lost_out)
2788 		tp->retransmit_high = tp->snd_una;
2789 
2790 	if (tp->retransmit_skb_hint) {
2791 		skb = tp->retransmit_skb_hint;
2792 		last_lost = TCP_SKB_CB(skb)->end_seq;
2793 		if (after(last_lost, tp->retransmit_high))
2794 			last_lost = tp->retransmit_high;
2795 	} else {
2796 		skb = tcp_write_queue_head(sk);
2797 		last_lost = tp->snd_una;
2798 	}
2799 
2800 	tcp_for_write_queue_from(skb, sk) {
2801 		__u8 sacked = TCP_SKB_CB(skb)->sacked;
2802 
2803 		if (skb == tcp_send_head(sk))
2804 			break;
2805 		/* we could do better than to assign each time */
2806 		if (!hole)
2807 			tp->retransmit_skb_hint = skb;
2808 
2809 		/* Assume this retransmit will generate
2810 		 * only one packet for congestion window
2811 		 * calculation purposes.  This works because
2812 		 * tcp_retransmit_skb() will chop up the
2813 		 * packet to be MSS sized and all the
2814 		 * packet counting works out.
2815 		 */
2816 		if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2817 			return;
2818 
2819 		if (fwd_rexmitting) {
2820 begin_fwd:
2821 			if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2822 				break;
2823 			mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2824 
2825 		} else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2826 			tp->retransmit_high = last_lost;
2827 			if (!tcp_can_forward_retransmit(sk))
2828 				break;
2829 			/* Backtrack if necessary to non-L'ed skb */
2830 			if (hole) {
2831 				skb = hole;
2832 				hole = NULL;
2833 			}
2834 			fwd_rexmitting = 1;
2835 			goto begin_fwd;
2836 
2837 		} else if (!(sacked & TCPCB_LOST)) {
2838 			if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2839 				hole = skb;
2840 			continue;
2841 
2842 		} else {
2843 			last_lost = TCP_SKB_CB(skb)->end_seq;
2844 			if (icsk->icsk_ca_state != TCP_CA_Loss)
2845 				mib_idx = LINUX_MIB_TCPFASTRETRANS;
2846 			else
2847 				mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2848 		}
2849 
2850 		if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2851 			continue;
2852 
2853 		if (tcp_retransmit_skb(sk, skb))
2854 			return;
2855 
2856 		NET_INC_STATS_BH(sock_net(sk), mib_idx);
2857 
2858 		if (tcp_in_cwnd_reduction(sk))
2859 			tp->prr_out += tcp_skb_pcount(skb);
2860 
2861 		if (skb == tcp_write_queue_head(sk))
2862 			inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2863 						  inet_csk(sk)->icsk_rto,
2864 						  TCP_RTO_MAX);
2865 	}
2866 }
2867 
2868 /* We allow to exceed memory limits for FIN packets to expedite
2869  * connection tear down and (memory) recovery.
2870  * Otherwise tcp_send_fin() could be tempted to either delay FIN
2871  * or even be forced to close flow without any FIN.
2872  * In general, we want to allow one skb per socket to avoid hangs
2873  * with edge trigger epoll()
2874  */
sk_forced_mem_schedule(struct sock * sk,int size)2875 void sk_forced_mem_schedule(struct sock *sk, int size)
2876 {
2877 	int amt, status;
2878 
2879 	if (size <= sk->sk_forward_alloc)
2880 		return;
2881 	amt = sk_mem_pages(size);
2882 	sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
2883 	sk_memory_allocated_add(sk, amt, &status);
2884 }
2885 
2886 /* Send a FIN. The caller locks the socket for us.
2887  * We should try to send a FIN packet really hard, but eventually give up.
2888  */
tcp_send_fin(struct sock * sk)2889 void tcp_send_fin(struct sock *sk)
2890 {
2891 	struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
2892 	struct tcp_sock *tp = tcp_sk(sk);
2893 
2894 	/* Optimization, tack on the FIN if we have one skb in write queue and
2895 	 * this skb was not yet sent, or we are under memory pressure.
2896 	 * Note: in the latter case, FIN packet will be sent after a timeout,
2897 	 * as TCP stack thinks it has already been transmitted.
