xref: /linux/net/ipv4/tcp_output.c (revision 634ec1fc7982efeeeeed4a7688b0004827b43a21)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * INET		An implementation of the TCP/IP protocol suite for the LINUX
4  *		operating system.  INET is implemented using the  BSD Socket
5  *		interface as the means of communication with the user level.
6  *
7  *		Implementation of the Transmission Control Protocol(TCP).
8  *
9  * Authors:	Ross Biro
10  *		Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11  *		Mark Evans, <evansmp@uhura.aston.ac.uk>
12  *		Corey Minyard <wf-rch!minyard@relay.EU.net>
13  *		Florian La Roche, <flla@stud.uni-sb.de>
14  *		Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15  *		Linus Torvalds, <torvalds@cs.helsinki.fi>
16  *		Alan Cox, <gw4pts@gw4pts.ampr.org>
17  *		Matthew Dillon, <dillon@apollo.west.oic.com>
18  *		Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19  *		Jorge Cwik, <jorge@laser.satlink.net>
20  */
21 
22 /*
23  * Changes:	Pedro Roque	:	Retransmit queue handled by TCP.
24  *				:	Fragmentation on mtu decrease
25  *				:	Segment collapse on retransmit
26  *				:	AF independence
27  *
28  *		Linus Torvalds	:	send_delayed_ack
29  *		David S. Miller	:	Charge memory using the right skb
30  *					during syn/ack processing.
31  *		David S. Miller :	Output engine completely rewritten.
32  *		Andrea Arcangeli:	SYNACK carry ts_recent in tsecr.
33  *		Cacophonix Gaul :	draft-minshall-nagle-01
34  *		J Hadi Salim	:	ECN support
35  *
36  */
37 
38 #define pr_fmt(fmt) "TCP: " fmt
39 
40 #include <net/tcp.h>
41 #include <net/tcp_ecn.h>
42 #include <net/mptcp.h>
43 #include <net/proto_memory.h>
44 #include <net/psp.h>
45 
46 #include <linux/compiler.h>
47 #include <linux/gfp.h>
48 #include <linux/module.h>
49 #include <linux/static_key.h>
50 #include <linux/skbuff_ref.h>
51 
52 #include <trace/events/tcp.h>
53 
54 /* Refresh clocks of a TCP socket,
55  * ensuring monotically increasing values.
56  */
tcp_mstamp_refresh(struct tcp_sock * tp)57 void tcp_mstamp_refresh(struct tcp_sock *tp)
58 {
59 	u64 val = tcp_clock_ns();
60 
61 	tp->tcp_clock_cache = val;
62 	tp->tcp_mstamp = div_u64(val, NSEC_PER_USEC);
63 }
64 
65 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
66 			   int push_one, gfp_t gfp);
67 
68 /* Account for new data that has been sent to the network. */
tcp_event_new_data_sent(struct sock * sk,struct sk_buff * skb)69 static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
70 {
71 	struct inet_connection_sock *icsk = inet_csk(sk);
72 	struct tcp_sock *tp = tcp_sk(sk);
73 	unsigned int prior_packets = tp->packets_out;
74 
75 	WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(skb)->end_seq);
76 
77 	__skb_unlink(skb, &sk->sk_write_queue);
78 	tcp_rbtree_insert(&sk->tcp_rtx_queue, skb);
79 
80 	if (tp->highest_sack == NULL)
81 		tp->highest_sack = skb;
82 
83 	tp->packets_out += tcp_skb_pcount(skb);
84 	if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
85 		tcp_rearm_rto(sk);
86 
87 	NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
88 		      tcp_skb_pcount(skb));
89 	tcp_check_space(sk);
90 }
91 
92 /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
93  * window scaling factor due to loss of precision.
94  * If window has been shrunk, what should we make? It is not clear at all.
95  * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
96  * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
97  * invalid. OK, let's make this for now:
98  */
tcp_acceptable_seq(const struct sock * sk)99 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
100 {
101 	const struct tcp_sock *tp = tcp_sk(sk);
102 
103 	if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
104 	    (tp->rx_opt.wscale_ok &&
105 	     ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
106 		return tp->snd_nxt;
107 	else
108 		return tcp_wnd_end(tp);
109 }
110 
111 /* Calculate mss to advertise in SYN segment.
112  * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
113  *
114  * 1. It is independent of path mtu.
115  * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
116  * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
117  *    attached devices, because some buggy hosts are confused by
118  *    large MSS.
119  * 4. We do not make 3, we advertise MSS, calculated from first
120  *    hop device mtu, but allow to raise it to ip_rt_min_advmss.
121  *    This may be overridden via information stored in routing table.
122  * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
123  *    probably even Jumbo".
124  */
tcp_advertise_mss(struct sock * sk)125 static __u16 tcp_advertise_mss(struct sock *sk)
126 {
127 	struct tcp_sock *tp = tcp_sk(sk);
128 	const struct dst_entry *dst = __sk_dst_get(sk);
129 	int mss = tp->advmss;
130 
131 	if (dst) {
132 		unsigned int metric = dst_metric_advmss(dst);
133 
134 		if (metric < mss) {
135 			mss = metric;
136 			tp->advmss = mss;
137 		}
138 	}
139 
140 	return (__u16)mss;
141 }
142 
143 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
144  * This is the first part of cwnd validation mechanism.
145  */
tcp_cwnd_restart(struct sock * sk,s32 delta)146 void tcp_cwnd_restart(struct sock *sk, s32 delta)
147 {
148 	struct tcp_sock *tp = tcp_sk(sk);
149 	u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
150 	u32 cwnd = tcp_snd_cwnd(tp);
151 
152 	tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
153 
154 	tp->snd_ssthresh = tcp_current_ssthresh(sk);
155 	restart_cwnd = min(restart_cwnd, cwnd);
156 
157 	while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
158 		cwnd >>= 1;
159 	tcp_snd_cwnd_set(tp, max(cwnd, restart_cwnd));
160 	tp->snd_cwnd_stamp = tcp_jiffies32;
161 	tp->snd_cwnd_used = 0;
162 }
163 
164 /* Congestion state accounting after a packet has been sent. */
tcp_event_data_sent(struct tcp_sock * tp,struct sock * sk)165 static void tcp_event_data_sent(struct tcp_sock *tp,
166 				struct sock *sk)
167 {
168 	struct inet_connection_sock *icsk = inet_csk(sk);
169 	const u32 now = tcp_jiffies32;
170 
171 	if (tcp_packets_in_flight(tp) == 0)
172 		tcp_ca_event(sk, CA_EVENT_TX_START);
173 
174 	tp->lsndtime = now;
175 
176 	/* If it is a reply for ato after last received
177 	 * packet, increase pingpong count.
178 	 */
179 	if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
180 		inet_csk_inc_pingpong_cnt(sk);
181 }
182 
183 /* Account for an ACK we sent. */
tcp_event_ack_sent(struct sock * sk,u32 rcv_nxt)184 static inline void tcp_event_ack_sent(struct sock *sk, u32 rcv_nxt)
185 {
186 	struct tcp_sock *tp = tcp_sk(sk);
187 
188 	if (unlikely(tp->compressed_ack)) {
189 		NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
190 			      tp->compressed_ack);
191 		tp->compressed_ack = 0;
192 		if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
193 			__sock_put(sk);
194 	}
195 
196 	if (unlikely(rcv_nxt != tp->rcv_nxt))
197 		return;  /* Special ACK sent by DCTCP to reflect ECN */
198 	tcp_dec_quickack_mode(sk);
199 	inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
200 }
201 
202 /* Determine a window scaling and initial window to offer.
203  * Based on the assumption that the given amount of space
204  * will be offered. Store the results in the tp structure.
205  * NOTE: for smooth operation initial space offering should
206  * be a multiple of mss if possible. We assume here that mss >= 1.
207  * This MUST be enforced by all callers.
208  */
tcp_select_initial_window(const struct sock * sk,int __space,__u32 mss,__u32 * rcv_wnd,__u32 * __window_clamp,int wscale_ok,__u8 * rcv_wscale,__u32 init_rcv_wnd)209 void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
210 			       __u32 *rcv_wnd, __u32 *__window_clamp,
211 			       int wscale_ok, __u8 *rcv_wscale,
212 			       __u32 init_rcv_wnd)
213 {
214 	unsigned int space = (__space < 0 ? 0 : __space);
215 	u32 window_clamp = READ_ONCE(*__window_clamp);
216 
217 	/* If no clamp set the clamp to the max possible scaled window */
218 	if (window_clamp == 0)
219 		window_clamp = (U16_MAX << TCP_MAX_WSCALE);
220 	space = min(window_clamp, space);
221 
222 	/* Quantize space offering to a multiple of mss if possible. */
223 	if (space > mss)
224 		space = rounddown(space, mss);
225 
226 	/* NOTE: offering an initial window larger than 32767
227 	 * will break some buggy TCP stacks. If the admin tells us
228 	 * it is likely we could be speaking with such a buggy stack
229 	 * we will truncate our initial window offering to 32K-1
230 	 * unless the remote has sent us a window scaling option,
231 	 * which we interpret as a sign the remote TCP is not
232 	 * misinterpreting the window field as a signed quantity.
233 	 */
234 	if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows))
235 		(*rcv_wnd) = min(space, MAX_TCP_WINDOW);
236 	else
237 		(*rcv_wnd) = space;
238 
239 	if (init_rcv_wnd)
240 		*rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
241 
242 	*rcv_wscale = 0;
243 	if (wscale_ok) {
244 		/* Set window scaling on max possible window */
245 		space = max_t(u32, space, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
246 		space = max_t(u32, space, READ_ONCE(sysctl_rmem_max));
247 		space = min_t(u32, space, window_clamp);
248 		*rcv_wscale = clamp_t(int, ilog2(space) - 15,
249 				      0, TCP_MAX_WSCALE);
250 	}
251 	/* Set the clamp no higher than max representable value */
252 	WRITE_ONCE(*__window_clamp,
253 		   min_t(__u32, U16_MAX << (*rcv_wscale), window_clamp));
254 }
255 EXPORT_IPV6_MOD(tcp_select_initial_window);
256 
257 /* Chose a new window to advertise, update state in tcp_sock for the
258  * socket, and return result with RFC1323 scaling applied.  The return
259  * value can be stuffed directly into th->window for an outgoing
260  * frame.
261  */
tcp_select_window(struct sock * sk)262 static u16 tcp_select_window(struct sock *sk)
263 {
264 	struct tcp_sock *tp = tcp_sk(sk);
265 	struct net *net = sock_net(sk);
266 	u32 old_win = tp->rcv_wnd;
267 	u32 cur_win, new_win;
268 
269 	/* Make the window 0 if we failed to queue the data because we
270 	 * are out of memory.
271 	 */
272 	if (unlikely(inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOMEM)) {
273 		tp->pred_flags = 0;
274 		tp->rcv_wnd = 0;
275 		tp->rcv_wup = tp->rcv_nxt;
276 		return 0;
277 	}
278 
279 	cur_win = tcp_receive_window(tp);
280 	new_win = __tcp_select_window(sk);
281 	if (new_win < cur_win) {
282 		/* Danger Will Robinson!
283 		 * Don't update rcv_wup/rcv_wnd here or else
284 		 * we will not be able to advertise a zero
285 		 * window in time.  --DaveM
286 		 *
287 		 * Relax Will Robinson.
288 		 */
289 		if (!READ_ONCE(net->ipv4.sysctl_tcp_shrink_window) || !tp->rx_opt.rcv_wscale) {
290 			/* Never shrink the offered window */
291 			if (new_win == 0)
292 				NET_INC_STATS(net, LINUX_MIB_TCPWANTZEROWINDOWADV);
293 			new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
294 		}
295 	}
296 
297 	tp->rcv_wnd = new_win;
298 	tp->rcv_wup = tp->rcv_nxt;
299 
300 	/* Make sure we do not exceed the maximum possible
301 	 * scaled window.
302 	 */
303 	if (!tp->rx_opt.rcv_wscale &&
304 	    READ_ONCE(net->ipv4.sysctl_tcp_workaround_signed_windows))
305 		new_win = min(new_win, MAX_TCP_WINDOW);
306 	else
307 		new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
308 
309 	/* RFC1323 scaling applied */
310 	new_win >>= tp->rx_opt.rcv_wscale;
311 
312 	/* If we advertise zero window, disable fast path. */
313 	if (new_win == 0) {
314 		tp->pred_flags = 0;
315 		if (old_win)
316 			NET_INC_STATS(net, LINUX_MIB_TCPTOZEROWINDOWADV);
317 	} else if (old_win == 0) {
318 		NET_INC_STATS(net, LINUX_MIB_TCPFROMZEROWINDOWADV);
319 	}
320 
321 	return new_win;
322 }
323 
324 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
325  * be sent.
326  */
tcp_ecn_send(struct sock * sk,struct sk_buff * skb,struct tcphdr * th,int tcp_header_len)327 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
328 			 struct tcphdr *th, int tcp_header_len)
329 {
330 	struct tcp_sock *tp = tcp_sk(sk);
331 
332 	if (!tcp_ecn_mode_any(tp))
333 		return;
334 
335 	if (tcp_ecn_mode_accecn(tp)) {
336 		if (!tcp_accecn_ace_fail_recv(tp))
337 			INET_ECN_xmit(sk);
338 		tcp_accecn_set_ace(tp, skb, th);
339 		skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ACCECN;
340 	} else {
341 		/* Not-retransmitted data segment: set ECT and inject CWR. */
342 		if (skb->len != tcp_header_len &&
343 		    !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
344 			INET_ECN_xmit(sk);
345 			if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
346 				tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
347 				th->cwr = 1;
348 				skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
349 			}
350 		} else if (!tcp_ca_needs_ecn(sk)) {
351 			/* ACK or retransmitted segment: clear ECT|CE */
352 			INET_ECN_dontxmit(sk);
353 		}
354 		if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
355 			th->ece = 1;
356 	}
357 }
358 
359 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
360  * auto increment end seqno.
361  */
tcp_init_nondata_skb(struct sk_buff * skb,struct sock * sk,u32 seq,u16 flags)362 static void tcp_init_nondata_skb(struct sk_buff *skb, struct sock *sk,
363 				 u32 seq, u16 flags)
364 {
365 	skb->ip_summed = CHECKSUM_PARTIAL;
366 
367 	TCP_SKB_CB(skb)->tcp_flags = flags;
368 
369 	tcp_skb_pcount_set(skb, 1);
370 	psp_enqueue_set_decrypted(sk, skb);
371 
372 	TCP_SKB_CB(skb)->seq = seq;
373 	if (flags & (TCPHDR_SYN | TCPHDR_FIN))
374 		seq++;
375 	TCP_SKB_CB(skb)->end_seq = seq;
376 }
377 
tcp_urg_mode(const struct tcp_sock * tp)378 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
379 {
380 	return tp->snd_una != tp->snd_up;
381 }
382 
383 #define OPTION_SACK_ADVERTISE	BIT(0)
384 #define OPTION_TS		BIT(1)
385 #define OPTION_MD5		BIT(2)
386 #define OPTION_WSCALE		BIT(3)
387 #define OPTION_FAST_OPEN_COOKIE	BIT(8)
388 #define OPTION_SMC		BIT(9)
389 #define OPTION_MPTCP		BIT(10)
390 #define OPTION_AO		BIT(11)
391 #define OPTION_ACCECN		BIT(12)
392 
smc_options_write(__be32 * ptr,u16 * options)393 static void smc_options_write(__be32 *ptr, u16 *options)
394 {
395 #if IS_ENABLED(CONFIG_SMC)
396 	if (static_branch_unlikely(&tcp_have_smc)) {
397 		if (unlikely(OPTION_SMC & *options)) {
398 			*ptr++ = htonl((TCPOPT_NOP  << 24) |
399 				       (TCPOPT_NOP  << 16) |
400 				       (TCPOPT_EXP <<  8) |
401 				       (TCPOLEN_EXP_SMC_BASE));
402 			*ptr++ = htonl(TCPOPT_SMC_MAGIC);
403 		}
404 	}
405 #endif
406 }
407 
408 struct tcp_out_options {
409 	u16 options;		/* bit field of OPTION_* */
410 	u16 mss;		/* 0 to disable */
411 	u8 ws;			/* window scale, 0 to disable */
412 	u8 num_sack_blocks;	/* number of SACK blocks to include */
413 	u8 num_accecn_fields:7,	/* number of AccECN fields needed */
414 	   use_synack_ecn_bytes:1; /* Use synack_ecn_bytes or not */
415 	u8 hash_size;		/* bytes in hash_location */
416 	u8 bpf_opt_len;		/* length of BPF hdr option */
417 	__u8 *hash_location;	/* temporary pointer, overloaded */
418 	__u32 tsval, tsecr;	/* need to include OPTION_TS */
419 	struct tcp_fastopen_cookie *fastopen_cookie;	/* Fast open cookie */
420 	struct mptcp_out_options mptcp;
421 };
422 
mptcp_options_write(struct tcphdr * th,__be32 * ptr,struct tcp_sock * tp,struct tcp_out_options * opts)423 static void mptcp_options_write(struct tcphdr *th, __be32 *ptr,
424 				struct tcp_sock *tp,
425 				struct tcp_out_options *opts)
426 {
427 #if IS_ENABLED(CONFIG_MPTCP)
428 	if (unlikely(OPTION_MPTCP & opts->options))
429 		mptcp_write_options(th, ptr, tp, &opts->mptcp);
430 #endif
431 }
432 
433 #ifdef CONFIG_CGROUP_BPF
bpf_skops_write_hdr_opt_arg0(struct sk_buff * skb,enum tcp_synack_type synack_type)434 static int bpf_skops_write_hdr_opt_arg0(struct sk_buff *skb,
435 					enum tcp_synack_type synack_type)
436 {
437 	if (unlikely(!skb))
438 		return BPF_WRITE_HDR_TCP_CURRENT_MSS;
439 
440 	if (unlikely(synack_type == TCP_SYNACK_COOKIE))
441 		return BPF_WRITE_HDR_TCP_SYNACK_COOKIE;
442 
443 	return 0;
444 }
445 
446 /* req, syn_skb and synack_type are used when writing synack */
bpf_skops_hdr_opt_len(struct sock * sk,struct sk_buff * skb,struct request_sock * req,struct sk_buff * syn_skb,enum tcp_synack_type synack_type,struct tcp_out_options * opts,unsigned int * remaining)447 static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb,
448 				  struct request_sock *req,
449 				  struct sk_buff *syn_skb,
450 				  enum tcp_synack_type synack_type,
451 				  struct tcp_out_options *opts,
452 				  unsigned int *remaining)
453 {
454 	struct bpf_sock_ops_kern sock_ops;
455 	int err;
456 
457 	if (likely(!BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
458 					   BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG)) ||
459 	    !*remaining)
460 		return;
461 
462 	/* *remaining has already been aligned to 4 bytes, so *remaining >= 4 */
463 
464 	/* init sock_ops */
465 	memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
466 
467 	sock_ops.op = BPF_SOCK_OPS_HDR_OPT_LEN_CB;
468 
469 	if (req) {
470 		/* The listen "sk" cannot be passed here because
471 		 * it is not locked.  It would not make too much
472 		 * sense to do bpf_setsockopt(listen_sk) based
473 		 * on individual connection request also.
474 		 *
475 		 * Thus, "req" is passed here and the cgroup-bpf-progs
476 		 * of the listen "sk" will be run.
477 		 *
478 		 * "req" is also used here for fastopen even the "sk" here is
479 		 * a fullsock "child" sk.  It is to keep the behavior
480 		 * consistent between fastopen and non-fastopen on
481 		 * the bpf programming side.