2898 	 */
2899 	if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) {
2900 coalesce:
2901 		TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
2902 		TCP_SKB_CB(tskb)->end_seq++;
2903 		tp->write_seq++;
2904 		if (!tcp_send_head(sk)) {
2905 			/* This means tskb was already sent.
2906 			 * Pretend we included the FIN on previous transmit.
2907 			 * We need to set tp->snd_nxt to the value it would have
2908 			 * if FIN had been sent. This is because retransmit path
2909 			 * does not change tp->snd_nxt.
2910 			 */
2911 			tp->snd_nxt++;
2912 			return;
2913 		}
2914 	} else {
2915 		skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
2916 		if (unlikely(!skb)) {
2917 			if (tskb)
2918 				goto coalesce;
2919 			return;
2920 		}
2921 		skb_reserve(skb, MAX_TCP_HEADER);
2922 		sk_forced_mem_schedule(sk, skb->truesize);
2923 		/* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2924 		tcp_init_nondata_skb(skb, tp->write_seq,
2925 				     TCPHDR_ACK | TCPHDR_FIN);
2926 		tcp_queue_skb(sk, skb);
2927 	}
2928 	__tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
2929 }
2930 
2931 /* We get here when a process closes a file descriptor (either due to
2932  * an explicit close() or as a byproduct of exit()'ing) and there
2933  * was unread data in the receive queue.  This behavior is recommended
2934  * by RFC 2525, section 2.17.  -DaveM
2935  */
tcp_send_active_reset(struct sock * sk,gfp_t priority)2936 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2937 {
2938 	struct sk_buff *skb;
2939 
2940 	/* NOTE: No TCP options attached and we never retransmit this. */
2941 	skb = alloc_skb(MAX_TCP_HEADER, priority);
2942 	if (!skb) {
2943 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2944 		return;
2945 	}
2946 
2947 	/* Reserve space for headers and prepare control bits. */
2948 	skb_reserve(skb, MAX_TCP_HEADER);
2949 	tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2950 			     TCPHDR_ACK | TCPHDR_RST);
2951 	skb_mstamp_get(&skb->skb_mstamp);
2952 	/* Send it off. */
2953 	if (tcp_transmit_skb(sk, skb, 0, priority))
2954 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2955 
2956 	TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2957 }
2958 
2959 /* Send a crossed SYN-ACK during socket establishment.
2960  * WARNING: This routine must only be called when we have already sent
2961  * a SYN packet that crossed the incoming SYN that caused this routine
2962  * to get called. If this assumption fails then the initial rcv_wnd
2963  * and rcv_wscale values will not be correct.
2964  */
tcp_send_synack(struct sock * sk)2965 int tcp_send_synack(struct sock *sk)
2966 {
2967 	struct sk_buff *skb;
2968 
2969 	skb = tcp_write_queue_head(sk);
2970 	if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2971 		pr_debug("%s: wrong queue state\n", __func__);
2972 		return -EFAULT;
2973 	}
2974 	if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2975 		if (skb_cloned(skb)) {
2976 			struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2977 			if (!nskb)
2978 				return -ENOMEM;
2979 			tcp_unlink_write_queue(skb, sk);
2980 			__skb_header_release(nskb);
2981 			__tcp_add_write_queue_head(sk, nskb);
2982 			sk_wmem_free_skb(sk, skb);
2983 			sk->sk_wmem_queued += nskb->truesize;
2984 			sk_mem_charge(sk, nskb->truesize);
2985 			skb = nskb;
2986 		}
2987 
2988 		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2989 		tcp_ecn_send_synack(sk, skb);
2990 	}
2991 	return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2992 }
2993 
2994 /**
2995  * tcp_make_synack - Prepare a SYN-ACK.
2996  * sk: listener socket
2997  * dst: dst entry attached to the SYNACK
2998  * req: request_sock pointer
2999  *
3000  * Allocate one skb and build a SYNACK packet.
3001  * @dst is consumed : Caller should not use it again.