482 		 */
483 		sock_ops.sk = (struct sock *)req;
484 		sock_ops.syn_skb = syn_skb;
485 	} else {
486 		sock_owned_by_me(sk);
487 
488 		sock_ops.is_fullsock = 1;
489 		sock_ops.is_locked_tcp_sock = 1;
490 		sock_ops.sk = sk;
491 	}
492 
493 	sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type);
494 	sock_ops.remaining_opt_len = *remaining;
495 	/* tcp_current_mss() does not pass a skb */
496 	if (skb)
497 		bpf_skops_init_skb(&sock_ops, skb, 0);
498 
499 	err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk);
500 
501 	if (err || sock_ops.remaining_opt_len == *remaining)
502 		return;
503 
504 	opts->bpf_opt_len = *remaining - sock_ops.remaining_opt_len;
505 	/* round up to 4 bytes */
506 	opts->bpf_opt_len = (opts->bpf_opt_len + 3) & ~3;
507 
508 	*remaining -= opts->bpf_opt_len;
509 }
510 
bpf_skops_write_hdr_opt(struct sock * sk,struct sk_buff * skb,struct request_sock * req,struct sk_buff * syn_skb,enum tcp_synack_type synack_type,struct tcp_out_options * opts)511 static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb,
512 				    struct request_sock *req,
513 				    struct sk_buff *syn_skb,
514 				    enum tcp_synack_type synack_type,
515 				    struct tcp_out_options *opts)
516 {
517 	u8 first_opt_off, nr_written, max_opt_len = opts->bpf_opt_len;
518 	struct bpf_sock_ops_kern sock_ops;
519 	int err;
520 
521 	if (likely(!max_opt_len))
522 		return;
523 
524 	memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
525 
526 	sock_ops.op = BPF_SOCK_OPS_WRITE_HDR_OPT_CB;
527 
528 	if (req) {
529 		sock_ops.sk = (struct sock *)req;
530 		sock_ops.syn_skb = syn_skb;
531 	} else {
532 		sock_owned_by_me(sk);
533 
534 		sock_ops.is_fullsock = 1;
535 		sock_ops.is_locked_tcp_sock = 1;
536 		sock_ops.sk = sk;
537 	}
538 
539 	sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type);
540 	sock_ops.remaining_opt_len = max_opt_len;
541 	first_opt_off = tcp_hdrlen(skb) - max_opt_len;
542 	bpf_skops_init_skb(&sock_ops, skb, first_opt_off);
543 
544 	err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk);
545 
546 	if (err)
547 		nr_written = 0;
548 	else
549 		nr_written = max_opt_len - sock_ops.remaining_opt_len;
550 
551 	if (nr_written < max_opt_len)
552 		memset(skb->data + first_opt_off + nr_written, TCPOPT_NOP,
553 		       max_opt_len - nr_written);
554 }
555 #else
bpf_skops_hdr_opt_len(struct sock * sk,struct sk_buff * skb,struct request_sock * req,struct sk_buff * syn_skb,enum tcp_synack_type synack_type,struct tcp_out_options * opts,unsigned int * remaining)556 static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb,
557 				  struct request_sock *req,
558 				  struct sk_buff *syn_skb,
559 				  enum tcp_synack_type synack_type,
560 				  struct tcp_out_options *opts,
561 				  unsigned int *remaining)
562 {
563 }
564 
bpf_skops_write_hdr_opt(struct sock * sk,struct sk_buff * skb,struct request_sock * req,struct sk_buff * syn_skb,enum tcp_synack_type synack_type,struct tcp_out_options * opts)565 static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb,
566 				    struct request_sock *req,
567 				    struct sk_buff *syn_skb,
568 				    enum tcp_synack_type synack_type,
569 				    struct tcp_out_options *opts)
570 {
571 }
572 #endif
573 
process_tcp_ao_options(struct tcp_sock * tp,const struct tcp_request_sock * tcprsk,struct tcp_out_options * opts,struct tcp_key * key,__be32 * ptr)574 static __be32 *process_tcp_ao_options(struct tcp_sock *tp,
575 				      const struct tcp_request_sock *tcprsk,
576 				      struct tcp_out_options *opts,
577 				      struct tcp_key *key, __be32 *ptr)
578 {
579 #ifdef CONFIG_TCP_AO
580 	u8 maclen = tcp_ao_maclen(key->ao_key);
581 
582 	if (tcprsk) {
583 		u8 aolen = maclen + sizeof(struct tcp_ao_hdr);
584 
585 		*ptr++ = htonl((TCPOPT_AO << 24) | (aolen << 16) |
586 			       (tcprsk->ao_keyid << 8) |
587 			       (tcprsk->ao_rcv_next));
588 	} else {
589 		struct tcp_ao_key *rnext_key;
590 		struct tcp_ao_info *ao_info;
591 
592 		ao_info = rcu_dereference_check(tp->ao_info,
593 			lockdep_sock_is_held(&tp->inet_conn.icsk_inet.sk));
594 		rnext_key = READ_ONCE(ao_info->rnext_key);
595 		if (WARN_ON_ONCE(!rnext_key))
596 			return ptr;
597 		*ptr++ = htonl((TCPOPT_AO << 24) |
598 			       (tcp_ao_len(key->ao_key) << 16) |
599 			       (key->ao_key->sndid << 8) |
600 			       (rnext_key->rcvid));
601 	}
602 	opts->hash_location = (__u8 *)ptr;
603 	ptr += maclen / sizeof(*ptr);
604 	if (unlikely(maclen % sizeof(*ptr))) {
605 		memset(ptr, TCPOPT_NOP, sizeof(*ptr));
606 		ptr++;
607 	}
608 #endif
609 	return ptr;
610 }
611 
612 /* Initial values for AccECN option, ordered is based on ECN field bits
613  * similar to received_ecn_bytes. Used for SYN/ACK AccECN option.
614  */
615 static const u32 synack_ecn_bytes[3] = { 0, 0, 0 };
616 
617 /* Write previously computed TCP options to the packet.
618  *
619  * Beware: Something in the Internet is very sensitive to the ordering of
620  * TCP options, we learned this through the hard way, so be careful here.
621  * Luckily we can at least blame others for their non-compliance but from
622  * inter-operability perspective it seems that we're somewhat stuck with
623  * the ordering which we have been using if we want to keep working with
624  * those broken things (not that it currently hurts anybody as there isn't
625  * particular reason why the ordering would need to be changed).
626  *
627  * At least SACK_PERM as the first option is known to lead to a disaster
628  * (but it may well be that other scenarios fail similarly).
629  */
tcp_options_write(struct tcphdr * th,struct tcp_sock * tp,const struct tcp_request_sock * tcprsk,struct tcp_out_options * opts,struct tcp_key * key)630 static void tcp_options_write(struct tcphdr *th, struct tcp_sock *tp,
631 			      const struct tcp_request_sock *tcprsk,
632 			      struct tcp_out_options *opts,
633 			      struct tcp_key *key)
634 {
635 	u8 leftover_highbyte = TCPOPT_NOP; /* replace 1st NOP if avail */
636 	u8 leftover_lowbyte = TCPOPT_NOP;  /* replace 2nd NOP in succession */
637 	__be32 *ptr = (__be32 *)(th + 1);
638 	u16 options = opts->options;	/* mungable copy */
639 
640 	if (tcp_key_is_md5(key)) {
641 		*ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
642 			       (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
643 		/* overload cookie hash location */
644 		opts->hash_location = (__u8 *)ptr;
645 		ptr += 4;
646 	} else if (tcp_key_is_ao(key)) {
647 		ptr = process_tcp_ao_options(tp, tcprsk, opts, key, ptr);
648 	}
649 	if (unlikely(opts->mss)) {
650 		*ptr++ = htonl((TCPOPT_MSS << 24) |
651 			       (TCPOLEN_MSS << 16) |
652 			       opts->mss);
653 	}
654 
655 	if (likely(OPTION_TS & options)) {
656 		if (unlikely(OPTION_SACK_ADVERTISE & options)) {
657 			*ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
658 				       (TCPOLEN_SACK_PERM << 16) |
659 				       (TCPOPT_TIMESTAMP << 8) |
660 				       TCPOLEN_TIMESTAMP);
661 			options &= ~OPTION_SACK_ADVERTISE;
662 		} else {
663 			*ptr++ = htonl((TCPOPT_NOP << 24) |
664 				       (TCPOPT_NOP << 16) |
665 				       (TCPOPT_TIMESTAMP << 8) |
666 				       TCPOLEN_TIMESTAMP);
667 		}
668 		*ptr++ = htonl(opts->tsval);
669 		*ptr++ = htonl(opts->tsecr);
670 	}
671 
672 	if (OPTION_ACCECN & options) {
673 		const u32 *ecn_bytes = opts->use_synack_ecn_bytes ?
674 				       synack_ecn_bytes :
675 				       tp->received_ecn_bytes;
676 		const u8 ect0_idx = INET_ECN_ECT_0 - 1;
677 		const u8 ect1_idx = INET_ECN_ECT_1 - 1;
678 		const u8 ce_idx = INET_ECN_CE - 1;
679 		u32 e0b;
680 		u32 e1b;
681 		u32 ceb;
682 		u8 len;
683 
684 		e0b = ecn_bytes[ect0_idx] + TCP_ACCECN_E0B_INIT_OFFSET;
685 		e1b = ecn_bytes[ect1_idx] + TCP_ACCECN_E1B_INIT_OFFSET;
686 		ceb = ecn_bytes[ce_idx] + TCP_ACCECN_CEB_INIT_OFFSET;
687 		len = TCPOLEN_ACCECN_BASE +
688 		      opts->num_accecn_fields * TCPOLEN_ACCECN_PERFIELD;
689 
690 		if (opts->num_accecn_fields == 2) {
691 			*ptr++ = htonl((TCPOPT_ACCECN1 << 24) | (len << 16) |
692 				       ((e1b >> 8) & 0xffff));
693 			*ptr++ = htonl(((e1b & 0xff) << 24) |
694 				       (ceb & 0xffffff));
695 		} else if (opts->num_accecn_fields == 1) {
696 			*ptr++ = htonl((TCPOPT_ACCECN1 << 24) | (len << 16) |
697 				       ((e1b >> 8) & 0xffff));
698 			leftover_highbyte = e1b & 0xff;
699 			leftover_lowbyte = TCPOPT_NOP;
700 		} else if (opts->num_accecn_fields == 0) {
701 			leftover_highbyte = TCPOPT_ACCECN1;
702 			leftover_lowbyte = len;
703 		} else if (opts->num_accecn_fields == 3) {
704 			*ptr++ = htonl((TCPOPT_ACCECN1 << 24) | (len << 16) |
705 				       ((e1b >> 8) & 0xffff));
706 			*ptr++ = htonl(((e1b & 0xff) << 24) |
707 				       (ceb & 0xffffff));
708 			*ptr++ = htonl(((e0b & 0xffffff) << 8) |
709 				       TCPOPT_NOP);
710 		}
711 		if (tp) {
712 			tp->accecn_minlen = 0;
713 			tp->accecn_opt_tstamp = tp->tcp_mstamp;
714 			if (tp->accecn_opt_demand)
715 				tp->accecn_opt_demand--;
716 		}
717 	}
718 
719 	if (unlikely(OPTION_SACK_ADVERTISE & options)) {
720 		*ptr++ = htonl((leftover_highbyte << 24) |
721 			       (leftover_lowbyte << 16) |
722 			       (TCPOPT_SACK_PERM << 8) |
723 			       TCPOLEN_SACK_PERM);
724 		leftover_highbyte = TCPOPT_NOP;
725 		leftover_lowbyte = TCPOPT_NOP;
726 	}
727 
728 	if (unlikely(OPTION_WSCALE & options)) {
729 		u8 highbyte = TCPOPT_NOP;
730 
731 		/* Do not split the leftover 2-byte to fit into a single
732 		 * NOP, i.e., replace this NOP only when 1 byte is leftover
733 		 * within leftover_highbyte.
734 		 */
735 		if (unlikely(leftover_highbyte != TCPOPT_NOP &&
736 			     leftover_lowbyte == TCPOPT_NOP)) {
737 			highbyte = leftover_highbyte;
738 			leftover_highbyte = TCPOPT_NOP;
739 		}
740 		*ptr++ = htonl((highbyte << 24) |
741 			       (TCPOPT_WINDOW << 16) |
742 			       (TCPOLEN_WINDOW << 8) |
743 			       opts->ws);
744 	}
745 
746 	if (unlikely(opts->num_sack_blocks)) {
747 		struct tcp_sack_block *sp = tp->rx_opt.dsack ?
748 			tp->duplicate_sack : tp->selective_acks;
749 		int this_sack;
750 
751 		*ptr++ = htonl((leftover_highbyte << 24) |
752 			       (leftover_lowbyte << 16) |
753 			       (TCPOPT_SACK <<  8) |
754 			       (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
755 						     TCPOLEN_SACK_PERBLOCK)));
756 		leftover_highbyte = TCPOPT_NOP;
757 		leftover_lowbyte = TCPOPT_NOP;
758 
759 		for (this_sack = 0; this_sack < opts->num_sack_blocks;
760 		     ++this_sack) {
761 			*ptr++ = htonl(sp[this_sack].start_seq);
762 			*ptr++ = htonl(sp[this_sack].end_seq);
763 		}
764 
765 		tp->rx_opt.dsack = 0;
766 	} else if (unlikely(leftover_highbyte != TCPOPT_NOP ||
767 			    leftover_lowbyte != TCPOPT_NOP)) {
768 		*ptr++ = htonl((leftover_highbyte << 24) |
769 			       (leftover_lowbyte << 16) |
770 			       (TCPOPT_NOP << 8) |
771 			       TCPOPT_NOP);
772 		leftover_highbyte = TCPOPT_NOP;
773 		leftover_lowbyte = TCPOPT_NOP;
774 	}
775 
776 	if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
777 		struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
778 		u8 *p = (u8 *)ptr;
779 		u32 len; /* Fast Open option length */
780 
781 		if (foc->exp) {
782 			len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
783 			*ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
784 				     TCPOPT_FASTOPEN_MAGIC);
785 			p += TCPOLEN_EXP_FASTOPEN_BASE;
786 		} else {
787 			len = TCPOLEN_FASTOPEN_BASE + foc->len;
788 			*p++ = TCPOPT_FASTOPEN;
789 			*p++ = len;
790 		}
791 
792 		memcpy(p, foc->val, foc->len);
793 		if ((len & 3) == 2) {
794 			p[foc->len] = TCPOPT_NOP;
795 			p[foc->len + 1] = TCPOPT_NOP;
796 		}
797 		ptr += (len + 3) >> 2;
798 	}
799 
800 	smc_options_write(ptr, &options);
801 
802 	mptcp_options_write(th, ptr, tp, opts);
803 }
804 
smc_set_option(const struct tcp_sock * tp,struct tcp_out_options * opts,unsigned int * remaining)805 static void smc_set_option(const struct tcp_sock *tp,
806 			   struct tcp_out_options *opts,
807 			   unsigned int *remaining)
808 {
809 #if IS_ENABLED(CONFIG_SMC)
810 	if (static_branch_unlikely(&tcp_have_smc)) {
811 		if (tp->syn_smc) {
812 			if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
813 				opts->options |= OPTION_SMC;
814 				*remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
815 			}
816 		}
817 	}
818 #endif
819 }
820 
smc_set_option_cond(const struct tcp_sock * tp,const struct inet_request_sock * ireq,struct tcp_out_options * opts,unsigned int * remaining)821 static void smc_set_option_cond(const struct tcp_sock *tp,
822 				const struct inet_request_sock *ireq,
823 				struct tcp_out_options *opts,
824 				unsigned int *remaining)
825 {
826 #if IS_ENABLED(CONFIG_SMC)
827 	if (static_branch_unlikely(&tcp_have_smc)) {
828 		if (tp->syn_smc && ireq->smc_ok) {
829 			if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
830 				opts->options |= OPTION_SMC;
831 				*remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
832 			}
833 		}
834 	}
835 #endif
836 }
837 
mptcp_set_option_cond(const struct request_sock * req,struct tcp_out_options * opts,unsigned int * remaining)838 static void mptcp_set_option_cond(const struct request_sock *req,
839 				  struct tcp_out_options *opts,
840 				  unsigned int *remaining)
841 {
842 	if (rsk_is_mptcp(req)) {
843 		unsigned int size;
844 
845 		if (mptcp_synack_options(req, &size, &opts->mptcp)) {
846 			if (*remaining >= size) {
847 				opts->options |= OPTION_MPTCP;
848 				*remaining -= size;
849 			}
850 		}
851 	}
852 }
853 
tcp_synack_options_combine_saving(struct tcp_out_options * opts)854 static u32 tcp_synack_options_combine_saving(struct tcp_out_options *opts)
855 {
856 	/* How much there's room for combining with the alignment padding? */
857 	if ((opts->options & (OPTION_SACK_ADVERTISE | OPTION_TS)) ==
858 	    OPTION_SACK_ADVERTISE)
859 		return 2;
860 	else if (opts->options & OPTION_WSCALE)
861 		return 1;
862 	return 0;
863 }
864 
865 /* Calculates how long AccECN option will fit to @remaining option space.
866  *
867  * AccECN option can sometimes replace NOPs used for alignment of other
868  * TCP options (up to @max_combine_saving available).
869  *
870  * Only solutions with at least @required AccECN fields are accepted.
871  *
872  * Returns: The size of the AccECN option excluding space repurposed from
873  * the alignment of the other options.
874  */
tcp_options_fit_accecn(struct tcp_out_options * opts,int required,int remaining)875 static int tcp_options_fit_accecn(struct tcp_out_options *opts, int required,
876 				  int remaining)
877 {
878 	int size = TCP_ACCECN_MAXSIZE;
879 	int sack_blocks_reduce = 0;
880 	int max_combine_saving;
881 	int rem = remaining;
882 	int align_size;
883 
884 	if (opts->use_synack_ecn_bytes)
885 		max_combine_saving = tcp_synack_options_combine_saving(opts);
886 	else
887 		max_combine_saving = opts->num_sack_blocks > 0 ? 2 : 0;
888 	opts->num_accecn_fields = TCP_ACCECN_NUMFIELDS;
889 	while (opts->num_accecn_fields >= required) {
890 		/* Pad to dword if cannot combine */
891 		if ((size & 0x3) > max_combine_saving)
892 			align_size = ALIGN(size, 4);
893 		else
894 			align_size = ALIGN_DOWN(size, 4);
895 
896 		if (rem >= align_size) {
897 			size = align_size;
898 			break;
899 		} else if (opts->num_accecn_fields == required &&
900 			   opts->num_sack_blocks > 2 &&
901 			   required > 0) {
902 			/* Try to fit the option by removing one SACK block */
903 			opts->num_sack_blocks--;
904 			sack_blocks_reduce++;
905 			rem = rem + TCPOLEN_SACK_PERBLOCK;
906 
907 			opts->num_accecn_fields = TCP_ACCECN_NUMFIELDS;
908 			size = TCP_ACCECN_MAXSIZE;
909 			continue;
910 		}
911 
912 		opts->num_accecn_fields--;
913 		size -= TCPOLEN_ACCECN_PERFIELD;
914 	}
915 	if (sack_blocks_reduce > 0) {
916 		if (opts->num_accecn_fields >= required)
917 			size -= sack_blocks_reduce * TCPOLEN_SACK_PERBLOCK;
918 		else
919 			opts->num_sack_blocks += sack_blocks_reduce;
920 	}
921 	if (opts->num_accecn_fields < required)
922 		return 0;
923 
924 	opts->options |= OPTION_ACCECN;
925 	return size;
926 }
927 
928 /* Compute TCP options for SYN packets. This is not the final
929  * network wire format yet.
930  */
tcp_syn_options(struct sock * sk,struct sk_buff * skb,struct tcp_out_options * opts,struct tcp_key * key)931 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
932 				struct tcp_out_options *opts,
933 				struct tcp_key *key)
934 {
935 	struct tcp_sock *tp = tcp_sk(sk);
936 	unsigned int remaining = MAX_TCP_OPTION_SPACE;
937 	struct tcp_fastopen_request *fastopen = tp->fastopen_req;
938 	bool timestamps;
939 
940 	/* Better than switch (key.type) as it has static branches */
941 	if (tcp_key_is_md5(key)) {
942 		timestamps = false;
943 		opts->options |= OPTION_MD5;
944 		remaining -= TCPOLEN_MD5SIG_ALIGNED;
945 	} else {
946 		timestamps = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps);
947 		if (tcp_key_is_ao(key)) {
948 			opts->options |= OPTION_AO;
949 			remaining -= tcp_ao_len_aligned(key->ao_key);
950 		}
951 	}
952 
953 	/* We always get an MSS option.  The option bytes which will be seen in
954 	 * normal data packets should timestamps be used, must be in the MSS
955 	 * advertised.  But we subtract them from tp->mss_cache so that
956 	 * calculations in tcp_sendmsg are simpler etc.  So account for this
957 	 * fact here if necessary.  If we don't do this correctly, as a
958 	 * receiver we won't recognize data packets as being full sized when we
959 	 * should, and thus we won't abide by the delayed ACK rules correctly.
960 	 * SACKs don't matter, we never delay an ACK when we have any of those
961 	 * going out.  */
962 	opts->mss = tcp_advertise_mss(sk);
963 	remaining -= TCPOLEN_MSS_ALIGNED;
964 
965 	if (likely(timestamps)) {
966 		opts->options |= OPTION_TS;
967 		opts->tsval = tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb) + tp->tsoffset;
968 		opts->tsecr = tp->rx_opt.ts_recent;
969 		remaining -= TCPOLEN_TSTAMP_ALIGNED;
970 	}
971 	if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling))) {
972 		opts->ws = tp->rx_opt.rcv_wscale;
973 		opts->options |= OPTION_WSCALE;
974 		remaining -= TCPOLEN_WSCALE_ALIGNED;
975 	}
976 	if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_sack))) {
977 		opts->options |= OPTION_SACK_ADVERTISE;
978 		if (unlikely(!(OPTION_TS & opts->options)))
979 			remaining -= TCPOLEN_SACKPERM_ALIGNED;
980 	}
981 
982 	if (fastopen && fastopen->cookie.len >= 0) {
983 		u32 need = fastopen->cookie.len;
984 
985 		need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
986 					       TCPOLEN_FASTOPEN_BASE;
987 		need = (need + 3) & ~3U;  /* Align to 32 bits */
988 		if (remaining >= need) {
989 			opts->options |= OPTION_FAST_OPEN_COOKIE;
990 			opts->fastopen_cookie = &fastopen->cookie;
991 			remaining -= need;
992 			tp->syn_fastopen = 1;
993 			tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
994 		}
995 	}
996 
997 	smc_set_option(tp, opts, &remaining);
998 
999 	if (sk_is_mptcp(sk)) {
1000 		unsigned int size;
1001 
1002 		if (mptcp_syn_options(sk, skb, &size, &opts->mptcp)) {
1003 			if (remaining >= size) {
1004 				opts->options |= OPTION_MPTCP;
1005 				remaining -= size;
1006 			}
1007 		}
1008 	}
1009 
1010 	/* Simultaneous open SYN/ACK needs AccECN option but not SYN.
1011 	 * It is attempted to negotiate the use of AccECN also on the first
1012 	 * retransmitted SYN, as mentioned in "3.1.4.1. Retransmitted SYNs"
1013 	 * of AccECN draft.
1014 	 */
1015 	if (unlikely((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK) &&
1016 		     tcp_ecn_mode_accecn(tp) &&
1017 		     inet_csk(sk)->icsk_retransmits < 2 &&
1018 		     READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn_option) &&
1019 		     remaining >= TCPOLEN_ACCECN_BASE)) {
1020 		opts->use_synack_ecn_bytes = 1;
1021 		remaining -= tcp_options_fit_accecn(opts, 0, remaining);
1022 	}
1023 
1024 	bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining);
1025 
1026 	return MAX_TCP_OPTION_SPACE - remaining;
1027 }
1028 
1029 /* Set up TCP options for SYN-ACKs. */
tcp_synack_options(const struct sock * sk,struct request_sock * req,unsigned int mss,struct sk_buff * skb,struct tcp_out_options * opts,const struct tcp_key * key,struct tcp_fastopen_cookie * foc,enum tcp_synack_type synack_type,struct sk_buff * syn_skb)1030 static unsigned int tcp_synack_options(const struct sock *sk,
1031 				       struct request_sock *req,
1032 				       unsigned int mss, struct sk_buff *skb,
1033 				       struct tcp_out_options *opts,
1034 				       const struct tcp_key *key,
1035 				       struct tcp_fastopen_cookie *foc,
1036 				       enum tcp_synack_type synack_type,
1037 				       struct sk_buff *syn_skb)
1038 {
1039 	struct inet_request_sock *ireq = inet_rsk(req);
1040 	unsigned int remaining = MAX_TCP_OPTION_SPACE;
1041 	struct tcp_request_sock *treq = tcp_rsk(req);
1042 
1043 	if (tcp_key_is_md5(key)) {
1044 		opts->options |= OPTION_MD5;
1045 		remaining -= TCPOLEN_MD5SIG_ALIGNED;
1046 
1047 		/* We can't fit any SACK blocks in a packet with MD5 + TS
1048 		 * options. There was discussion about disabling SACK
1049 		 * rather than TS in order to fit in better with old,
1050 		 * buggy kernels, but that was deemed to be unnecessary.
1051 		 */
1052 		if (synack_type != TCP_SYNACK_COOKIE)
1053 			ireq->tstamp_ok &= !ireq->sack_ok;
1054 	} else if (tcp_key_is_ao(key)) {
1055 		opts->options |= OPTION_AO;
1056 		remaining -= tcp_ao_len_aligned(key->ao_key);
1057 		ireq->tstamp_ok &= !ireq->sack_ok;
1058 	}
1059 
1060 	/* We always send an MSS option. */
1061 	opts->mss = mss;
1062 	remaining -= TCPOLEN_MSS_ALIGNED;
1063 
1064 	if (likely(ireq->wscale_ok)) {
1065 		opts->ws = ireq->rcv_wscale;
1066 		opts->options |= OPTION_WSCALE;
1067 		remaining -= TCPOLEN_WSCALE_ALIGNED;
1068 	}
1069 	if (likely(ireq->tstamp_ok)) {
1070 		opts->options |= OPTION_TS;
1071 		opts->tsval = tcp_skb_timestamp_ts(tcp_rsk(req)->req_usec_ts, skb) +
1072 			      tcp_rsk(req)->ts_off;
1073 		if (!tcp_rsk(req)->snt_tsval_first) {
1074 			if (!opts->tsval)
1075 				opts->tsval = ~0U;
1076 			tcp_rsk(req)->snt_tsval_first = opts->tsval;
1077 		}
1078 		WRITE_ONCE(tcp_rsk(req)->snt_tsval_last, opts->tsval);
1079 		opts->tsecr = req->ts_recent;
1080 		remaining -= TCPOLEN_TSTAMP_ALIGNED;
1081 	}
1082 	if (likely(ireq->sack_ok)) {
1083 		opts->options |= OPTION_SACK_ADVERTISE;
1084 		if (unlikely(!ireq->tstamp_ok))
1085 			remaining -= TCPOLEN_SACKPERM_ALIGNED;
1086 	}
1087 	if (foc != NULL && foc->len >= 0) {
1088 		u32 need = foc->len;
1089 
1090 		need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
1091 				   TCPOLEN_FASTOPEN_BASE;
1092 		need = (need + 3) & ~3U;  /* Align to 32 bits */
1093 		if (remaining >= need) {
1094 			opts->options |= OPTION_FAST_OPEN_COOKIE;
1095 			opts->fastopen_cookie = foc;
1096 			remaining -= need;
1097 		}
1098 	}
1099 
1100 	mptcp_set_option_cond(req, opts, &remaining);
1101 
1102 	smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
1103 
1104 	if (treq->accecn_ok &&
1105 	    READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn_option) &&
1106 	    req->num_timeout < 1 && remaining >= TCPOLEN_ACCECN_BASE) {
1107 		opts->use_synack_ecn_bytes = 1;
1108 		remaining -= tcp_options_fit_accecn(opts, 0, remaining);
1109 	}
1110 
1111 	bpf_skops_hdr_opt_len((struct sock *)sk, skb, req, syn_skb,
1112 			      synack_type, opts, &remaining);
1113 
1114 	return MAX_TCP_OPTION_SPACE - remaining;
1115 }
1116 
1117 /* Compute TCP options for ESTABLISHED sockets. This is not the
1118  * final wire format yet.
1119  */
tcp_established_options(struct sock * sk,struct sk_buff * skb,struct tcp_out_options * opts,struct tcp_key * key)1120 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
1121 					struct tcp_out_options *opts,
1122 					struct tcp_key *key)
1123 {
1124 	struct tcp_sock *tp = tcp_sk(sk);
1125 	unsigned int size = 0;
1126 	unsigned int eff_sacks;
1127 
1128 	opts->options = 0;
1129 
1130 	/* Better than switch (key.type) as it has static branches */
1131 	if (tcp_key_is_md5(key)) {
1132 		opts->options |= OPTION_MD5;
1133 		size += TCPOLEN_MD5SIG_ALIGNED;
1134 	} else if (tcp_key_is_ao(key)) {
1135 		opts->options |= OPTION_AO;
1136 		size += tcp_ao_len_aligned(key->ao_key);
1137 	}
1138 
1139 	if (likely(tp->rx_opt.tstamp_ok)) {
1140 		opts->options |= OPTION_TS;
1141 		opts->tsval = skb ? tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb) +
1142 				tp->tsoffset : 0;
1143 		opts->tsecr = tp->rx_opt.ts_recent;
1144 		size += TCPOLEN_TSTAMP_ALIGNED;
1145 	}
1146 
1147 	/* MPTCP options have precedence over SACK for the limited TCP
1148 	 * option space because a MPTCP connection would be forced to
1149 	 * fall back to regular TCP if a required multipath option is
1150 	 * missing. SACK still gets a chance to use whatever space is
1151 	 * left.
1152 	 */
1153 	if (sk_is_mptcp(sk)) {
1154 		unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
1155 		unsigned int opt_size = 0;
1156 
1157 		if (mptcp_established_options(sk, skb, &opt_size, remaining,
1158 					      &opts->mptcp)) {
1159 			opts->options |= OPTION_MPTCP;
1160 			size += opt_size;
1161 		}
1162 	}
1163 
1164 	eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
1165 	if (unlikely(eff_sacks)) {
1166 		const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
1167 		if (likely(remaining >= TCPOLEN_SACK_BASE_ALIGNED +
1168 					TCPOLEN_SACK_PERBLOCK)) {
1169 			opts->num_sack_blocks =
1170 				min_t(unsigned int, eff_sacks,
1171 				      (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
1172 				      TCPOLEN_SACK_PERBLOCK);
1173 
1174 			size += TCPOLEN_SACK_BASE_ALIGNED +
1175 				opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
1176 		} else {
1177 			opts->num_sack_blocks = 0;
1178 		}
1179 	} else {
1180 		opts->num_sack_blocks = 0;
1181 	}
1182 
1183 	if (tcp_ecn_mode_accecn(tp)) {
1184 		int ecn_opt = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn_option);
1185 
1186 		if (ecn_opt && tp->saw_accecn_opt && !tcp_accecn_opt_fail_send(tp) &&
1187 		    (ecn_opt >= TCP_ACCECN_OPTION_FULL || tp->accecn_opt_demand ||
1188 		     tcp_accecn_option_beacon_check(sk))) {
1189 			opts->use_synack_ecn_bytes = 0;
1190 			size += tcp_options_fit_accecn(opts, tp->accecn_minlen,
1191 						       MAX_TCP_OPTION_SPACE - size);
1192 		}
1193 	}
1194 
1195 	if (unlikely(BPF_SOCK_OPS_TEST_FLAG(tp,
1196 					    BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG))) {
1197 		unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
1198 
1199 		bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining);
1200 
1201 		size = MAX_TCP_OPTION_SPACE - remaining;
1202 	}
1203 
1204 	return size;
1205 }
1206 
1207 
1208 /* TCP SMALL QUEUES (TSQ)
1209  *
1210  * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
1211  * to reduce RTT and bufferbloat.
1212  * We do this using a special skb destructor (tcp_wfree).
1213  *
1214  * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
1215  * needs to be reallocated in a driver.
1216  * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
1217  *
1218  * Since transmit from skb destructor is forbidden, we use a BH work item
1219  * to process all sockets that eventually need to send more skbs.
1220  * We use one work item per cpu, with its own queue of sockets.
1221  */
1222 struct tsq_work {
1223 	struct work_struct	work;
1224 	struct list_head	head; /* queue of tcp sockets */
1225 };
1226 static DEFINE_PER_CPU(struct tsq_work, tsq_work);
1227 
tcp_tsq_write(struct sock * sk)1228 static void tcp_tsq_write(struct sock *sk)
1229 {
1230 	if ((1 << sk->sk_state) &
1231 	    (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
1232 	     TCPF_CLOSE_WAIT  | TCPF_LAST_ACK)) {
1233 		struct tcp_sock *tp = tcp_sk(sk);
1234 
1235 		if (tp->lost_out > tp->retrans_out &&
1236 		    tcp_snd_cwnd(tp) > tcp_packets_in_flight(tp)) {
1237 			tcp_mstamp_refresh(tp);
1238 			tcp_xmit_retransmit_queue(sk);
1239 		}
1240 
1241 		tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
1242 			       0, GFP_ATOMIC);
1243 	}
1244 }
1245 
tcp_tsq_handler(struct sock * sk)1246 static void tcp_tsq_handler(struct sock *sk)
1247 {
1248 	bh_lock_sock(sk);
1249 	if (!sock_owned_by_user(sk))
1250 		tcp_tsq_write(sk);
1251 	else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
1252 		sock_hold(sk);
1253 	bh_unlock_sock(sk);
1254 }
1255 /*
1256  * One work item per cpu tries to send more skbs.
1257  * We run in BH context but need to disable irqs when
1258  * transferring tsq->head because tcp_wfree() might
1259  * interrupt us (non NAPI drivers)
1260  */
tcp_tsq_workfn(struct work_struct * work)1261 static void tcp_tsq_workfn(struct work_struct *work)
1262 {
1263 	struct tsq_work *tsq = container_of(work, struct tsq_work, work);
1264 	LIST_HEAD(list);
1265 	unsigned long flags;
1266 	struct list_head *q, *n;
1267 	struct tcp_sock *tp;
1268 	struct sock *sk;
1269 
1270 	local_irq_save(flags);
1271 	list_splice_init(&tsq->head, &list);
1272 	local_irq_restore(flags);
1273 
1274 	list_for_each_safe(q, n, &list) {
1275 		tp = list_entry(q, struct tcp_sock, tsq_node);
1276 		list_del(&tp->tsq_node);
1277 
1278 		sk = (struct sock *)tp;
1279 		smp_mb__before_atomic();
1280 		clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
1281 
1282 		tcp_tsq_handler(sk);
1283 		sk_free(sk);
1284 	}
1285 }
1286 
1287 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED |		\
1288 			  TCPF_WRITE_TIMER_DEFERRED |	\
1289 			  TCPF_DELACK_TIMER_DEFERRED |	\
1290 			  TCPF_MTU_REDUCED_DEFERRED |	\
1291 			  TCPF_ACK_DEFERRED)
1292 /**
1293  * tcp_release_cb - tcp release_sock() callback
1294  * @sk: socket
1295  *
1296  * called from release_sock() to perform protocol dependent
1297  * actions before socket release.
1298  */
tcp_release_cb(struct sock * sk)1299 void tcp_release_cb(struct sock *sk)
1300 {
1301 	unsigned long flags = smp_load_acquire(&sk->sk_tsq_flags);
1302 	unsigned long nflags;
1303 
1304 	/* perform an atomic operation only if at least one flag is set */
1305 	do {
1306 		if (!(flags & TCP_DEFERRED_ALL))
1307 			return;
1308 		nflags = flags & ~TCP_DEFERRED_ALL;
1309 	} while (!try_cmpxchg(&sk->sk_tsq_flags, &flags, nflags));
1310 
1311 	if (flags & TCPF_TSQ_DEFERRED) {
1312 		tcp_tsq_write(sk);
1313 		__sock_put(sk);
1314 	}
1315 
1316 	if (flags & TCPF_WRITE_TIMER_DEFERRED) {
1317 		tcp_write_timer_handler(sk);
1318 		__sock_put(sk);
1319 	}
1320 	if (flags & TCPF_DELACK_TIMER_DEFERRED) {
1321 		tcp_delack_timer_handler(sk);
1322 		__sock_put(sk);
1323 	}
1324 	if (flags & TCPF_MTU_REDUCED_DEFERRED) {
1325 		inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
1326 		__sock_put(sk);
1327 	}
1328 	if ((flags & TCPF_ACK_DEFERRED) && inet_csk_ack_scheduled(sk))
1329 		tcp_send_ack(sk);
1330 }
1331 EXPORT_IPV6_MOD(tcp_release_cb);
1332 
tcp_tsq_work_init(void)1333 void __init tcp_tsq_work_init(void)
1334 {
1335 	int i;
1336 
1337 	for_each_possible_cpu(i) {
1338 		struct tsq_work *tsq = &per_cpu(tsq_work, i);
1339 
1340 		INIT_LIST_HEAD(&tsq->head);
1341 		INIT_WORK(&tsq->work, tcp_tsq_workfn);
1342 	}
1343 }
1344 
1345 /*
1346  * Write buffer destructor automatically called from kfree_skb.
1347  * We can't xmit new skbs from this context, as we might already
1348  * hold qdisc lock.
1349  */
tcp_wfree(struct sk_buff * skb)1350 void tcp_wfree(struct sk_buff *skb)
1351 {
1352 	struct sock *sk = skb->sk;
1353 	struct tcp_sock *tp = tcp_sk(sk);
1354 	unsigned long flags, nval, oval;
1355 	struct tsq_work *tsq;
1356 	bool empty;
1357 
1358 	/* Keep one reference on sk_wmem_alloc.
1359 	 * Will be released by sk_free() from here or tcp_tsq_workfn()
1360 	 */
1361 	WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
1362 
1363 	/* If this softirq is serviced by ksoftirqd, we are likely under stress.
1364 	 * Wait until our queues (qdisc + devices) are drained.
1365 	 * This gives :
1366 	 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
1367 	 * - chance for incoming ACK (processed by another cpu maybe)
1368 	 *   to migrate this flow (skb->ooo_okay will be eventually set)
1369 	 */
1370 	if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
1371 		goto out;
1372 
1373 	oval = smp_load_acquire(&sk->sk_tsq_flags);
1374 	do {
1375 		if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
1376 			goto out;
1377 
1378 		nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
1379 	} while (!try_cmpxchg(&sk->sk_tsq_flags, &oval, nval));
1380 
1381 	/* queue this socket to BH workqueue */
1382 	local_irq_save(flags);
1383 	tsq = this_cpu_ptr(&tsq_work);
1384 	empty = list_empty(&tsq->head);
1385 	list_add(&tp->tsq_node, &tsq->head);
1386 	if (empty)
1387 		queue_work(system_bh_wq, &tsq->work);
1388 	local_irq_restore(flags);
1389 	return;
1390 out:
1391 	sk_free(sk);
1392 }
1393 
1394 /* Note: Called under soft irq.