3002  */
tcp_make_synack(const struct sock * sk,struct dst_entry * dst,struct request_sock * req,struct tcp_fastopen_cookie * foc,bool attach_req)3003 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3004 				struct request_sock *req,
3005 				struct tcp_fastopen_cookie *foc,
3006 				bool attach_req)
3007 {
3008 	struct inet_request_sock *ireq = inet_rsk(req);
3009 	const struct tcp_sock *tp = tcp_sk(sk);
3010 	struct tcp_md5sig_key *md5 = NULL;
3011 	struct tcp_out_options opts;
3012 	struct sk_buff *skb;
3013 	int tcp_header_size;
3014 	struct tcphdr *th;
3015 	u16 user_mss;
3016 	int mss;
3017 
3018 	skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3019 	if (unlikely(!skb)) {
3020 		dst_release(dst);
3021 		return NULL;
3022 	}
3023 	/* Reserve space for headers. */
3024 	skb_reserve(skb, MAX_TCP_HEADER);
3025 
3026 	if (attach_req) {
3027 		skb_set_owner_w(skb, req_to_sk(req));
3028 	} else {
3029 		/* sk is a const pointer, because we want to express multiple
3030 		 * cpu might call us concurrently.
3031 		 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3032 		 */
3033 		skb_set_owner_w(skb, (struct sock *)sk);
3034 	}
3035 	skb_dst_set(skb, dst);
3036 
3037 	mss = dst_metric_advmss(dst);
3038 	user_mss = READ_ONCE(tp->rx_opt.user_mss);
3039 	if (user_mss && user_mss < mss)
3040 		mss = user_mss;
3041 
3042 	memset(&opts, 0, sizeof(opts));
3043 #ifdef CONFIG_SYN_COOKIES
3044 	if (unlikely(req->cookie_ts))
3045 		skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req);
3046 	else
3047 #endif
3048 	skb_mstamp_get(&skb->skb_mstamp);
3049 
3050 #ifdef CONFIG_TCP_MD5SIG
3051 	rcu_read_lock();
3052 	md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3053 #endif
3054 	skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3055 	tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5, foc) +
3056 			  sizeof(*th);
3057 
3058 	skb_push(skb, tcp_header_size);
3059 	skb_reset_transport_header(skb);
3060 
3061 	th = tcp_hdr(skb);
3062 	memset(th, 0, sizeof(struct tcphdr));
3063 	th->syn = 1;
3064 	th->ack = 1;
3065 	tcp_ecn_make_synack(req, th);
3066 	th->source = htons(ireq->ir_num);
3067 	th->dest = ireq->ir_rmt_port;
3068 	skb->ip_summed = CHECKSUM_PARTIAL;
3069 	th->seq = htonl(tcp_rsk(req)->snt_isn);
3070 	/* XXX data is queued and acked as is. No buffer/window check */
3071 	th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3072 
3073 	/* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3074 	th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3075 	tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3076 	th->doff = (tcp_header_size >> 2);
3077 	TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_OUTSEGS);
3078 
3079 #ifdef CONFIG_TCP_MD5SIG
3080 	/* Okay, we have all we need - do the md5 hash if needed */
3081 	if (md5)
3082 		tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3083 					       md5, req_to_sk(req), skb);
3084 	rcu_read_unlock();
3085 #endif
3086 
3087 	/* Do not fool tcpdump (if any), clean our debris */
3088 	skb->tstamp.tv64 = 0;
3089 	return skb;
3090 }
3091 EXPORT_SYMBOL(tcp_make_synack);
3092 
tcp_ca_dst_init(struct sock * sk,const struct dst_entry * dst)3093 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3094 {
3095 	struct inet_connection_sock *icsk = inet_csk(sk);
3096 	const struct tcp_congestion_ops *ca;
3097 	u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3098 
3099 	if (ca_key == TCP_CA_UNSPEC)
3100 		return;
3101 
3102 	rcu_read_lock();
3103 	ca = tcp_ca_find_key(ca_key);
3104 	if (likely(ca && try_module_get(ca->owner))) {
3105 		module_put(icsk->icsk_ca_ops->owner);
3106 		icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3107 		icsk->icsk_ca_ops = ca;
3108 	}
3109 	rcu_read_unlock();
3110 }
3111 
3112 /* Do all connect socket setups that can be done AF independent. */
tcp_connect_init(struct sock * sk)3113 static void tcp_connect_init(struct sock *sk)
3114 {
3115 	const struct dst_entry *dst = __sk_dst_get(sk);
3116 	struct tcp_sock *tp = tcp_sk(sk);
3117 	__u8 rcv_wscale;
3118 
3119 	/* We'll fix this up when we get a response from the other end.