1395  * We can call TCP stack right away, unless socket is owned by user.
1396  */
tcp_pace_kick(struct hrtimer * timer)1397 enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
1398 {
1399 	struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
1400 	struct sock *sk = (struct sock *)tp;
1401 
1402 	tcp_tsq_handler(sk);
1403 	sock_put(sk);
1404 
1405 	return HRTIMER_NORESTART;
1406 }
1407 
tcp_update_skb_after_send(struct sock * sk,struct sk_buff * skb,u64 prior_wstamp)1408 static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb,
1409 				      u64 prior_wstamp)
1410 {
1411 	struct tcp_sock *tp = tcp_sk(sk);
1412 
1413 	if (sk->sk_pacing_status != SK_PACING_NONE) {
1414 		unsigned long rate = READ_ONCE(sk->sk_pacing_rate);
1415 
1416 		/* Original sch_fq does not pace first 10 MSS
1417 		 * Note that tp->data_segs_out overflows after 2^32 packets,
1418 		 * this is a minor annoyance.
1419 		 */
1420 		if (rate != ~0UL && rate && tp->data_segs_out >= 10) {
1421 			u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate);
1422 			u64 credit = tp->tcp_wstamp_ns - prior_wstamp;
1423 
1424 			/* take into account OS jitter */
1425 			len_ns -= min_t(u64, len_ns / 2, credit);
1426 			tp->tcp_wstamp_ns += len_ns;
1427 		}
1428 	}
1429 	list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1430 }
1431 
1432 INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1433 INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1434 INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb));
1435 
1436 /* This routine actually transmits TCP packets queued in by
1437  * tcp_do_sendmsg().  This is used by both the initial
1438  * transmission and possible later retransmissions.
1439  * All SKB's seen here are completely headerless.  It is our
1440  * job to build the TCP header, and pass the packet down to
1441  * IP so it can do the same plus pass the packet off to the
1442  * device.
1443  *
1444  * We are working here with either a clone of the original
1445  * SKB, or a fresh unique copy made by the retransmit engine.
1446  */
__tcp_transmit_skb(struct sock * sk,struct sk_buff * skb,int clone_it,gfp_t gfp_mask,u32 rcv_nxt)1447 static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
1448 			      int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
1449 {
1450 	const struct inet_connection_sock *icsk = inet_csk(sk);
1451 	struct inet_sock *inet;
1452 	struct tcp_sock *tp;
1453 	struct tcp_skb_cb *tcb;
1454 	struct tcp_out_options opts;
1455 	unsigned int tcp_options_size, tcp_header_size;
1456 	struct sk_buff *oskb = NULL;
1457 	struct tcp_key key;
1458 	struct tcphdr *th;
1459 	u64 prior_wstamp;
1460 	int err;
1461 
1462 	BUG_ON(!skb || !tcp_skb_pcount(skb));
1463 	tp = tcp_sk(sk);
1464 	prior_wstamp = tp->tcp_wstamp_ns;
1465 	tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache);
1466 	skb_set_delivery_time(skb, tp->tcp_wstamp_ns, SKB_CLOCK_MONOTONIC);
1467 	if (clone_it) {
1468 		oskb = skb;
1469 
1470 		tcp_skb_tsorted_save(oskb) {
1471 			if (unlikely(skb_cloned(oskb)))
1472 				skb = pskb_copy(oskb, gfp_mask);
1473 			else
1474 				skb = skb_clone(oskb, gfp_mask);
1475 		} tcp_skb_tsorted_restore(oskb);
1476 
1477 		if (unlikely(!skb))
1478 			return -ENOBUFS;
1479 		/* retransmit skbs might have a non zero value in skb->dev
1480 		 * because skb->dev is aliased with skb->rbnode.rb_left
1481 		 */
1482 		skb->dev = NULL;
1483 	}
1484 
1485 	inet = inet_sk(sk);
1486 	tcb = TCP_SKB_CB(skb);
1487 	memset(&opts, 0, sizeof(opts));
1488 
1489 	tcp_get_current_key(sk, &key);
1490 	if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
1491 		tcp_options_size = tcp_syn_options(sk, skb, &opts, &key);
1492 	} else {
1493 		tcp_options_size = tcp_established_options(sk, skb, &opts, &key);
1494 		/* Force a PSH flag on all (GSO) packets to expedite GRO flush
1495 		 * at receiver : This slightly improve GRO performance.
1496 		 * Note that we do not force the PSH flag for non GSO packets,
1497 		 * because they might be sent under high congestion events,
1498 		 * and in this case it is better to delay the delivery of 1-MSS
1499 		 * packets and thus the corresponding ACK packet that would
1500 		 * release the following packet.
1501 		 */
1502 		if (tcp_skb_pcount(skb) > 1)
1503 			tcb->tcp_flags |= TCPHDR_PSH;
1504 	}
1505 	tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1506 
1507 	/* We set skb->ooo_okay to one if this packet can select
1508 	 * a different TX queue than prior packets of this flow,
1509 	 * to avoid self inflicted reorders.
1510 	 * The 'other' queue decision is based on current cpu number
1511 	 * if XPS is enabled, or sk->sk_txhash otherwise.
1512 	 * We can switch to another (and better) queue if:
1513 	 * 1) No packet with payload is in qdisc/device queues.
1514 	 *    Delays in TX completion can defeat the test
1515 	 *    even if packets were already sent.
1516 	 * 2) Or rtx queue is empty.
1517 	 *    This mitigates above case if ACK packets for
1518 	 *    all prior packets were already processed.
1519 	 */
1520 	skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1) ||
1521 			tcp_rtx_queue_empty(sk);
1522 
1523 	/* If we had to use memory reserve to allocate this skb,
1524 	 * this might cause drops if packet is looped back :
1525 	 * Other socket might not have SOCK_MEMALLOC.
1526 	 * Packets not looped back do not care about pfmemalloc.
1527 	 */
1528 	skb->pfmemalloc = 0;
1529 
1530 	skb_push(skb, tcp_header_size);
1531 	skb_reset_transport_header(skb);
1532 
1533 	skb_orphan(skb);
1534 	skb->sk = sk;
1535 	skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1536 	refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1537 
1538 	skb_set_dst_pending_confirm(skb, READ_ONCE(sk->sk_dst_pending_confirm));
1539 
1540 	/* Build TCP header and checksum it. */
1541 	th = (struct tcphdr *)skb->data;
1542 	th->source		= inet->inet_sport;
1543 	th->dest		= inet->inet_dport;
1544 	th->seq			= htonl(tcb->seq);
1545 	th->ack_seq		= htonl(rcv_nxt);
1546 	*(((__be16 *)th) + 6)	= htons(((tcp_header_size >> 2) << 12) |
1547 					(tcb->tcp_flags & TCPHDR_FLAGS_MASK));
1548 
1549 	th->check		= 0;
1550 	th->urg_ptr		= 0;
1551 
1552 	/* The urg_mode check is necessary during a below snd_una win probe */
1553 	if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1554 		if (before(tp->snd_up, tcb->seq + 0x10000)) {
1555 			th->urg_ptr = htons(tp->snd_up - tcb->seq);
1556 			th->urg = 1;
1557 		} else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1558 			th->urg_ptr = htons(0xFFFF);
1559 			th->urg = 1;
1560 		}
1561 	}
1562 
1563 	skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1564 	if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1565 		th->window      = htons(tcp_select_window(sk));
1566 		tcp_ecn_send(sk, skb, th, tcp_header_size);
1567 	} else {
1568 		/* RFC1323: The window in SYN & SYN/ACK segments
1569 		 * is never scaled.
1570 		 */
1571 		th->window	= htons(min(tp->rcv_wnd, 65535U));
1572 	}
1573 
1574 	tcp_options_write(th, tp, NULL, &opts, &key);
1575 
1576 	if (tcp_key_is_md5(&key)) {
1577 #ifdef CONFIG_TCP_MD5SIG
1578 		/* Calculate the MD5 hash, as we have all we need now */
1579 		sk_gso_disable(sk);
1580 		tp->af_specific->calc_md5_hash(opts.hash_location,
1581 					       key.md5_key, sk, skb);
1582 #endif
1583 	} else if (tcp_key_is_ao(&key)) {
1584 		int err;
1585 
1586 		err = tcp_ao_transmit_skb(sk, skb, key.ao_key, th,
1587 					  opts.hash_location);
1588 		if (err) {
1589 			sk_skb_reason_drop(sk, skb, SKB_DROP_REASON_NOT_SPECIFIED);
1590 			return -ENOMEM;
1591 		}
1592 	}
1593 
1594 	/* BPF prog is the last one writing header option */
1595 	bpf_skops_write_hdr_opt(sk, skb, NULL, NULL, 0, &opts);
1596 
1597 	INDIRECT_CALL_INET(icsk->icsk_af_ops->send_check,
1598 			   tcp_v6_send_check, tcp_v4_send_check,
1599 			   sk, skb);
1600 
1601 	if (likely(tcb->tcp_flags & TCPHDR_ACK))
1602 		tcp_event_ack_sent(sk, rcv_nxt);
1603 
1604 	if (skb->len != tcp_header_size) {
1605 		tcp_event_data_sent(tp, sk);
1606 		tp->data_segs_out += tcp_skb_pcount(skb);
1607 		tp->bytes_sent += skb->len - tcp_header_size;
1608 	}
1609 
1610 	if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1611 		TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1612 			      tcp_skb_pcount(skb));
1613 
1614 	tp->segs_out += tcp_skb_pcount(skb);
1615 	skb_set_hash_from_sk(skb, sk);
1616 	/* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1617 	skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1618 	skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1619 
1620 	/* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */
1621 
1622 	/* Cleanup our debris for IP stacks */
1623 	memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1624 			       sizeof(struct inet6_skb_parm)));
1625 
1626 	tcp_add_tx_delay(skb, tp);
1627 
1628 	err = INDIRECT_CALL_INET(icsk->icsk_af_ops->queue_xmit,
1629 				 inet6_csk_xmit, ip_queue_xmit,
1630 				 sk, skb, &inet->cork.fl);
1631 
1632 	if (unlikely(err > 0)) {
1633 		tcp_enter_cwr(sk);
1634 		err = net_xmit_eval(err);
1635 	}
1636 	if (!err && oskb) {
1637 		tcp_update_skb_after_send(sk, oskb, prior_wstamp);
1638 		tcp_rate_skb_sent(sk, oskb);
1639 	}
1640 	return err;
1641 }
1642 
tcp_transmit_skb(struct sock * sk,struct sk_buff * skb,int clone_it,gfp_t gfp_mask)1643 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1644 			    gfp_t gfp_mask)
1645 {
1646 	return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1647 				  tcp_sk(sk)->rcv_nxt);
1648 }
1649 
1650 /* This routine just queues the buffer for sending.
1651  *
1652  * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1653  * otherwise socket can stall.
1654  */
tcp_queue_skb(struct sock * sk,struct sk_buff * skb)1655 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1656 {
1657 	struct tcp_sock *tp = tcp_sk(sk);
1658 
1659 	/* Advance write_seq and place onto the write_queue. */
1660 	WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq);
1661 	__skb_header_release(skb);
1662 	psp_enqueue_set_decrypted(sk, skb);
1663 	tcp_add_write_queue_tail(sk, skb);
1664 	sk_wmem_queued_add(sk, skb->truesize);
1665 	sk_mem_charge(sk, skb->truesize);
1666 }
1667 
1668 /* Initialize TSO segments for a packet. */
tcp_set_skb_tso_segs(struct sk_buff * skb,unsigned int mss_now)1669 static int tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1670 {
1671 	int tso_segs;
1672 
1673 	if (skb->len <= mss_now) {
1674 		/* Avoid the costly divide in the normal
1675 		 * non-TSO case.
1676 		 */
1677 		TCP_SKB_CB(skb)->tcp_gso_size = 0;
1678 		tcp_skb_pcount_set(skb, 1);
1679 		return 1;
1680 	}
1681 	TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1682 	tso_segs = DIV_ROUND_UP(skb->len, mss_now);
1683 	tcp_skb_pcount_set(skb, tso_segs);
1684 	return tso_segs;
1685 }
1686 
1687 /* Pcount in the middle of the write queue got changed, we need to do various
1688  * tweaks to fix counters
1689  */
tcp_adjust_pcount(struct sock * sk,const struct sk_buff * skb,int decr)1690 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1691 {
1692 	struct tcp_sock *tp = tcp_sk(sk);
1693 
1694 	tp->packets_out -= decr;
1695 
1696 	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1697 		tp->sacked_out -= decr;
1698 	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1699 		tp->retrans_out -= decr;
1700 	if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1701 		tp->lost_out -= decr;
1702 
1703 	/* Reno case is special. Sigh... */
1704 	if (tcp_is_reno(tp) && decr > 0)
1705 		tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1706 
1707 	tcp_verify_left_out(tp);
1708 }
1709 
tcp_has_tx_tstamp(const struct sk_buff * skb)1710 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1711 {
1712 	return TCP_SKB_CB(skb)->txstamp_ack ||
1713 		(skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1714 }
1715 
tcp_fragment_tstamp(struct sk_buff * skb,struct sk_buff * skb2)1716 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1717 {
1718 	struct skb_shared_info *shinfo = skb_shinfo(skb);
1719 
1720 	if (unlikely(tcp_has_tx_tstamp(skb)) &&
1721 	    !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1722 		struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1723 		u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1724 
1725 		shinfo->tx_flags &= ~tsflags;
1726 		shinfo2->tx_flags |= tsflags;
1727 		swap(shinfo->tskey, shinfo2->tskey);
1728 		TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1729 		TCP_SKB_CB(skb)->txstamp_ack = 0;
1730 	}
1731 }
1732 
tcp_skb_fragment_eor(struct sk_buff * skb,struct sk_buff * skb2)1733 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1734 {
1735 	TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1736 	TCP_SKB_CB(skb)->eor = 0;
1737 }
1738 
1739 /* Insert buff after skb on the write or rtx queue of sk.  */
tcp_insert_write_queue_after(struct sk_buff * skb,struct sk_buff * buff,struct sock * sk,enum tcp_queue tcp_queue)1740 static void tcp_insert_write_queue_after(struct sk_buff *skb,
1741 					 struct sk_buff *buff,
1742 					 struct sock *sk,
1743 					 enum tcp_queue tcp_queue)
1744 {
1745 	if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1746 		__skb_queue_after(&sk->sk_write_queue, skb, buff);
1747 	else
1748 		tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1749 }
1750 
1751 /* Function to create two new TCP segments.  Shrinks the given segment
1752  * to the specified size and appends a new segment with the rest of the
1753  * packet to the list.  This won't be called frequently, I hope.
1754  * Remember, these are still headerless SKBs at this point.
1755  */
tcp_fragment(struct sock * sk,enum tcp_queue tcp_queue,struct sk_buff * skb,u32 len,unsigned int mss_now,gfp_t gfp)1756 int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1757 		 struct sk_buff *skb, u32 len,
1758 		 unsigned int mss_now, gfp_t gfp)
1759 {
1760 	struct tcp_sock *tp = tcp_sk(sk);
1761 	struct sk_buff *buff;
1762 	int old_factor;
1763 	long limit;
1764 	u16 flags;
1765 	int nlen;
1766 
1767 	if (WARN_ON(len > skb->len))
1768 		return -EINVAL;
1769 
1770 	DEBUG_NET_WARN_ON_ONCE(skb_headlen(skb));
1771 
1772 	/* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1773 	 * We need some allowance to not penalize applications setting small
1774 	 * SO_SNDBUF values.
1775 	 * Also allow first and last skb in retransmit queue to be split.
1776 	 */
1777 	limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_LEGACY_MAX_SIZE);
1778 	if (unlikely((sk->sk_wmem_queued >> 1) > limit &&
1779 		     tcp_queue != TCP_FRAG_IN_WRITE_QUEUE &&
1780 		     skb != tcp_rtx_queue_head(sk) &&
1781 		     skb != tcp_rtx_queue_tail(sk))) {
1782 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
1783 		return -ENOMEM;
1784 	}
1785 
1786 	if (skb_unclone_keeptruesize(skb, gfp))
1787 		return -ENOMEM;
1788 
1789 	/* Get a new skb... force flag on. */
1790 	buff = tcp_stream_alloc_skb(sk, gfp, true);
1791 	if (!buff)
1792 		return -ENOMEM; /* We'll just try again later. */
1793 	skb_copy_decrypted(buff, skb);
1794 	mptcp_skb_ext_copy(buff, skb);
1795 
1796 	sk_wmem_queued_add(sk, buff->truesize);
1797 	sk_mem_charge(sk, buff->truesize);
1798 	nlen = skb->len - len;
1799 	buff->truesize += nlen;
1800 	skb->truesize -= nlen;
1801 
1802 	/* Correct the sequence numbers. */
1803 	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1804 	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1805 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1806 
1807 	/* PSH and FIN should only be set in the second packet. */
1808 	flags = TCP_SKB_CB(skb)->tcp_flags;
1809 	TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1810 	TCP_SKB_CB(buff)->tcp_flags = flags;
1811 	TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1812 	tcp_skb_fragment_eor(skb, buff);
1813 
1814 	skb_split(skb, buff, len);
1815 
1816 	skb_set_delivery_time(buff, skb->tstamp, SKB_CLOCK_MONOTONIC);
1817 	tcp_fragment_tstamp(skb, buff);
1818 
1819 	old_factor = tcp_skb_pcount(skb);
1820 
1821 	/* Fix up tso_factor for both original and new SKB.  */
1822 	tcp_set_skb_tso_segs(skb, mss_now);
1823 	tcp_set_skb_tso_segs(buff, mss_now);
1824 
1825 	/* Update delivered info for the new segment */
1826 	TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1827 
1828 	/* If this packet has been sent out already, we must
1829 	 * adjust the various packet counters.
1830 	 */
1831 	if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1832 		int diff = old_factor - tcp_skb_pcount(skb) -
1833 			tcp_skb_pcount(buff);
1834 
1835 		if (diff)
1836 			tcp_adjust_pcount(sk, skb, diff);
1837 	}
1838 
1839 	/* Link BUFF into the send queue. */
1840 	__skb_header_release(buff);
1841 	tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1842 	if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1843 		list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1844 
1845 	return 0;
1846 }
1847 
1848 /* This is similar to __pskb_pull_tail(). The difference is that pulled
1849  * data is not copied, but immediately discarded.