3120 	 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3121 	 */
3122 	tp->tcp_header_len = sizeof(struct tcphdr) +
3123 		(sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
3124 
3125 #ifdef CONFIG_TCP_MD5SIG
3126 	if (tp->af_specific->md5_lookup(sk, sk))
3127 		tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3128 #endif
3129 
3130 	/* If user gave his TCP_MAXSEG, record it to clamp */
3131 	if (tp->rx_opt.user_mss)
3132 		tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3133 	tp->max_window = 0;
3134 	tcp_mtup_init(sk);
3135 	tcp_sync_mss(sk, dst_mtu(dst));
3136 
3137 	tcp_ca_dst_init(sk, dst);
3138 
3139 	if (!tp->window_clamp)
3140 		tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3141 	tp->advmss = dst_metric_advmss(dst);
3142 	if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
3143 		tp->advmss = tp->rx_opt.user_mss;
3144 
3145 	tcp_initialize_rcv_mss(sk);
3146 
3147 	/* limit the window selection if the user enforce a smaller rx buffer */
3148 	if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3149 	    (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3150 		tp->window_clamp = tcp_full_space(sk);
3151 
3152 	tcp_select_initial_window(tcp_full_space(sk),
3153 				  tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3154 				  &tp->rcv_wnd,
3155 				  &tp->window_clamp,
3156 				  sysctl_tcp_window_scaling,
3157 				  &rcv_wscale,
3158 				  dst_metric(dst, RTAX_INITRWND));
3159 
3160 	tp->rx_opt.rcv_wscale = rcv_wscale;
3161 	tp->rcv_ssthresh = tp->rcv_wnd;
3162 
3163 	sk->sk_err = 0;
3164 	sock_reset_flag(sk, SOCK_DONE);
3165 	tp->snd_wnd = 0;
3166 	tcp_init_wl(tp, 0);
3167 	tcp_write_queue_purge(sk);
3168 	tp->snd_una = tp->write_seq;
3169 	tp->snd_sml = tp->write_seq;
3170 	tp->snd_up = tp->write_seq;
3171 	tp->snd_nxt = tp->write_seq;
3172 
3173 	if (likely(!tp->repair))
3174 		tp->rcv_nxt = 0;
3175 	else
3176 		tp->rcv_tstamp = tcp_time_stamp;
3177 	tp->rcv_wup = tp->rcv_nxt;
3178 	tp->copied_seq = tp->rcv_nxt;
3179 
3180 	inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
3181 	inet_csk(sk)->icsk_retransmits = 0;
3182 	tcp_clear_retrans(tp);
3183 }
3184 
tcp_connect_queue_skb(struct sock * sk,struct sk_buff * skb)3185 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3186 {
3187 	struct tcp_sock *tp = tcp_sk(sk);
3188 	struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3189 
3190 	tcb->end_seq += skb->len;
3191 	__skb_header_release(skb);
3192 	__tcp_add_write_queue_tail(sk, skb);
3193 	sk->sk_wmem_queued += skb->truesize;
3194 	sk_mem_charge(sk, skb->truesize);
3195 	tp->write_seq = tcb->end_seq;
3196 	tp->packets_out += tcp_skb_pcount(skb);
3197 }
3198 
3199 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3200  * queue a data-only packet after the regular SYN, such that regular SYNs
3201  * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3202  * only the SYN sequence, the data are retransmitted in the first ACK.
3203  * If cookie is not cached or other error occurs, falls back to send a
3204  * regular SYN with Fast Open cookie request option.
3205  */
tcp_send_syn_data(struct sock * sk,struct sk_buff * syn)3206 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3207 {
3208 	struct tcp_sock *tp = tcp_sk(sk);
3209 	struct tcp_fastopen_request *fo = tp->fastopen_req;
3210 	int syn_loss = 0, space, err = 0;
3211 	unsigned long last_syn_loss = 0;
3212 	struct sk_buff *syn_data;
3213 
3214 	tp->rx_opt.mss_clamp = tp->advmss;  /* If MSS is not cached */
3215 	tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
3216 			       &syn_loss, &last_syn_loss);
3217 	/* Recurring FO SYN losses: revert to regular handshake temporarily */
3218 	if (syn_loss > 1 &&
3219 	    time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
3220 		fo->cookie.len = -1;
3221 		goto fallback;
3222 	}
3223 
3224 	if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
3225 		fo->cookie.len = -1;
3226 	else if (fo->cookie.len <= 0)
3227 		goto fallback;
3228 
3229 	/* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3230 	 * user-MSS. Reserve maximum option space for middleboxes that add
3231 	 * private TCP options. The cost is reduced data space in SYN :(
3232 	 */
3233 	if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
3234 		tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3235 	space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3236 		MAX_TCP_OPTION_SPACE;
3237 
3238 	space = min_t(size_t, space, fo->size);
3239 
3240 	/* limit to order-0 allocations */
3241 	space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3242 
3243 	syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3244 	if (!syn_data)
3245 		goto fallback;
3246 	syn_data->ip_summed = CHECKSUM_PARTIAL;
3247 	memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3248 	if (space) {
3249 		int copied = copy_from_iter(skb_put(syn_data, space), space,
3250 					    &fo->data->msg_iter);
3251 		if (unlikely(!copied)) {
3252 			kfree_skb(syn_data);
3253 			goto fallback;
3254 		}
3255 		if (copied != space) {
3256 			skb_trim(syn_data, copied);
3257 			space = copied;
3258 		}
3259 	}
3260 	/* No more data pending in inet_wait_for_connect() */
3261 	if (space == fo->size)
3262 		fo->data = NULL;
3263 	fo->copied = space;
3264 
3265 	tcp_connect_queue_skb(sk, syn_data);
3266 
3267 	err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3268 
3269 	syn->skb_mstamp = syn_data->skb_mstamp;
3270 
3271 	/* Now full SYN+DATA was cloned and sent (or not),
3272 	 * remove the SYN from the original skb (syn_data)
3273 	 * we keep in write queue in case of a retransmit, as we
3274 	 * also have the SYN packet (with no data) in the same queue.
3275 	 */
3276 	TCP_SKB_CB(syn_data)->seq++;
3277 	TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3278 	if (!err) {
3279 		tp->syn_data = (fo->copied > 0);
3280 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3281 		goto done;
3282 	}
3283 
3284 fallback:
3285 	/* Send a regular SYN with Fast Open cookie request option */
3286 	if (fo->cookie.len > 0)
3287 		fo->cookie.len = 0;
3288 	err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3289 	if (err)
3290 		tp->syn_fastopen = 0;
3291 done:
3292 	fo->cookie.len = -1;  /* Exclude Fast Open option for SYN retries */
3293 	return err;
3294 }
3295 
3296 /* Build a SYN and send it off. */
tcp_connect(struct sock * sk)3297 int tcp_connect(struct sock *sk)
3298 {
3299 	struct tcp_sock *tp = tcp_sk(sk);
3300 	struct sk_buff *buff;
3301 	int err;
3302 
3303 	if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3304 		return -EHOSTUNREACH; /* Routing failure or similar. */
3305 
3306 	tcp_connect_init(sk);
3307 
3308 	if (unlikely(tp->repair)) {
3309 		tcp_finish_connect(sk, NULL);
3310 		return 0;
3311 	}
3312 
3313 	buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3314 	if (unlikely(!buff))
3315 		return -ENOBUFS;
3316 
3317 	tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3318 	tp->retrans_stamp = tcp_time_stamp;
3319 	tcp_connect_queue_skb(sk, buff);
3320 	tcp_ecn_send_syn(sk, buff);
3321 
3322 	/* Send off SYN; include data in Fast Open. */
3323 	err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3324 	      tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3325 	if (err == -ECONNREFUSED)
3326 		return err;
3327 
3328 	/* We change tp->snd_nxt after the tcp_transmit_skb() call
3329 	 * in order to make this packet get counted in tcpOutSegs.
3330 	 */
3331 	tp->snd_nxt = tp->write_seq;
3332 	tp->pushed_seq = tp->write_seq;
3333 	TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3334 
3335 	/* Timer for repeating the SYN until an answer. */
3336 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3337 				  inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3338 	return 0;
3339 }
3340 EXPORT_SYMBOL(tcp_connect);
3341 
3342 /* Send out a delayed ack, the caller does the policy checking
3343  * to see if we should even be here.  See tcp_input.c:tcp_ack_snd_check()
3344  * for details.