1850  */
__pskb_trim_head(struct sk_buff * skb,int len)1851 static int __pskb_trim_head(struct sk_buff *skb, int len)
1852 {
1853 	struct skb_shared_info *shinfo;
1854 	int i, k, eat;
1855 
1856 	DEBUG_NET_WARN_ON_ONCE(skb_headlen(skb));
1857 	eat = len;
1858 	k = 0;
1859 	shinfo = skb_shinfo(skb);
1860 	for (i = 0; i < shinfo->nr_frags; i++) {
1861 		int size = skb_frag_size(&shinfo->frags[i]);
1862 
1863 		if (size <= eat) {
1864 			skb_frag_unref(skb, i);
1865 			eat -= size;
1866 		} else {
1867 			shinfo->frags[k] = shinfo->frags[i];
1868 			if (eat) {
1869 				skb_frag_off_add(&shinfo->frags[k], eat);
1870 				skb_frag_size_sub(&shinfo->frags[k], eat);
1871 				eat = 0;
1872 			}
1873 			k++;
1874 		}
1875 	}
1876 	shinfo->nr_frags = k;
1877 
1878 	skb->data_len -= len;
1879 	skb->len = skb->data_len;
1880 	return len;
1881 }
1882 
1883 /* Remove acked data from a packet in the transmit queue. */
tcp_trim_head(struct sock * sk,struct sk_buff * skb,u32 len)1884 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1885 {
1886 	u32 delta_truesize;
1887 
1888 	if (skb_unclone_keeptruesize(skb, GFP_ATOMIC))
1889 		return -ENOMEM;
1890 
1891 	delta_truesize = __pskb_trim_head(skb, len);
1892 
1893 	TCP_SKB_CB(skb)->seq += len;
1894 
1895 	skb->truesize	   -= delta_truesize;
1896 	sk_wmem_queued_add(sk, -delta_truesize);
1897 	if (!skb_zcopy_pure(skb))
1898 		sk_mem_uncharge(sk, delta_truesize);
1899 
1900 	/* Any change of skb->len requires recalculation of tso factor. */
1901 	if (tcp_skb_pcount(skb) > 1)
1902 		tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1903 
1904 	return 0;
1905 }
1906 
1907 /* Calculate MSS not accounting any TCP options.  */
__tcp_mtu_to_mss(struct sock * sk,int pmtu)1908 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1909 {
1910 	const struct tcp_sock *tp = tcp_sk(sk);
1911 	const struct inet_connection_sock *icsk = inet_csk(sk);
1912 	int mss_now;
1913 
1914 	/* Calculate base mss without TCP options:
1915 	   It is MMS_S - sizeof(tcphdr) of rfc1122
1916 	 */
1917 	mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1918 
1919 	/* Clamp it (mss_clamp does not include tcp options) */
1920 	if (mss_now > tp->rx_opt.mss_clamp)
1921 		mss_now = tp->rx_opt.mss_clamp;
1922 
1923 	/* Now subtract optional transport overhead */
1924 	mss_now -= icsk->icsk_ext_hdr_len;
1925 
1926 	/* Then reserve room for full set of TCP options and 8 bytes of data */
1927 	mss_now = max(mss_now,
1928 		      READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss));
1929 	return mss_now;
1930 }
1931 
1932 /* Calculate MSS. Not accounting for SACKs here.  */
tcp_mtu_to_mss(struct sock * sk,int pmtu)1933 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1934 {
1935 	/* Subtract TCP options size, not including SACKs */
1936 	return __tcp_mtu_to_mss(sk, pmtu) -
1937 	       (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1938 }
1939 EXPORT_IPV6_MOD(tcp_mtu_to_mss);
1940 
1941 /* Inverse of above */
tcp_mss_to_mtu(struct sock * sk,int mss)1942 int tcp_mss_to_mtu(struct sock *sk, int mss)
1943 {
1944 	const struct tcp_sock *tp = tcp_sk(sk);
1945 	const struct inet_connection_sock *icsk = inet_csk(sk);
1946 
1947 	return mss +
1948 	      tp->tcp_header_len +
1949 	      icsk->icsk_ext_hdr_len +
1950 	      icsk->icsk_af_ops->net_header_len;
1951 }
1952 EXPORT_SYMBOL(tcp_mss_to_mtu);
1953 
1954 /* MTU probing init per socket */
tcp_mtup_init(struct sock * sk)1955 void tcp_mtup_init(struct sock *sk)
1956 {
1957 	struct tcp_sock *tp = tcp_sk(sk);
1958 	struct inet_connection_sock *icsk = inet_csk(sk);
1959 	struct net *net = sock_net(sk);
1960 
1961 	icsk->icsk_mtup.enabled = READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing) > 1;
1962 	icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1963 			       icsk->icsk_af_ops->net_header_len;
1964 	icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, READ_ONCE(net->ipv4.sysctl_tcp_base_mss));
1965 	icsk->icsk_mtup.probe_size = 0;
1966 	if (icsk->icsk_mtup.enabled)
1967 		icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1968 }
1969 
1970 /* This function synchronize snd mss to current pmtu/exthdr set.
1971 
1972    tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1973    for TCP options, but includes only bare TCP header.
1974 
1975    tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1976    It is minimum of user_mss and mss received with SYN.
1977    It also does not include TCP options.
1978 
1979    inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1980 
1981    tp->mss_cache is current effective sending mss, including
1982    all tcp options except for SACKs. It is evaluated,
1983    taking into account current pmtu, but never exceeds
1984    tp->rx_opt.mss_clamp.
1985 
1986    NOTE1. rfc1122 clearly states that advertised MSS
1987    DOES NOT include either tcp or ip options.
1988 
1989    NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1990    are READ ONLY outside this function.		--ANK (980731)
1991  */
tcp_sync_mss(struct sock * sk,u32 pmtu)1992 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1993 {
1994 	struct tcp_sock *tp = tcp_sk(sk);
1995 	struct inet_connection_sock *icsk = inet_csk(sk);
1996 	int mss_now;
1997 
1998 	if (icsk->icsk_mtup.search_high > pmtu)
1999 		icsk->icsk_mtup.search_high = pmtu;
2000 
2001 	mss_now = tcp_mtu_to_mss(sk, pmtu);
2002 	mss_now = tcp_bound_to_half_wnd(tp, mss_now);
2003 
2004 	/* And store cached results */
2005 	icsk->icsk_pmtu_cookie = pmtu;
2006 	if (icsk->icsk_mtup.enabled)
2007 		mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
2008 	tp->mss_cache = mss_now;
2009 
2010 	return mss_now;
2011 }
2012 EXPORT_IPV6_MOD(tcp_sync_mss);
2013 
2014 /* Compute the current effective MSS, taking SACKs and IP options,
2015  * and even PMTU discovery events into account.
2016  */
tcp_current_mss(struct sock * sk)2017 unsigned int tcp_current_mss(struct sock *sk)
2018 {
2019 	const struct tcp_sock *tp = tcp_sk(sk);
2020 	const struct dst_entry *dst = __sk_dst_get(sk);
2021 	u32 mss_now;
2022 	unsigned int header_len;
2023 	struct tcp_out_options opts;
2024 	struct tcp_key key;
2025 
2026 	mss_now = tp->mss_cache;
2027 
2028 	if (dst) {
2029 		u32 mtu = dst_mtu(dst);
2030 		if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
2031 			mss_now = tcp_sync_mss(sk, mtu);
2032 	}
2033 	tcp_get_current_key(sk, &key);
2034 	header_len = tcp_established_options(sk, NULL, &opts, &key) +
2035 		     sizeof(struct tcphdr);
2036 	/* The mss_cache is sized based on tp->tcp_header_len, which assumes
2037 	 * some common options. If this is an odd packet (because we have SACK
2038 	 * blocks etc) then our calculated header_len will be different, and
2039 	 * we have to adjust mss_now correspondingly */
2040 	if (header_len != tp->tcp_header_len) {
2041 		int delta = (int) header_len - tp->tcp_header_len;
2042 		mss_now -= delta;
2043 	}
2044 
2045 	return mss_now;
2046 }
2047 
2048 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
2049  * As additional protections, we do not touch cwnd in retransmission phases,
2050  * and if application hit its sndbuf limit recently.
2051  */
tcp_cwnd_application_limited(struct sock * sk)2052 static void tcp_cwnd_application_limited(struct sock *sk)
2053 {
2054 	struct tcp_sock *tp = tcp_sk(sk);
2055 
2056 	if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
2057 	    sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
2058 		/* Limited by application or receiver window. */
2059 		u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
2060 		u32 win_used = max(tp->snd_cwnd_used, init_win);
2061 		if (win_used < tcp_snd_cwnd(tp)) {
2062 			tp->snd_ssthresh = tcp_current_ssthresh(sk);
2063 			tcp_snd_cwnd_set(tp, (tcp_snd_cwnd(tp) + win_used) >> 1);
2064 		}
2065 		tp->snd_cwnd_used = 0;
2066 	}
2067 	tp->snd_cwnd_stamp = tcp_jiffies32;
2068 }
2069 
tcp_cwnd_validate(struct sock * sk,bool is_cwnd_limited)2070 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
2071 {
2072 	const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2073 	struct tcp_sock *tp = tcp_sk(sk);
2074 
2075 	/* Track the strongest available signal of the degree to which the cwnd
2076 	 * is fully utilized. If cwnd-limited then remember that fact for the
2077 	 * current window. If not cwnd-limited then track the maximum number of
2078 	 * outstanding packets in the current window. (If cwnd-limited then we
2079 	 * chose to not update tp->max_packets_out to avoid an extra else
2080 	 * clause with no functional impact.)
2081 	 */
2082 	if (!before(tp->snd_una, tp->cwnd_usage_seq) ||
2083 	    is_cwnd_limited ||
2084 	    (!tp->is_cwnd_limited &&
2085 	     tp->packets_out > tp->max_packets_out)) {
2086 		tp->is_cwnd_limited = is_cwnd_limited;
2087 		tp->max_packets_out = tp->packets_out;
2088 		tp->cwnd_usage_seq = tp->snd_nxt;
2089 	}
2090 
2091 	if (tcp_is_cwnd_limited(sk)) {
2092 		/* Network is feed fully. */
2093 		tp->snd_cwnd_used = 0;
2094 		tp->snd_cwnd_stamp = tcp_jiffies32;
2095 	} else {
2096 		/* Network starves. */
2097 		if (tp->packets_out > tp->snd_cwnd_used)
2098 			tp->snd_cwnd_used = tp->packets_out;
2099 
2100 		if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle) &&
2101 		    (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
2102 		    !ca_ops->cong_control)
2103 			tcp_cwnd_application_limited(sk);
2104 
2105 		/* The following conditions together indicate the starvation
2106 		 * is caused by insufficient sender buffer:
2107 		 * 1) just sent some data (see tcp_write_xmit)
2108 		 * 2) not cwnd limited (this else condition)
2109 		 * 3) no more data to send (tcp_write_queue_empty())
2110 		 * 4) application is hitting buffer limit (SOCK_NOSPACE)
2111 		 */
2112 		if (tcp_write_queue_empty(sk) && sk->sk_socket &&
2113 		    test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
2114 		    (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
2115 			tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
2116 	}
2117 }
2118 
2119 /* Minshall's variant of the Nagle send check. */
tcp_minshall_check(const struct tcp_sock * tp)2120 static bool tcp_minshall_check(const struct tcp_sock *tp)
2121 {
2122 	return after(tp->snd_sml, tp->snd_una) &&
2123 		!after(tp->snd_sml, tp->snd_nxt);
2124 }
2125 
2126 /* Update snd_sml if this skb is under mss
2127  * Note that a TSO packet might end with a sub-mss segment
2128  * The test is really :
2129  * if ((skb->len % mss) != 0)
2130  *        tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
2131  * But we can avoid doing the divide again given we already have
2132  *  skb_pcount = skb->len / mss_now
2133  */
tcp_minshall_update(struct tcp_sock * tp,unsigned int mss_now,const struct sk_buff * skb)2134 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
2135 				const struct sk_buff *skb)
2136 {
2137 	if (skb->len < tcp_skb_pcount(skb) * mss_now)
2138 		tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
2139 }
2140 
2141 /* Return false, if packet can be sent now without violation Nagle's rules:
2142  * 1. It is full sized. (provided by caller in %partial bool)
2143  * 2. Or it contains FIN. (already checked by caller)
2144  * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
2145  * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
2146  *    With Minshall's modification: all sent small packets are ACKed.
2147  */
tcp_nagle_check(bool partial,const struct tcp_sock * tp,int nonagle)2148 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
2149 			    int nonagle)
2150 {
2151 	return partial &&
2152 		((nonagle & TCP_NAGLE_CORK) ||
2153 		 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
2154 }
2155 
2156 /* Return how many segs we'd like on a TSO packet,
2157  * depending on current pacing rate, and how close the peer is.
2158  *
2159  * Rationale is:
2160  * - For close peers, we rather send bigger packets to reduce
2161  *   cpu costs, because occasional losses will be repaired fast.
2162  * - For long distance/rtt flows, we would like to get ACK clocking
2163  *   with 1 ACK per ms.
2164  *
2165  * Use min_rtt to help adapt TSO burst size, with smaller min_rtt resulting
2166  * in bigger TSO bursts. We we cut the RTT-based allowance in half
2167  * for every 2^9 usec (aka 512 us) of RTT, so that the RTT-based allowance
2168  * is below 1500 bytes after 6 * ~500 usec = 3ms.
2169  */
tcp_tso_autosize(const struct sock * sk,unsigned int mss_now,int min_tso_segs)2170 static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
2171 			    int min_tso_segs)
2172 {
2173 	unsigned long bytes;
2174 	u32 r;
2175 
2176 	bytes = READ_ONCE(sk->sk_pacing_rate) >> READ_ONCE(sk->sk_pacing_shift);
2177 
2178 	r = tcp_min_rtt(tcp_sk(sk)) >> READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_rtt_log);
2179 	if (r < BITS_PER_TYPE(sk->sk_gso_max_size))
2180 		bytes += sk->sk_gso_max_size >> r;
2181 
2182 	bytes = min_t(unsigned long, bytes, sk->sk_gso_max_size);
2183 
2184 	return max_t(u32, bytes / mss_now, min_tso_segs);
2185 }
2186 
2187 /* Return the number of segments we want in the skb we are transmitting.
2188  * See if congestion control module wants to decide; otherwise, autosize.
2189  */
tcp_tso_segs(struct sock * sk,unsigned int mss_now)2190 static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
2191 {
2192 	const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2193 	u32 min_tso, tso_segs;
2194 
2195 	min_tso = ca_ops->min_tso_segs ?
2196 			ca_ops->min_tso_segs(sk) :
2197 			READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs);
2198 
2199 	tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
2200 	return min_t(u32, tso_segs, sk->sk_gso_max_segs);
2201 }
2202 
2203 /* 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)2204 static unsigned int tcp_mss_split_point(const struct sock *sk,
2205 					const struct sk_buff *skb,
2206 					unsigned int mss_now,
2207 					unsigned int max_segs,
2208 					int nonagle)
2209 {
2210 	const struct tcp_sock *tp = tcp_sk(sk);
2211 	u32 partial, needed, window, max_len;
2212 
2213 	window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2214 	max_len = mss_now * max_segs;
2215 
2216 	if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
2217 		return max_len;
2218 
2219 	needed = min(skb->len, window);
2220 
2221 	if (max_len <= needed)
2222 		return max_len;
2223 
2224 	partial = needed % mss_now;
2225 	/* If last segment is not a full MSS, check if Nagle rules allow us
2226 	 * to include this last segment in this skb.
2227 	 * Otherwise, we'll split the skb at last MSS boundary
2228 	 */
2229 	if (tcp_nagle_check(partial != 0, tp, nonagle))
2230 		return needed - partial;
2231 
2232 	return needed;
2233 }
2234 
2235 /* Can at least one segment of SKB be sent right now, according to the
2236  * congestion window rules?  If so, return how many segments are allowed.
2237  */
tcp_cwnd_test(const struct tcp_sock * tp)2238 static u32 tcp_cwnd_test(const struct tcp_sock *tp)
2239 {
2240 	u32 in_flight, cwnd, halfcwnd;
2241 
2242 	in_flight = tcp_packets_in_flight(tp);
2243 	cwnd = tcp_snd_cwnd(tp);
2244 	if (in_flight >= cwnd)
2245 		return 0;
2246 
2247 	/* For better scheduling, ensure we have at least
2248 	 * 2 GSO packets in flight.
2249 	 */
2250 	halfcwnd = max(cwnd >> 1, 1U);
2251 	return min(halfcwnd, cwnd - in_flight);
2252 }
2253 
2254 /* Initialize TSO state of a skb.
2255  * This must be invoked the first time we consider transmitting
2256  * SKB onto the wire.
2257  */
tcp_init_tso_segs(struct sk_buff * skb,unsigned int mss_now)2258 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
2259 {
2260 	int tso_segs = tcp_skb_pcount(skb);
2261 
2262 	if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now))
2263 		return tcp_set_skb_tso_segs(skb, mss_now);
2264 
2265 	return tso_segs;
2266 }
2267 
2268 
2269 /* Return true if the Nagle test allows this packet to be
2270  * sent now.
2271  */
tcp_nagle_test(const struct tcp_sock * tp,const struct sk_buff * skb,unsigned int cur_mss,int nonagle)2272 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
2273 				  unsigned int cur_mss, int nonagle)
2274 {
2275 	/* Nagle rule does not apply to frames, which sit in the middle of the
2276 	 * write_queue (they have no chances to get new data).
2277 	 *
2278 	 * This is implemented in the callers, where they modify the 'nonagle'
2279 	 * argument based upon the location of SKB in the send queue.
2280 	 */
2281 	if (nonagle & TCP_NAGLE_PUSH)
2282 		return true;
2283 
2284 	/* Don't use the nagle rule for urgent data (or for the final FIN). */
2285 	if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
2286 		return true;
2287 
2288 	if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
2289 		return true;
2290 
2291 	return false;
2292 }
2293 
2294 /* 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)2295 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
2296 			     const struct sk_buff *skb,
2297 			     unsigned int cur_mss)
2298 {
2299 	u32 end_seq = TCP_SKB_CB(skb)->end_seq;
2300 
2301 	if (skb->len > cur_mss)
2302 		end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
2303 
2304 	return !after(end_seq, tcp_wnd_end(tp));
2305 }
2306 
2307 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
2308  * which is put after SKB on the list.  It is very much like
2309  * tcp_fragment() except that it may make several kinds of assumptions
2310  * in order to speed up the splitting operation.  In particular, we
2311  * know that all the data is in scatter-gather pages, and that the
2312  * packet has never been sent out before (and thus is not cloned).
2313  */
tso_fragment(struct sock * sk,struct sk_buff * skb,unsigned int len,unsigned int mss_now,gfp_t gfp)2314 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
2315 			unsigned int mss_now, gfp_t gfp)
2316 {
2317 	int nlen = skb->len - len;
2318 	struct sk_buff *buff;
2319 	u16 flags;
2320 
2321 	/* All of a TSO frame must be composed of paged data.  */
2322 	DEBUG_NET_WARN_ON_ONCE(skb->len != skb->data_len);
2323 
2324 	buff = tcp_stream_alloc_skb(sk, gfp, true);
2325 	if (unlikely(!buff))
2326 		return -ENOMEM;
2327 	skb_copy_decrypted(buff, skb);
2328 	mptcp_skb_ext_copy(buff, skb);
2329 
2330 	sk_wmem_queued_add(sk, buff->truesize);
2331 	sk_mem_charge(sk, buff->truesize);
2332 	buff->truesize += nlen;
2333 	skb->truesize -= nlen;
2334 
2335 	/* Correct the sequence numbers. */
2336 	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
2337 	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
2338 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
2339 
2340 	/* PSH and FIN should only be set in the second packet. */
2341 	flags = TCP_SKB_CB(skb)->tcp_flags;
2342 	TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
2343 	TCP_SKB_CB(buff)->tcp_flags = flags;
2344 
2345 	tcp_skb_fragment_eor(skb, buff);
2346 
2347 	skb_split(skb, buff, len);
2348 	tcp_fragment_tstamp(skb, buff);
2349 
2350 	/* Fix up tso_factor for both original and new SKB.  */
2351 	tcp_set_skb_tso_segs(skb, mss_now);
2352 	tcp_set_skb_tso_segs(buff, mss_now);
2353 
2354 	/* Link BUFF into the send queue. */
2355 	__skb_header_release(buff);
2356 	tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE);
2357 
2358 	return 0;
2359 }
2360 
2361 /* Try to defer sending, if possible, in order to minimize the amount
2362  * of TSO splitting we do.  View it as a kind of TSO Nagle test.