3345  */
tcp_send_delayed_ack(struct sock * sk)3346 void tcp_send_delayed_ack(struct sock *sk)
3347 {
3348 	struct inet_connection_sock *icsk = inet_csk(sk);
3349 	int ato = icsk->icsk_ack.ato;
3350 	unsigned long timeout;
3351 
3352 	if (ato > TCP_DELACK_MIN) {
3353 		const struct tcp_sock *tp = tcp_sk(sk);
3354 		int max_ato = HZ / 2;
3355 
3356 		if (icsk->icsk_ack.pingpong ||
3357 		    (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3358 			max_ato = TCP_DELACK_MAX;
3359 
3360 		/* Slow path, intersegment interval is "high". */
3361 
3362 		/* If some rtt estimate is known, use it to bound delayed ack.
3363 		 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3364 		 * directly.
3365 		 */
3366 		if (tp->srtt_us) {
3367 			int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3368 					TCP_DELACK_MIN);
3369 
3370 			if (rtt < max_ato)
3371 				max_ato = rtt;
3372 		}
3373 
3374 		ato = min(ato, max_ato);
3375 	}
3376 
3377 	/* Stay within the limit we were given */
3378 	timeout = jiffies + ato;
3379 
3380 	/* Use new timeout only if there wasn't a older one earlier. */
3381 	if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3382 		/* If delack timer was blocked or is about to expire,
3383 		 * send ACK now.
3384 		 */
3385 		if (icsk->icsk_ack.blocked ||
3386 		    time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3387 			tcp_send_ack(sk);
3388 			return;
3389 		}
3390 
3391 		if (!time_before(timeout, icsk->icsk_ack.timeout))
3392 			timeout = icsk->icsk_ack.timeout;
3393 	}
3394 	icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3395 	icsk->icsk_ack.timeout = timeout;
3396 	sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3397 }
3398 
3399 /* This routine sends an ack and also updates the window. */
__tcp_send_ack(struct sock * sk,u32 rcv_nxt)3400 void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
3401 {
3402 	struct sk_buff *buff;
3403 
3404 	/* If we have been reset, we may not send again. */
3405 	if (sk->sk_state == TCP_CLOSE)
3406 		return;
3407 
3408 	/* We are not putting this on the write queue, so
3409 	 * tcp_transmit_skb() will set the ownership to this
3410 	 * sock.
3411 	 */
3412 	buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3413 	if (!buff) {
3414 		inet_csk_schedule_ack(sk);
3415 		inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3416 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3417 					  TCP_DELACK_MAX, TCP_RTO_MAX);
3418 		return;
3419 	}
3420 
3421 	/* Reserve space for headers and prepare control bits. */
3422 	skb_reserve(buff, MAX_TCP_HEADER);
3423 	tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3424 
3425 	/* We do not want pure acks influencing TCP Small Queues or fq/pacing
3426 	 * too much.
3427 	 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3428 	 * We also avoid tcp_wfree() overhead (cache line miss accessing
3429 	 * tp->tsq_flags) by using regular sock_wfree()
3430 	 */
3431 	skb_set_tcp_pure_ack(buff);
3432 
3433 	/* Send it off, this clears delayed acks for us. */
3434 	skb_mstamp_get(&buff->skb_mstamp);
3435 	__tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC), rcv_nxt);
3436 }
3437 EXPORT_SYMBOL_GPL(__tcp_send_ack);
3438 
tcp_send_ack(struct sock * sk)3439 void tcp_send_ack(struct sock *sk)
3440 {
3441 	__tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
3442 }
3443 
3444 /* This routine sends a packet with an out of date sequence
3445  * number. It assumes the other end will try to ack it.
3446  *
3447  * Question: what should we make while urgent mode?
3448  * 4.4BSD forces sending single byte of data. We cannot send
3449  * out of window data, because we have SND.NXT==SND.MAX...
3450  *
3451  * Current solution: to send TWO zero-length segments in urgent mode:
3452  * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3453  * out-of-date with SND.UNA-1 to probe window.
3454  */
tcp_xmit_probe_skb(struct sock * sk,int urgent,int mib)3455 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3456 {
3457 	struct tcp_sock *tp = tcp_sk(sk);
3458 	struct sk_buff *skb;
3459 
3460 	/* We don't queue it, tcp_transmit_skb() sets ownership. */
3461 	skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3462 	if (!skb)
3463 		return -1;
3464 
3465 	/* Reserve space for headers and set control bits. */
3466 	skb_reserve(skb, MAX_TCP_HEADER);
3467 	/* Use a previous sequence.  This should cause the other
3468 	 * end to send an ack.  Don't queue or clone SKB, just
3469 	 * send it.
3470 	 */
3471 	tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3472 	skb_mstamp_get(&skb->skb_mstamp);
3473 	NET_INC_STATS(sock_net(sk), mib);
3474 	return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
3475 }
3476 
tcp_send_window_probe(struct sock * sk)3477 void tcp_send_window_probe(struct sock *sk)
3478 {
3479 	if (sk->sk_state == TCP_ESTABLISHED) {
3480 		tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3481 		tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3482 	}
3483 }
3484 
3485 /* Initiate keepalive or window probe from timer. */
tcp_write_wakeup(struct sock * sk,int mib)3486 int tcp_write_wakeup(struct sock *sk, int mib)
3487 {
3488 	struct tcp_sock *tp = tcp_sk(sk);
3489 	struct sk_buff *skb;
3490 
3491 	if (sk->sk_state == TCP_CLOSE)
3492 		return -1;
3493 
3494 	skb = tcp_send_head(sk);
3495 	if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3496 		int err;
3497 		unsigned int mss = tcp_current_mss(sk);
3498 		unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3499 
3500 		if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3501 			tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3502 
3503 		/* We are probing the opening of a window
3504 		 * but the window size is != 0
3505 		 * must have been a result SWS avoidance ( sender )
3506 		 */
3507 		if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3508 		    skb->len > mss) {
3509 			seg_size = min(seg_size, mss);
3510 			TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3511 			if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC))
3512 				return -1;
3513 		} else if (!tcp_skb_pcount(skb))
3514 			tcp_set_skb_tso_segs(skb, mss);
3515 
3516 		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3517 		err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3518 		if (!err)
3519 			tcp_event_new_data_sent(sk, skb);
3520 		return err;
3521 	} else {
3522 		if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3523 			tcp_xmit_probe_skb(sk, 1, mib);
3524 		return tcp_xmit_probe_skb(sk, 0, mib);
3525 	}
3526 }
3527 
3528 /* A window probe timeout has occurred.  If window is not closed send
3529  * a partial packet else a zero probe.
3530  */
tcp_send_probe0(struct sock * sk)3531 void tcp_send_probe0(struct sock *sk)
3532 {
3533 	struct inet_connection_sock *icsk = inet_csk(sk);
3534 	struct tcp_sock *tp = tcp_sk(sk);
3535 	unsigned long probe_max;
3536 	int err;
3537 
3538 	err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3539 
3540 	if (tp->packets_out || !tcp_send_head(sk)) {
3541 		/* Cancel probe timer, if it is not required. */
3542 		icsk->icsk_probes_out = 0;
3543 		icsk->icsk_backoff = 0;
3544 		return;
3545 	}
3546 
3547 	if (err <= 0) {
3548 		if (icsk->icsk_backoff < sysctl_tcp_retries2)
3549 			icsk->icsk_backoff++;
3550 		icsk->icsk_probes_out++;
3551 		probe_max = TCP_RTO_MAX;
3552 	} else {
3553 		/* If packet was not sent due to local congestion,
3554 		 * do not backoff and do not remember icsk_probes_out.
3555 		 * Let local senders to fight for local resources.
3556 		 *
3557 		 * Use accumulated backoff yet.
3558 		 */
3559 		if (!icsk->icsk_probes_out)
3560 			icsk->icsk_probes_out = 1;
3561 		probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3562 	}
3563 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3564 				  tcp_probe0_when(sk, probe_max),
3565 				  TCP_RTO_MAX);
3566 }
3567 
tcp_rtx_synack(const struct sock * sk,struct request_sock * req)3568 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3569 {
3570 	const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3571 	struct flowi fl;
3572 	int res;
3573 
3574 	tcp_rsk(req)->txhash = net_tx_rndhash();
3575 	res = af_ops->send_synack(sk, NULL, &fl, req, NULL, true);
3576 	if (!res) {
3577 		TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_RETRANSSEGS);
3578 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3579 	}
3580 	return res;
3581 }
3582 EXPORT_SYMBOL(tcp_rtx_synack);
3583