2363  *
2364  * This algorithm is from John Heffner.
2365  */
tcp_tso_should_defer(struct sock * sk,struct sk_buff * skb,bool * is_cwnd_limited,bool * is_rwnd_limited,u32 max_segs)2366 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
2367 				 bool *is_cwnd_limited,
2368 				 bool *is_rwnd_limited,
2369 				 u32 max_segs)
2370 {
2371 	const struct inet_connection_sock *icsk = inet_csk(sk);
2372 	u32 send_win, cong_win, limit, in_flight, threshold;
2373 	u64 srtt_in_ns, expected_ack, how_far_is_the_ack;
2374 	struct tcp_sock *tp = tcp_sk(sk);
2375 	struct sk_buff *head;
2376 	int win_divisor;
2377 	s64 delta;
2378 
2379 	if (icsk->icsk_ca_state >= TCP_CA_Recovery)
2380 		goto send_now;
2381 
2382 	/* Avoid bursty behavior by allowing defer
2383 	 * only if the last write was recent (1 ms).
2384 	 * Note that tp->tcp_wstamp_ns can be in the future if we have
2385 	 * packets waiting in a qdisc or device for EDT delivery.
2386 	 */
2387 	delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC;
2388 	if (delta > 0)
2389 		goto send_now;
2390 
2391 	in_flight = tcp_packets_in_flight(tp);
2392 
2393 	BUG_ON(tcp_skb_pcount(skb) <= 1);
2394 	BUG_ON(tcp_snd_cwnd(tp) <= in_flight);
2395 
2396 	send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2397 
2398 	/* From in_flight test above, we know that cwnd > in_flight.  */
2399 	cong_win = (tcp_snd_cwnd(tp) - in_flight) * tp->mss_cache;
2400 
2401 	limit = min(send_win, cong_win);
2402 
2403 	/* If a full-sized TSO skb can be sent, do it. */
2404 	if (limit >= max_segs * tp->mss_cache)
2405 		goto send_now;
2406 
2407 	/* Middle in queue won't get any more data, full sendable already? */
2408 	if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
2409 		goto send_now;
2410 
2411 	win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
2412 	if (win_divisor) {
2413 		u32 chunk = min(tp->snd_wnd, tcp_snd_cwnd(tp) * tp->mss_cache);
2414 
2415 		/* If at least some fraction of a window is available,
2416 		 * just use it.
2417 		 */
2418 		chunk /= win_divisor;
2419 		if (limit >= chunk)
2420 			goto send_now;
2421 	} else {
2422 		/* Different approach, try not to defer past a single
2423 		 * ACK.  Receiver should ACK every other full sized
2424 		 * frame, so if we have space for more than 3 frames
2425 		 * then send now.
2426 		 */
2427 		if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
2428 			goto send_now;
2429 	}
2430 
2431 	/* TODO : use tsorted_sent_queue ? */
2432 	head = tcp_rtx_queue_head(sk);
2433 	if (!head)
2434 		goto send_now;
2435 
2436 	srtt_in_ns = (u64)(NSEC_PER_USEC >> 3) * tp->srtt_us;
2437 	/* When is the ACK expected ? */
2438 	expected_ack = head->tstamp + srtt_in_ns;
2439 	/* How far from now is the ACK expected ? */
2440 	how_far_is_the_ack = expected_ack - tp->tcp_clock_cache;
2441 
2442 	/* If next ACK is likely to come too late,
2443 	 * ie in more than min(1ms, half srtt), do not defer.
2444 	 */
2445 	threshold = min(srtt_in_ns >> 1, NSEC_PER_MSEC);
2446 
2447 	if ((s64)(how_far_is_the_ack - threshold) > 0)
2448 		goto send_now;
2449 
2450 	/* Ok, it looks like it is advisable to defer.
2451 	 * Three cases are tracked :
2452 	 * 1) We are cwnd-limited
2453 	 * 2) We are rwnd-limited
2454 	 * 3) We are application limited.
2455 	 */
2456 	if (cong_win < send_win) {
2457 		if (cong_win <= skb->len) {
2458 			*is_cwnd_limited = true;
2459 			return true;
2460 		}
2461 	} else {
2462 		if (send_win <= skb->len) {
2463 			*is_rwnd_limited = true;
2464 			return true;
2465 		}
2466 	}
2467 
2468 	/* If this packet won't get more data, do not wait. */
2469 	if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) ||
2470 	    TCP_SKB_CB(skb)->eor)
2471 		goto send_now;
2472 
2473 	return true;
2474 
2475 send_now:
2476 	return false;
2477 }
2478 
tcp_mtu_check_reprobe(struct sock * sk)2479 static inline void tcp_mtu_check_reprobe(struct sock *sk)
2480 {
2481 	struct inet_connection_sock *icsk = inet_csk(sk);
2482 	struct tcp_sock *tp = tcp_sk(sk);
2483 	struct net *net = sock_net(sk);
2484 	u32 interval;
2485 	s32 delta;
2486 
2487 	interval = READ_ONCE(net->ipv4.sysctl_tcp_probe_interval);
2488 	delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2489 	if (unlikely(delta >= interval * HZ)) {
2490 		int mss = tcp_current_mss(sk);
2491 
2492 		/* Update current search range */
2493 		icsk->icsk_mtup.probe_size = 0;
2494 		icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2495 			sizeof(struct tcphdr) +
2496 			icsk->icsk_af_ops->net_header_len;
2497 		icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2498 
2499 		/* Update probe time stamp */
2500 		icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2501 	}
2502 }
2503 
tcp_can_coalesce_send_queue_head(struct sock * sk,int len)2504 static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2505 {
2506 	struct sk_buff *skb, *next;
2507 
2508 	skb = tcp_send_head(sk);
2509 	tcp_for_write_queue_from_safe(skb, next, sk) {
2510 		if (len <= skb->len)
2511 			break;
2512 
2513 		if (tcp_has_tx_tstamp(skb) || !tcp_skb_can_collapse(skb, next))
2514 			return false;
2515 
2516 		len -= skb->len;
2517 	}
2518 
2519 	return true;
2520 }
2521 
tcp_clone_payload(struct sock * sk,struct sk_buff * to,int probe_size)2522 static int tcp_clone_payload(struct sock *sk, struct sk_buff *to,
2523 			     int probe_size)
2524 {
2525 	skb_frag_t *lastfrag = NULL, *fragto = skb_shinfo(to)->frags;
2526 	int i, todo, len = 0, nr_frags = 0;
2527 	const struct sk_buff *skb;
2528 
2529 	if (!sk_wmem_schedule(sk, to->truesize + probe_size))
2530 		return -ENOMEM;
2531 
2532 	skb_queue_walk(&sk->sk_write_queue, skb) {
2533 		const skb_frag_t *fragfrom = skb_shinfo(skb)->frags;
2534 
2535 		if (skb_headlen(skb))
2536 			return -EINVAL;
2537 
2538 		for (i = 0; i < skb_shinfo(skb)->nr_frags; i++, fragfrom++) {
2539 			if (len >= probe_size)
2540 				goto commit;
2541 			todo = min_t(int, skb_frag_size(fragfrom),
2542 				     probe_size - len);
2543 			len += todo;
2544 			if (lastfrag &&
2545 			    skb_frag_page(fragfrom) == skb_frag_page(lastfrag) &&
2546 			    skb_frag_off(fragfrom) == skb_frag_off(lastfrag) +
2547 						      skb_frag_size(lastfrag)) {
2548 				skb_frag_size_add(lastfrag, todo);
2549 				continue;
2550 			}
2551 			if (unlikely(nr_frags == MAX_SKB_FRAGS))
2552 				return -E2BIG;
2553 			skb_frag_page_copy(fragto, fragfrom);
2554 			skb_frag_off_copy(fragto, fragfrom);
2555 			skb_frag_size_set(fragto, todo);
2556 			nr_frags++;
2557 			lastfrag = fragto++;
2558 		}
2559 	}
2560 commit:
2561 	WARN_ON_ONCE(len != probe_size);
2562 	for (i = 0; i < nr_frags; i++)
2563 		skb_frag_ref(to, i);
2564 
2565 	skb_shinfo(to)->nr_frags = nr_frags;
2566 	to->truesize += probe_size;
2567 	to->len += probe_size;
2568 	to->data_len += probe_size;
2569 	__skb_header_release(to);
2570 	return 0;
2571 }
2572 
2573 /* tcp_mtu_probe() and tcp_grow_skb() can both eat an skb (src) if
2574  * all its payload was moved to another one (dst).
2575  * Make sure to transfer tcp_flags, eor, and tstamp.
2576  */
tcp_eat_one_skb(struct sock * sk,struct sk_buff * dst,struct sk_buff * src)2577 static void tcp_eat_one_skb(struct sock *sk,
2578 			    struct sk_buff *dst,
2579 			    struct sk_buff *src)
2580 {
2581 	TCP_SKB_CB(dst)->tcp_flags |= TCP_SKB_CB(src)->tcp_flags;
2582 	TCP_SKB_CB(dst)->eor = TCP_SKB_CB(src)->eor;
2583 	tcp_skb_collapse_tstamp(dst, src);
2584 	tcp_unlink_write_queue(src, sk);
2585 	tcp_wmem_free_skb(sk, src);
2586 }
2587 
2588 /* Create a new MTU probe if we are ready.
2589  * MTU probe is regularly attempting to increase the path MTU by
2590  * deliberately sending larger packets.  This discovers routing
2591  * changes resulting in larger path MTUs.
2592  *
2593  * Returns 0 if we should wait to probe (no cwnd available),
2594  *         1 if a probe was sent,
2595  *         -1 otherwise
2596  */
tcp_mtu_probe(struct sock * sk)2597 static int tcp_mtu_probe(struct sock *sk)
2598 {
2599 	struct inet_connection_sock *icsk = inet_csk(sk);
2600 	struct tcp_sock *tp = tcp_sk(sk);
2601 	struct sk_buff *skb, *nskb, *next;
2602 	struct net *net = sock_net(sk);
2603 	int probe_size;
2604 	int size_needed;
2605 	int copy, len;
2606 	int mss_now;
2607 	int interval;
2608 
2609 	/* Not currently probing/verifying,
2610 	 * not in recovery,
2611 	 * have enough cwnd, and
2612 	 * not SACKing (the variable headers throw things off)
2613 	 */
2614 	if (likely(!icsk->icsk_mtup.enabled ||
2615 		   icsk->icsk_mtup.probe_size ||
2616 		   inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2617 		   tcp_snd_cwnd(tp) < 11 ||
2618 		   tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2619 		return -1;
2620 
2621 	/* Use binary search for probe_size between tcp_mss_base,
2622 	 * and current mss_clamp. if (search_high - search_low)
2623 	 * smaller than a threshold, backoff from probing.
2624 	 */
2625 	mss_now = tcp_current_mss(sk);
2626 	probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2627 				    icsk->icsk_mtup.search_low) >> 1);
2628 	size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2629 	interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2630 	/* When misfortune happens, we are reprobing actively,
2631 	 * and then reprobe timer has expired. We stick with current
2632 	 * probing process by not resetting search range to its orignal.
2633 	 */
2634 	if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2635 	    interval < READ_ONCE(net->ipv4.sysctl_tcp_probe_threshold)) {
2636 		/* Check whether enough time has elaplased for
2637 		 * another round of probing.
2638 		 */
2639 		tcp_mtu_check_reprobe(sk);
2640 		return -1;
2641 	}
2642 
2643 	/* Have enough data in the send queue to probe? */
2644 	if (tp->write_seq - tp->snd_nxt < size_needed)
2645 		return -1;
2646 
2647 	if (tp->snd_wnd < size_needed)
2648 		return -1;
2649 	if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2650 		return 0;
2651 
2652 	/* Do we need to wait to drain cwnd? With none in flight, don't stall */
2653 	if (tcp_packets_in_flight(tp) + 2 > tcp_snd_cwnd(tp)) {
2654 		if (!tcp_packets_in_flight(tp))
2655 			return -1;
2656 		else
2657 			return 0;
2658 	}
2659 
2660 	if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2661 		return -1;
2662 
2663 	/* We're allowed to probe.  Build it now. */
2664 	nskb = tcp_stream_alloc_skb(sk, GFP_ATOMIC, false);
2665 	if (!nskb)
2666 		return -1;
2667 
2668 	/* build the payload, and be prepared to abort if this fails. */
2669 	if (tcp_clone_payload(sk, nskb, probe_size)) {
2670 		tcp_skb_tsorted_anchor_cleanup(nskb);
2671 		consume_skb(nskb);
2672 		return -1;
2673 	}
2674 	sk_wmem_queued_add(sk, nskb->truesize);
2675 	sk_mem_charge(sk, nskb->truesize);
2676 
2677 	skb = tcp_send_head(sk);
2678 	skb_copy_decrypted(nskb, skb);
2679 	mptcp_skb_ext_copy(nskb, skb);
2680 
2681 	TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2682 	TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2683 	TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2684 
2685 	tcp_insert_write_queue_before(nskb, skb, sk);
2686 	tcp_highest_sack_replace(sk, skb, nskb);
2687 
2688 	len = 0;
2689 	tcp_for_write_queue_from_safe(skb, next, sk) {
2690 		copy = min_t(int, skb->len, probe_size - len);
2691 
2692 		if (skb->len <= copy) {
2693 			tcp_eat_one_skb(sk, nskb, skb);
2694 		} else {
2695 			TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2696 						   ~(TCPHDR_FIN|TCPHDR_PSH);
2697 			__pskb_trim_head(skb, copy);
2698 			tcp_set_skb_tso_segs(skb, mss_now);
2699 			TCP_SKB_CB(skb)->seq += copy;
2700 		}
2701 
2702 		len += copy;
2703 
2704 		if (len >= probe_size)
2705 			break;
2706 	}
2707 	tcp_init_tso_segs(nskb, nskb->len);
2708 
2709 	/* We're ready to send.  If this fails, the probe will
2710 	 * be resegmented into mss-sized pieces by tcp_write_xmit().
2711 	 */
2712 	if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2713 		/* Decrement cwnd here because we are sending
2714 		 * effectively two packets. */
2715 		tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) - 1);
2716 		tcp_event_new_data_sent(sk, nskb);
2717 
2718 		icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2719 		tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2720 		tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2721 
2722 		return 1;
2723 	}
2724 
2725 	return -1;
2726 }
2727 
tcp_pacing_check(struct sock * sk)2728 static bool tcp_pacing_check(struct sock *sk)
2729 {
2730 	struct tcp_sock *tp = tcp_sk(sk);
2731 
2732 	if (!tcp_needs_internal_pacing(sk))
2733 		return false;
2734 
2735 	if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache)
2736 		return false;
2737 
2738 	if (!hrtimer_is_queued(&tp->pacing_timer)) {
2739 		hrtimer_start(&tp->pacing_timer,
2740 			      ns_to_ktime(tp->tcp_wstamp_ns),
2741 			      HRTIMER_MODE_ABS_PINNED_SOFT);
2742 		sock_hold(sk);
2743 	}
2744 	return true;
2745 }
2746 
tcp_rtx_queue_empty_or_single_skb(const struct sock * sk)2747 static bool tcp_rtx_queue_empty_or_single_skb(const struct sock *sk)
2748 {
2749 	const struct rb_node *node = sk->tcp_rtx_queue.rb_node;
2750 
2751 	/* No skb in the rtx queue. */
2752 	if (!node)
2753 		return true;
2754 
2755 	/* Only one skb in rtx queue. */
2756 	return !node->rb_left && !node->rb_right;
2757 }
2758 
2759 /* TCP Small Queues :
2760  * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2761  * (These limits are doubled for retransmits)
2762  * This allows for :
2763  *  - better RTT estimation and ACK scheduling
2764  *  - faster recovery
2765  *  - high rates
2766  * Alas, some drivers / subsystems require a fair amount
2767  * of queued bytes to ensure line rate.
2768  * One example is wifi aggregation (802.11 AMPDU)
2769  */
tcp_small_queue_check(struct sock * sk,const struct sk_buff * skb,unsigned int factor)2770 static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2771 				  unsigned int factor)
2772 {
2773 	unsigned long limit;
2774 
2775 	limit = max_t(unsigned long,
2776 		      2 * skb->truesize,
2777 		      READ_ONCE(sk->sk_pacing_rate) >> READ_ONCE(sk->sk_pacing_shift));
2778 	limit = min_t(unsigned long, limit,
2779 		      READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes));
2780 	limit <<= factor;
2781 
2782 	if (static_branch_unlikely(&tcp_tx_delay_enabled) &&
2783 	    tcp_sk(sk)->tcp_tx_delay) {
2784 		u64 extra_bytes = (u64)READ_ONCE(sk->sk_pacing_rate) *
2785 				  tcp_sk(sk)->tcp_tx_delay;
2786 
2787 		/* TSQ is based on skb truesize sum (sk_wmem_alloc), so we
2788 		 * approximate our needs assuming an ~100% skb->truesize overhead.
2789 		 * USEC_PER_SEC is approximated by 2^20.
2790 		 * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift.
2791 		 */
2792 		extra_bytes >>= (20 - 1);
2793 		limit += extra_bytes;
2794 	}
2795 	if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2796 		/* Always send skb if rtx queue is empty or has one skb.
2797 		 * No need to wait for TX completion to call us back,
2798 		 * after softirq schedule.
2799 		 * This helps when TX completions are delayed too much.
2800 		 */
2801 		if (tcp_rtx_queue_empty_or_single_skb(sk))
2802 			return false;
2803 
2804 		set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2805 		/* It is possible TX completion already happened
2806 		 * before we set TSQ_THROTTLED, so we must
2807 		 * test again the condition.
2808 		 */
2809 		smp_mb__after_atomic();
2810 		if (refcount_read(&sk->sk_wmem_alloc) > limit)
2811 			return true;
2812 	}
2813 	return false;
2814 }
2815 
tcp_chrono_set(struct tcp_sock * tp,const enum tcp_chrono new)2816 static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2817 {
2818 	const u32 now = tcp_jiffies32;
2819 	enum tcp_chrono old = tp->chrono_type;
2820 
2821 	if (old > TCP_CHRONO_UNSPEC)
2822 		tp->chrono_stat[old - 1] += now - tp->chrono_start;
2823 	tp->chrono_start = now;
2824 	tp->chrono_type = new;
2825 }
2826 
tcp_chrono_start(struct sock * sk,const enum tcp_chrono type)2827 void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2828 {
2829 	struct tcp_sock *tp = tcp_sk(sk);
2830 
2831 	/* If there are multiple conditions worthy of tracking in a
2832 	 * chronograph then the highest priority enum takes precedence
2833 	 * over the other conditions. So that if something "more interesting"
2834 	 * starts happening, stop the previous chrono and start a new one.
2835 	 */
2836 	if (type > tp->chrono_type)
2837 		tcp_chrono_set(tp, type);
2838 }
2839 
tcp_chrono_stop(struct sock * sk,const enum tcp_chrono type)2840 void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2841 {
2842 	struct tcp_sock *tp = tcp_sk(sk);
2843 
2844 
2845 	/* There are multiple conditions worthy of tracking in a
2846 	 * chronograph, so that the highest priority enum takes
2847 	 * precedence over the other conditions (see tcp_chrono_start).
2848 	 * If a condition stops, we only stop chrono tracking if
2849 	 * it's the "most interesting" or current chrono we are
2850 	 * tracking and starts busy chrono if we have pending data.
2851 	 */
2852 	if (tcp_rtx_and_write_queues_empty(sk))
2853 		tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2854 	else if (type == tp->chrono_type)
2855 		tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2856 }
2857 
2858 /* First skb in the write queue is smaller than ideal packet size.
2859  * Check if we can move payload from the second skb in the queue.
2860  */
tcp_grow_skb(struct sock * sk,struct sk_buff * skb,int amount)2861 static void tcp_grow_skb(struct sock *sk, struct sk_buff *skb, int amount)
2862 {
2863 	struct sk_buff *next_skb = skb->next;
2864 	unsigned int nlen;
2865 
2866 	if (tcp_skb_is_last(sk, skb))
2867 		return;
2868 
2869 	if (!tcp_skb_can_collapse(skb, next_skb))
2870 		return;
2871 
2872 	nlen = min_t(u32, amount, next_skb->len);
2873 	if (!nlen || !skb_shift(skb, next_skb, nlen))
2874 		return;
2875 
2876 	TCP_SKB_CB(skb)->end_seq += nlen;
2877 	TCP_SKB_CB(next_skb)->seq += nlen;
2878 
2879 	if (!next_skb->len) {
2880 		/* In case FIN is set, we need to update end_seq */
2881 		TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2882 
2883 		tcp_eat_one_skb(sk, skb, next_skb);
2884 	}
2885 }
2886 
2887 /* This routine writes packets to the network.  It advances the
2888  * send_head.  This happens as incoming acks open up the remote
2889  * window for us.
2890  *
2891  * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2892  * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2893  * account rare use of URG, this is not a big flaw.
2894  *
2895  * Send at most one packet when push_one > 0. Temporarily ignore
2896  * cwnd limit to force at most one packet out when push_one == 2.
2897 
2898  * Returns true, if no segments are in flight and we have queued segments,
2899  * but cannot send anything now because of SWS or another problem.
2900  */
tcp_write_xmit(struct sock * sk,unsigned int mss_now,int nonagle,int push_one,gfp_t gfp)2901 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2902 			   int push_one, gfp_t gfp)
2903 {
2904 	struct tcp_sock *tp = tcp_sk(sk);
2905 	struct sk_buff *skb;
2906 	unsigned int tso_segs, sent_pkts;
2907 	u32 cwnd_quota, max_segs;
2908 	int result;
2909 	bool is_cwnd_limited = false, is_rwnd_limited = false;
2910 
2911 	sent_pkts = 0;
2912 
2913 	tcp_mstamp_refresh(tp);
2914 
2915 	/* AccECN option beacon depends on mstamp, it may change mss */
2916 	if (tcp_ecn_mode_accecn(tp) && tcp_accecn_option_beacon_check(sk))
2917 		mss_now = tcp_current_mss(sk);
2918 
2919 	if (!push_one) {
2920 		/* Do MTU probing. */
2921 		result = tcp_mtu_probe(sk);
2922 		if (!result) {
2923 			return false;
2924 		} else if (result > 0) {
2925 			sent_pkts = 1;
2926 		}
2927 	}
2928 
2929 	max_segs = tcp_tso_segs(sk, mss_now);
2930 	while ((skb = tcp_send_head(sk))) {
2931 		unsigned int limit;
2932 		int missing_bytes;
2933 
2934 		if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2935 			/* "skb_mstamp_ns" is used as a start point for the retransmit timer */
2936 			tp->tcp_wstamp_ns = tp->tcp_clock_cache;
2937 			skb_set_delivery_time(skb, tp->tcp_wstamp_ns, SKB_CLOCK_MONOTONIC);
2938 			list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
2939 			tcp_init_tso_segs(skb, mss_now);
2940 			goto repair; /* Skip network transmission */
2941 		}
2942 
2943 		if (tcp_pacing_check(sk))
2944 			break;
2945 
2946 		cwnd_quota = tcp_cwnd_test(tp);
2947 		if (!cwnd_quota) {
2948 			if (push_one == 2)
2949 				/* Force out a loss probe pkt. */
2950 				cwnd_quota = 1;
2951 			else
2952 				break;
2953 		}
2954 		cwnd_quota = min(cwnd_quota, max_segs);
2955 		missing_bytes = cwnd_quota * mss_now - skb->len;
2956 		if (missing_bytes > 0)
2957 			tcp_grow_skb(sk, skb, missing_bytes);
2958 
2959 		tso_segs = tcp_set_skb_tso_segs(skb, mss_now);
2960 
2961 		if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2962 			is_rwnd_limited = true;
2963 			break;
2964 		}
2965 
2966 		if (tso_segs == 1) {
2967 			if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2968 						     (tcp_skb_is_last(sk, skb) ?
2969 						      nonagle : TCP_NAGLE_PUSH))))
2970 				break;
2971 		} else {
2972 			if (!push_one &&
2973 			    tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2974 						 &is_rwnd_limited, max_segs))
2975 				break;
2976 		}
2977 
2978 		limit = mss_now;
2979 		if (tso_segs > 1 && !tcp_urg_mode(tp))
2980 			limit = tcp_mss_split_point(sk, skb, mss_now,
2981 						    cwnd_quota,
2982 						    nonagle);
2983 
2984 		if (skb->len > limit &&
2985 		    unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2986 			break;
2987 
2988 		if (tcp_small_queue_check(sk, skb, 0))
2989 			break;
2990 
2991 		/* Argh, we hit an empty skb(), presumably a thread
2992 		 * is sleeping in sendmsg()/sk_stream_wait_memory().
2993 		 * We do not want to send a pure-ack packet and have
2994 		 * a strange looking rtx queue with empty packet(s).
2995 		 */
2996 		if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq)
2997 			break;
2998 
2999 		if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
3000 			break;
3001 
3002 repair:
3003 		/* Advance the send_head.  This one is sent out.
3004 		 * This call will increment packets_out.
3005 		 */
3006 		tcp_event_new_data_sent(sk, skb);
3007 
3008 		tcp_minshall_update(tp, mss_now, skb);
3009 		sent_pkts += tcp_skb_pcount(skb);
3010 
3011 		if (push_one)
3012 			break;
3013 	}
3014 
3015 	if (is_rwnd_limited)
3016 		tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
3017 	else
3018 		tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
3019 
3020 	is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp));
3021 	if (likely(sent_pkts || is_cwnd_limited))
3022 		tcp_cwnd_validate(sk, is_cwnd_limited);
3023 
3024 	if (likely(sent_pkts)) {
3025 		if (tcp_in_cwnd_reduction(sk))
3026 			tp->prr_out += sent_pkts;
3027 
3028 		/* Send one loss probe per tail loss episode. */
3029 		if (push_one != 2)
3030 			tcp_schedule_loss_probe(sk, false);
3031 		return false;
3032 	}
3033 	return !tp->packets_out && !tcp_write_queue_empty(sk);
3034 }
3035 
tcp_schedule_loss_probe(struct sock * sk,bool advancing_rto)3036 bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
3037 {
3038 	struct inet_connection_sock *icsk = inet_csk(sk);
3039 	struct tcp_sock *tp = tcp_sk(sk);
3040 	u32 timeout, timeout_us, rto_delta_us;
3041 	int early_retrans;
3042 
3043 	/* Don't do any loss probe on a Fast Open connection before 3WHS
3044 	 * finishes.
3045 	 */
3046 	if (rcu_access_pointer(tp->fastopen_rsk))
3047 		return false;
3048 
3049 	early_retrans = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_early_retrans);
3050 	/* Schedule a loss probe in 2*RTT for SACK capable connections
3051 	 * not in loss recovery, that are either limited by cwnd or application.
3052 	 */
3053 	if ((early_retrans != 3 && early_retrans != 4) ||
3054 	    !tp->packets_out || !tcp_is_sack(tp) ||
3055 	    (icsk->icsk_ca_state != TCP_CA_Open &&
3056 	     icsk->icsk_ca_state != TCP_CA_CWR))
3057 		return false;
3058 
3059 	/* Probe timeout is 2*rtt. Add minimum RTO to account
3060 	 * for delayed ack when there's one outstanding packet. If no RTT
3061 	 * sample is available then probe after TCP_TIMEOUT_INIT.
3062 	 */
3063 	if (tp->srtt_us) {
3064 		timeout_us = tp->srtt_us >> 2;
3065 		if (tp->packets_out == 1)
3066 			timeout_us += tcp_rto_min_us(sk);
3067 		else
3068 			timeout_us += TCP_TIMEOUT_MIN_US;
3069 		timeout = usecs_to_jiffies(timeout_us);
3070 	} else {
3071 		timeout = TCP_TIMEOUT_INIT;
3072 	}
3073 
3074 	/* If the RTO formula yields an earlier time, then use that time. */
3075 	rto_delta_us = advancing_rto ?
3076 			jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
3077 			tcp_rto_delta_us(sk);  /* How far in future is RTO? */
3078 	if (rto_delta_us > 0)
3079 		timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
3080 
3081 	tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, true);
3082 	return true;
3083 }
3084 
3085 /* Thanks to skb fast clones, we can detect if a prior transmit of
3086  * a packet is still in a qdisc or driver queue.
3087  * In this case, there is very little point doing a retransmit !
3088  */
skb_still_in_host_queue(struct sock * sk,const struct sk_buff * skb)3089 static bool skb_still_in_host_queue(struct sock *sk,
3090 				    const struct sk_buff *skb)
3091 {
3092 	if (unlikely(skb_fclone_busy(sk, skb))) {
3093 		set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
3094 		smp_mb__after_atomic();
3095 		if (skb_fclone_busy(sk, skb)) {
3096 			NET_INC_STATS(sock_net(sk),
3097 				      LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
3098 			return true;
3099 		}
3100 	}
3101 	return false;
3102 }
3103 
3104 /* When probe timeout (PTO) fires, try send a new segment if possible, else
3105  * retransmit the last segment.
3106  */
tcp_send_loss_probe(struct sock * sk)3107 void tcp_send_loss_probe(struct sock *sk)
3108 {
3109 	struct tcp_sock *tp = tcp_sk(sk);
3110 	struct sk_buff *skb;
3111 	int pcount;
3112 	int mss = tcp_current_mss(sk);
3113 
3114 	/* At most one outstanding TLP */
3115 	if (tp->tlp_high_seq)
3116 		goto rearm_timer;
3117 
3118 	tp->tlp_retrans = 0;
3119 	skb = tcp_send_head(sk);
3120 	if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
3121 		pcount = tp->packets_out;
3122 		tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
3123 		if (tp->packets_out > pcount)
3124 			goto probe_sent;
3125 		goto rearm_timer;
3126 	}
3127 	skb = skb_rb_last(&sk->tcp_rtx_queue);
3128 	if (unlikely(!skb)) {
3129 		tcp_warn_once(sk, tp->packets_out, "invalid inflight: ");
3130 		smp_store_release(&inet_csk(sk)->icsk_pending, 0);
3131 		return;
3132 	}
3133 
3134 	if (skb_still_in_host_queue(sk, skb))
3135 		goto rearm_timer;
3136 
3137 	pcount = tcp_skb_pcount(skb);
3138 	if (WARN_ON(!pcount))
3139 		goto rearm_timer;
3140 
3141 	if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
3142 		if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
3143 					  (pcount - 1) * mss, mss,
3144 					  GFP_ATOMIC)))
3145 			goto rearm_timer;
3146 		skb = skb_rb_next(skb);
3147 	}
3148 
3149 	if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
3150 		goto rearm_timer;
3151 
3152 	if (__tcp_retransmit_skb(sk, skb, 1))
3153 		goto rearm_timer;
3154 
3155 	tp->tlp_retrans = 1;
3156 
3157 probe_sent:
3158 	/* Record snd_nxt for loss detection. */
3159 	tp->tlp_high_seq = tp->snd_nxt;
3160 
3161 	NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
3162 	/* Reset s.t. tcp_rearm_rto will restart timer from now */
3163 	smp_store_release(&inet_csk(sk)->icsk_pending, 0);
3164 rearm_timer:
3165 	tcp_rearm_rto(sk);
3166 }
3167 
3168 /* Push out any pending frames which were held back due to
3169  * TCP_CORK or attempt at coalescing tiny packets.
3170  * The socket must be locked by the caller.
3171  */
__tcp_push_pending_frames(struct sock * sk,unsigned int cur_mss,int nonagle)3172 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
3173 			       int nonagle)
3174 {
3175 	/* If we are closed, the bytes will have to remain here.
3176 	 * In time closedown will finish, we empty the write queue and
3177 	 * all will be happy.
3178 	 */
3179 	if (unlikely(sk->sk_state == TCP_CLOSE))
3180 		return;
3181 
3182 	if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
3183 			   sk_gfp_mask(sk, GFP_ATOMIC)))
3184 		tcp_check_probe_timer(sk);
3185 }
3186 
3187 /* Send _single_ skb sitting at the send head. This function requires
3188  * true push pending frames to setup probe timer etc.
3189  */
tcp_push_one(struct sock * sk,unsigned int mss_now)3190 void tcp_push_one(struct sock *sk, unsigned int mss_now)
3191 {
3192 	struct sk_buff *skb = tcp_send_head(sk);
3193 
3194 	BUG_ON(!skb || skb->len < mss_now);
3195 
3196 	tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
3197 }
3198 
3199 /* This function returns the amount that we can raise the
3200  * usable window based on the following constraints
3201  *
3202  * 1. The window can never be shrunk once it is offered (RFC 793)
3203  * 2. We limit memory per socket
3204  *
3205  * RFC 1122:
3206  * "the suggested [SWS] avoidance algorithm for the receiver is to keep
3207  *  RECV.NEXT + RCV.WIN fixed until:
3208  *  RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
3209  *
3210  * i.e. don't raise the right edge of the window until you can raise
3211  * it at least MSS bytes.
3212  *
3213  * Unfortunately, the recommended algorithm breaks header prediction,
3214  * since header prediction assumes th->window stays fixed.
3215  *
3216  * Strictly speaking, keeping th->window fixed violates the receiver
3217  * side SWS prevention criteria. The problem is that under this rule
3218  * a stream of single byte packets will cause the right side of the
3219  * window to always advance by a single byte.
3220  *
3221  * Of course, if the sender implements sender side SWS prevention
3222  * then this will not be a problem.
3223  *
3224  * BSD seems to make the following compromise:
3225  *
3226  *	If the free space is less than the 1/4 of the maximum
3227  *	space available and the free space is less than 1/2 mss,
3228  *	then set the window to 0.
3229  *	[ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
3230  *	Otherwise, just prevent the window from shrinking
3231  *	and from being larger than the largest representable value.
3232  *
3233  * This prevents incremental opening of the window in the regime
3234  * where TCP is limited by the speed of the reader side taking
3235  * data out of the TCP receive queue. It does nothing about
3236  * those cases where the window is constrained on the sender side
3237  * because the pipeline is full.
3238  *
3239  * BSD also seems to "accidentally" limit itself to windows that are a
3240  * multiple of MSS, at least until the free space gets quite small.
3241  * This would appear to be a side effect of the mbuf implementation.
3242  * Combining these two algorithms results in the observed behavior
3243  * of having a fixed window size at almost all times.
3244  *
3245  * Below we obtain similar behavior by forcing the offered window to
3246  * a multiple of the mss when it is feasible to do so.
3247  *
3248  * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
3249  * Regular options like TIMESTAMP are taken into account.
3250  */
__tcp_select_window(struct sock * sk)3251 u32 __tcp_select_window(struct sock *sk)
3252 {
3253 	struct inet_connection_sock *icsk = inet_csk(sk);
3254 	struct tcp_sock *tp = tcp_sk(sk);
3255 	struct net *net = sock_net(sk);
3256 	/* MSS for the peer's data.  Previous versions used mss_clamp
3257 	 * here.  I don't know if the value based on our guesses
3258 	 * of peer's MSS is better for the performance.  It's more correct
3259 	 * but may be worse for the performance because of rcv_mss
3260 	 * fluctuations.  --SAW  1998/11/1
3261 	 */
3262 	int mss = icsk->icsk_ack.rcv_mss;
3263 	int free_space = tcp_space(sk);
3264 	int allowed_space = tcp_full_space(sk);
3265 	int full_space, window;
3266 
3267 	if (sk_is_mptcp(sk))
3268 		mptcp_space(sk, &free_space, &allowed_space);
3269 
3270 	full_space = min_t(int, tp->window_clamp, allowed_space);
3271 
3272 	if (unlikely(mss > full_space)) {
3273 		mss = full_space;
3274 		if (mss <= 0)
3275 			return 0;
3276 	}
3277 
3278 	/* Only allow window shrink if the sysctl is enabled and we have
3279 	 * a non-zero scaling factor in effect.
3280 	 */
3281 	if (READ_ONCE(net->ipv4.sysctl_tcp_shrink_window) && tp->rx_opt.rcv_wscale)
3282 		goto shrink_window_allowed;
3283 
3284 	/* do not allow window to shrink */
3285 
3286 	if (free_space < (full_space >> 1)) {
3287 		icsk->icsk_ack.quick = 0;
3288 
3289 		if (tcp_under_memory_pressure(sk))
3290 			tcp_adjust_rcv_ssthresh(sk);
3291 
3292 		/* free_space might become our new window, make sure we don't
3293 		 * increase it due to wscale.
3294 		 */
3295 		free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
3296 
3297 		/* if free space is less than mss estimate, or is below 1/16th
3298 		 * of the maximum allowed, try to move to zero-window, else
3299 		 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
3300 		 * new incoming data is dropped due to memory limits.
3301 		 * With large window, mss test triggers way too late in order
3302 		 * to announce zero window in time before rmem limit kicks in.
3303 		 */
3304 		if (free_space < (allowed_space >> 4) || free_space < mss)
3305 			return 0;
3306 	}
3307 
3308 	if (free_space > tp->rcv_ssthresh)
3309 		free_space = tp->rcv_ssthresh;
3310 
3311 	/* Don't do rounding if we are using window scaling, since the
3312 	 * scaled window will not line up with the MSS boundary anyway.
3313 	 */
3314 	if (tp->rx_opt.rcv_wscale) {
3315 		window = free_space;
3316 
3317 		/* Advertise enough space so that it won't get scaled away.
3318 		 * Import case: prevent zero window announcement if
3319 		 * 1<<rcv_wscale > mss.
3320 		 */
3321 		window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
3322 	} else {
3323 		window = tp->rcv_wnd;
3324 		/* Get the largest window that is a nice multiple of mss.
3325 		 * Window clamp already applied above.
3326 		 * If our current window offering is within 1 mss of the
3327 		 * free space we just keep it. This prevents the divide
3328 		 * and multiply from happening most of the time.
3329 		 * We also don't do any window rounding when the free space
3330 		 * is too small.
3331 		 */
3332 		if (window <= free_space - mss || window > free_space)
3333 			window = rounddown(free_space, mss);
3334 		else if (mss == full_space &&
3335 			 free_space > window + (full_space >> 1))
3336 			window = free_space;
3337 	}
3338 
3339 	return window;
3340 
3341 shrink_window_allowed:
3342 	/* new window should always be an exact multiple of scaling factor */
3343 	free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
3344 
3345 	if (free_space < (full_space >> 1)) {
3346 		icsk->icsk_ack.quick = 0;
3347 
3348 		if (tcp_under_memory_pressure(sk))
3349 			tcp_adjust_rcv_ssthresh(sk);
3350 
3351 		/* if free space is too low, return a zero window */
3352 		if (free_space < (allowed_space >> 4) || free_space < mss ||
3353 			free_space < (1 << tp->rx_opt.rcv_wscale))
3354 			return 0;
3355 	}
3356 
3357 	if (free_space > tp->rcv_ssthresh) {
3358 		free_space = tp->rcv_ssthresh;
3359 		/* new window should always be an exact multiple of scaling factor
3360 		 *
3361 		 * For this case, we ALIGN "up" (increase free_space) because
3362 		 * we know free_space is not zero here, it has been reduced from
3363 		 * the memory-based limit, and rcv_ssthresh is not a hard limit
3364 		 * (unlike sk_rcvbuf).
3365 		 */
3366 		free_space = ALIGN(free_space, (1 << tp->rx_opt.rcv_wscale));
3367 	}
3368 
3369 	return free_space;
3370 }
3371 
tcp_skb_collapse_tstamp(struct sk_buff * skb,const struct sk_buff * next_skb)3372 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
3373 			     const struct sk_buff *next_skb)
3374 {
3375 	if (unlikely(tcp_has_tx_tstamp(next_skb))) {
3376 		const struct skb_shared_info *next_shinfo =
3377 			skb_shinfo(next_skb);
3378 		struct skb_shared_info *shinfo = skb_shinfo(skb);
3379 
3380 		shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
3381 		shinfo->tskey = next_shinfo->tskey;
3382 		TCP_SKB_CB(skb)->txstamp_ack |=
3383 			TCP_SKB_CB(next_skb)->txstamp_ack;
3384 	}
3385 }
3386 
3387 /* Collapses two adjacent SKB's during retransmission. */
tcp_collapse_retrans(struct sock * sk,struct sk_buff * skb)3388 static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
3389 {
3390 	struct tcp_sock *tp = tcp_sk(sk);
3391 	struct sk_buff *next_skb = skb_rb_next(skb);
3392 	int next_skb_size;
3393 
3394 	next_skb_size = next_skb->len;
3395 
3396 	BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
3397 
3398 	if (next_skb_size && !tcp_skb_shift(skb, next_skb, 1, next_skb_size))
3399 		return false;
3400 
3401 	tcp_highest_sack_replace(sk, next_skb, skb);
3402 
3403 	/* Update sequence range on original skb. */
3404 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
3405 
3406 	/* Merge over control information. This moves PSH/FIN etc. over */
3407 	TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
3408 
3409 	/* All done, get rid of second SKB and account for it so
3410 	 * packet counting does not break.
3411 	 */
3412 	TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
3413 	TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
3414 
3415 	/* changed transmit queue under us so clear hints */
3416 	if (next_skb == tp->retransmit_skb_hint)
3417 		tp->retransmit_skb_hint = skb;
3418 
3419 	tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
3420 
3421 	tcp_skb_collapse_tstamp(skb, next_skb);
3422 
3423 	tcp_rtx_queue_unlink_and_free(next_skb, sk);
3424 	return true;
3425 }
3426 
3427 /* Check if coalescing SKBs is legal. */
tcp_can_collapse(const struct sock * sk,const struct sk_buff * skb)3428 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
3429 {
3430 	if (tcp_skb_pcount(skb) > 1)
3431 		return false;
3432 	if (skb_cloned(skb))
3433 		return false;
3434 	if (!skb_frags_readable(skb))
3435 		return false;
3436 	/* Some heuristics for collapsing over SACK'd could be invented */
3437 	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3438 		return false;
3439 
3440 	return true;
3441 }
3442 
3443 /* Collapse packets in the retransmit queue to make to create
3444  * less packets on the wire. This is only done on retransmission.
3445  */
tcp_retrans_try_collapse(struct sock * sk,struct sk_buff * to,int space)3446 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
3447 				     int space)
3448 {
3449 	struct tcp_sock *tp = tcp_sk(sk);
3450 	struct sk_buff *skb = to, *tmp;
3451 	bool first = true;
3452 
3453 	if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse))
3454 		return;
3455 	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3456 		return;
3457 
3458 	skb_rbtree_walk_from_safe(skb, tmp) {
3459 		if (!tcp_can_collapse(sk, skb))
3460 			break;
3461 
3462 		if (!tcp_skb_can_collapse(to, skb))
3463 			break;
3464 
3465 		space -= skb->len;
3466 
3467 		if (first) {
3468 			first = false;
3469 			continue;
3470 		}
3471 
3472 		if (space < 0)
3473 			break;
3474 
3475 		if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
3476 			break;
3477 
3478 		if (!tcp_collapse_retrans(sk, to))
3479 			break;
3480 	}
3481 }
3482 
3483 /* This retransmits one SKB.  Policy decisions and retransmit queue
3484  * state updates are done by the caller.  Returns non-zero if an
3485  * error occurred which prevented the send.
3486  */
__tcp_retransmit_skb(struct sock * sk,struct sk_buff * skb,int segs)3487 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3488 {
3489 	struct inet_connection_sock *icsk = inet_csk(sk);
3490 	struct tcp_sock *tp = tcp_sk(sk);
3491 	unsigned int cur_mss;
3492 	int diff, len, err;
3493 	int avail_wnd;
3494 
3495 	/* Inconclusive MTU probe */
3496 	if (icsk->icsk_mtup.probe_size)
3497 		icsk->icsk_mtup.probe_size = 0;
3498 
3499 	if (skb_still_in_host_queue(sk, skb)) {
3500 		err = -EBUSY;
3501 		goto out;
3502 	}
3503 
3504 start:
3505 	if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
3506 		if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3507 			TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_SYN;
3508 			TCP_SKB_CB(skb)->seq++;
3509 			goto start;
3510 		}
3511 		if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
3512 			WARN_ON_ONCE(1);
3513 			err = -EINVAL;
3514 			goto out;
3515 		}
3516 		if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) {
3517 			err = -ENOMEM;
3518 			goto out;
3519 		}
3520 	}
3521 
3522 	if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) {
3523 		err = -EHOSTUNREACH; /* Routing failure or similar. */
3524 		goto out;
3525 	}
3526 
3527 	cur_mss = tcp_current_mss(sk);
3528 	avail_wnd = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3529 
3530 	/* If receiver has shrunk his window, and skb is out of
3531 	 * new window, do not retransmit it. The exception is the
3532 	 * case, when window is shrunk to zero. In this case
3533 	 * our retransmit of one segment serves as a zero window probe.
3534 	 */
3535 	if (avail_wnd <= 0) {
3536 		if (TCP_SKB_CB(skb)->seq != tp->snd_una) {
3537 			err = -EAGAIN;
3538 			goto out;
3539 		}
3540 		avail_wnd = cur_mss;
3541 	}
3542 
3543 	len = cur_mss * segs;
3544 	if (len > avail_wnd) {
3545 		len = rounddown(avail_wnd, cur_mss);
3546 		if (!len)
3547 			len = avail_wnd;
3548 	}
3549 	if (skb->len > len) {
3550 		if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
3551 				 cur_mss, GFP_ATOMIC)) {
3552 			err = -ENOMEM;  /* We'll try again later. */
3553 			goto out;
3554 		}
3555 	} else {
3556 		if (skb_unclone_keeptruesize(skb, GFP_ATOMIC)) {
3557 			err = -ENOMEM;
3558 			goto out;
3559 		}
3560 
3561 		diff = tcp_skb_pcount(skb);
3562 		tcp_set_skb_tso_segs(skb, cur_mss);
3563 		diff -= tcp_skb_pcount(skb);
3564 		if (diff)
3565 			tcp_adjust_pcount(sk, skb, diff);
3566 		avail_wnd = min_t(int, avail_wnd, cur_mss);
3567 		if (skb->len < avail_wnd)
3568 			tcp_retrans_try_collapse(sk, skb, avail_wnd);
3569 	}
3570 
3571 	/* RFC3168, section 6.1.1.1. ECN fallback
3572 	 * As AccECN uses the same SYN flags (+ AE), this check covers both
3573 	 * cases.
3574 	 */
3575 	if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
3576 		tcp_ecn_clear_syn(sk, skb);
3577 
3578 	/* Update global and local TCP statistics. */
3579 	segs = tcp_skb_pcount(skb);
3580 	TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
3581 	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3582 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3583 	tp->total_retrans += segs;
3584 	tp->bytes_retrans += skb->len;
3585 
3586 	/* make sure skb->data is aligned on arches that require it
3587 	 * and check if ack-trimming & collapsing extended the headroom
3588 	 * beyond what csum_start can cover.
3589 	 */
3590 	if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
3591 		     skb_headroom(skb) >= 0xFFFF)) {
3592 		struct sk_buff *nskb;
3593 
3594 		tcp_skb_tsorted_save(skb) {
3595 			nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
3596 			if (nskb) {
3597 				nskb->dev = NULL;
3598 				err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC);
3599 			} else {
3600 				err = -ENOBUFS;
3601 			}
3602 		} tcp_skb_tsorted_restore(skb);
3603 
3604 		if (!err) {
3605 			tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns);
3606 			tcp_rate_skb_sent(sk, skb);
3607 		}
3608 	} else {
3609 		err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3610 	}
3611 
3612 	if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
3613 		tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
3614 				  TCP_SKB_CB(skb)->seq, segs, err);
3615 
3616 	if (unlikely(err) && err != -EBUSY)
3617 		NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs);
3618 
3619 	/* To avoid taking spuriously low RTT samples based on a timestamp
3620 	 * for a transmit that never happened, always mark EVER_RETRANS
3621 	 */
3622 	TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
3623 
3624 out:
3625 	trace_tcp_retransmit_skb(sk, skb, err);
3626 	return err;
3627 }
3628 
tcp_retransmit_skb(struct sock * sk,struct sk_buff * skb,int segs)3629 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3630 {
3631 	struct tcp_sock *tp = tcp_sk(sk);
3632 	int err = __tcp_retransmit_skb(sk, skb, segs);
3633 
3634 	if (err == 0) {
3635 #if FASTRETRANS_DEBUG > 0
3636 		if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
3637 			net_dbg_ratelimited("retrans_out leaked\n");
3638 		}
3639 #endif
3640 		TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
3641 		tp->retrans_out += tcp_skb_pcount(skb);
3642 	}
3643 
3644 	/* Save stamp of the first (attempted) retransmit. */
3645 	if (!tp->retrans_stamp)
3646 		tp->retrans_stamp = tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb);
3647 
3648 	if (tp->undo_retrans < 0)
3649 		tp->undo_retrans = 0;
3650 	tp->undo_retrans += tcp_skb_pcount(skb);
3651 	return err;
3652 }
3653 
3654 /* This gets called after a retransmit timeout, and the initially
3655  * retransmitted data is acknowledged.  It tries to continue
3656  * resending the rest of the retransmit queue, until either
3657  * we've sent it all or the congestion window limit is reached.
3658  */
tcp_xmit_retransmit_queue(struct sock * sk)3659 void tcp_xmit_retransmit_queue(struct sock *sk)
3660 {
3661 	const struct inet_connection_sock *icsk = inet_csk(sk);
3662 	struct sk_buff *skb, *rtx_head, *hole = NULL;
3663 	struct tcp_sock *tp = tcp_sk(sk);
3664 	bool rearm_timer = false;
3665 	u32 max_segs;
3666 	int mib_idx;
3667 
3668 	if (!tp->packets_out)
3669 		return;
3670 
3671 	rtx_head = tcp_rtx_queue_head(sk);
3672 	skb = tp->retransmit_skb_hint ?: rtx_head;
3673 	max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
3674 	skb_rbtree_walk_from(skb) {
3675 		__u8 sacked;
3676 		int segs;
3677 
3678 		if (tcp_pacing_check(sk))
3679 			break;
3680 
3681 		/* we could do better than to assign each time */
3682 		if (!hole)
3683 			tp->retransmit_skb_hint = skb;
3684 
3685 		segs = tcp_snd_cwnd(tp) - tcp_packets_in_flight(tp);
3686 		if (segs <= 0)
3687 			break;
3688 		sacked = TCP_SKB_CB(skb)->sacked;
3689 		/* In case tcp_shift_skb_data() have aggregated large skbs,
3690 		 * we need to make sure not sending too bigs TSO packets
3691 		 */
3692 		segs = min_t(int, segs, max_segs);
3693 
3694 		if (tp->retrans_out >= tp->lost_out) {
3695 			break;
3696 		} else if (!(sacked & TCPCB_LOST)) {
3697 			if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
3698 				hole = skb;
3699 			continue;
3700 
3701 		} else {
3702 			if (icsk->icsk_ca_state != TCP_CA_Loss)
3703 				mib_idx = LINUX_MIB_TCPFASTRETRANS;
3704 			else
3705 				mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
3706 		}
3707 
3708 		if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
3709 			continue;
3710 
3711 		if (tcp_small_queue_check(sk, skb, 1))
3712 			break;
3713 
3714 		if (tcp_retransmit_skb(sk, skb, segs))
3715 			break;
3716 
3717 		NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3718 
3719 		if (tcp_in_cwnd_reduction(sk))
3720 			tp->prr_out += tcp_skb_pcount(skb);
3721 
3722 		if (skb == rtx_head &&
3723 		    icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3724 			rearm_timer = true;
3725 
3726 	}
3727 	if (rearm_timer)
3728 		tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3729 				     inet_csk(sk)->icsk_rto, true);
3730 }
3731 
3732 /* We allow to exceed memory limits for FIN packets to expedite
3733  * connection tear down and (memory) recovery.
3734  * Otherwise tcp_send_fin() could be tempted to either delay FIN
3735  * or even be forced to close flow without any FIN.
3736  * In general, we want to allow one skb per socket to avoid hangs
3737  * with edge trigger epoll()
3738  */
sk_forced_mem_schedule(struct sock * sk,int size)3739 void sk_forced_mem_schedule(struct sock *sk, int size)
3740 {
3741 	int delta, amt;
3742 
3743 	delta = size - sk->sk_forward_alloc;
3744 	if (delta <= 0)
3745 		return;
3746 	amt = sk_mem_pages(delta);
3747 	sk_forward_alloc_add(sk, amt << PAGE_SHIFT);
3748 	sk_memory_allocated_add(sk, amt);
3749 
3750 	if (mem_cgroup_sk_enabled(sk))
3751 		mem_cgroup_sk_charge(sk, amt, gfp_memcg_charge() | __GFP_NOFAIL);
3752 }
3753 
3754 /* Send a FIN. The caller locks the socket for us.
3755  * We should try to send a FIN packet really hard, but eventually give up.
3756  */
tcp_send_fin(struct sock * sk)3757 void tcp_send_fin(struct sock *sk)
3758 {
3759 	struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk);
3760 	struct tcp_sock *tp = tcp_sk(sk);
3761 
3762 	/* Optimization, tack on the FIN if we have one skb in write queue and
3763 	 * this skb was not yet sent, or we are under memory pressure.
3764 	 * Note: in the latter case, FIN packet will be sent after a timeout,
3765 	 * as TCP stack thinks it has already been transmitted.
3766 	 */
3767 	tskb = tail;
3768 	if (!tskb && tcp_under_memory_pressure(sk))
3769 		tskb = skb_rb_last(&sk->tcp_rtx_queue);
3770 
3771 	if (tskb) {
3772 		TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3773 		TCP_SKB_CB(tskb)->end_seq++;
3774 		tp->write_seq++;
3775 		if (!tail) {
3776 			/* This means tskb was already sent.
3777 			 * Pretend we included the FIN on previous transmit.
3778 			 * We need to set tp->snd_nxt to the value it would have
3779 			 * if FIN had been sent. This is because retransmit path
3780 			 * does not change tp->snd_nxt.
3781 			 */
3782 			WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1);
3783 			return;
3784 		}
3785 	} else {
3786 		skb = alloc_skb_fclone(MAX_TCP_HEADER,
3787 				       sk_gfp_mask(sk, GFP_ATOMIC |
3788 						       __GFP_NOWARN));
3789 		if (unlikely(!skb))
3790 			return;
3791 
3792 		INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3793 		skb_reserve(skb, MAX_TCP_HEADER);
3794 		sk_forced_mem_schedule(sk, skb->truesize);
3795 		/* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3796 		tcp_init_nondata_skb(skb, sk, tp->write_seq,
3797 				     TCPHDR_ACK | TCPHDR_FIN);
3798 		tcp_queue_skb(sk, skb);
3799 	}
3800 	__tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3801 }
3802 
3803 /* We get here when a process closes a file descriptor (either due to
3804  * an explicit close() or as a byproduct of exit()'ing) and there
3805  * was unread data in the receive queue.  This behavior is recommended
3806  * by RFC 2525, section 2.17.  -DaveM
3807  */
tcp_send_active_reset(struct sock * sk,gfp_t priority,enum sk_rst_reason reason)3808 void tcp_send_active_reset(struct sock *sk, gfp_t priority,
3809 			   enum sk_rst_reason reason)
3810 {
3811 	struct sk_buff *skb;
3812 
3813 	TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3814 
3815 	/* NOTE: No TCP options attached and we never retransmit this. */
3816 	skb = alloc_skb(MAX_TCP_HEADER, priority);
3817 	if (!skb) {
3818 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3819 		return;
3820 	}
3821 
3822 	/* Reserve space for headers and prepare control bits. */
3823 	skb_reserve(skb, MAX_TCP_HEADER);
3824 	tcp_init_nondata_skb(skb, sk, tcp_acceptable_seq(sk),
3825 			     TCPHDR_ACK | TCPHDR_RST);
3826 	tcp_mstamp_refresh(tcp_sk(sk));
3827 	/* Send it off. */
3828 	if (tcp_transmit_skb(sk, skb, 0, priority))
3829 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3830 
3831 	/* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3832 	 * skb here is different to the troublesome skb, so use NULL
3833 	 */
3834 	trace_tcp_send_reset(sk, NULL, reason);
3835 }
3836 
3837 /* Send a crossed SYN-ACK during socket establishment.
3838  * WARNING: This routine must only be called when we have already sent
3839  * a SYN packet that crossed the incoming SYN that caused this routine
3840  * to get called. If this assumption fails then the initial rcv_wnd
3841  * and rcv_wscale values will not be correct.
3842  */
tcp_send_synack(struct sock * sk)3843 int tcp_send_synack(struct sock *sk)
3844 {
3845 	struct sk_buff *skb;
3846 
3847 	skb = tcp_rtx_queue_head(sk);
3848 	if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3849 		pr_err("%s: wrong queue state\n", __func__);
3850 		return -EFAULT;
3851 	}
3852 	if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3853 		if (skb_cloned(skb)) {
3854 			struct sk_buff *nskb;
3855 
3856 			tcp_skb_tsorted_save(skb) {
3857 				nskb = skb_copy(skb, GFP_ATOMIC);
3858 			} tcp_skb_tsorted_restore(skb);
3859 			if (!nskb)
3860 				return -ENOMEM;
3861 			INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
3862 			tcp_highest_sack_replace(sk, skb, nskb);
3863 			tcp_rtx_queue_unlink_and_free(skb, sk);
3864 			__skb_header_release(nskb);
3865 			tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3866 			sk_wmem_queued_add(sk, nskb->truesize);
3867 			sk_mem_charge(sk, nskb->truesize);
3868 			skb = nskb;
3869 		}
3870 
3871 		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3872 		tcp_ecn_send_synack(sk, skb);
3873 	}
3874 	return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3875 }
3876 
3877 /**
3878  * tcp_make_synack - Allocate one skb and build a SYNACK packet.
3879  * @sk: listener socket
3880  * @dst: dst entry attached to the SYNACK. It is consumed and caller
3881  *       should not use it again.
3882  * @req: request_sock pointer
3883  * @foc: cookie for tcp fast open
3884  * @synack_type: Type of synack to prepare
3885  * @syn_skb: SYN packet just received.  It could be NULL for rtx case.
3886  */
tcp_make_synack(const struct sock * sk,struct dst_entry * dst,struct request_sock * req,struct tcp_fastopen_cookie * foc,enum tcp_synack_type synack_type,struct sk_buff * syn_skb)3887 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3888 				struct request_sock *req,
3889 				struct tcp_fastopen_cookie *foc,
3890 				enum tcp_synack_type synack_type,
3891 				struct sk_buff *syn_skb)
3892 {
3893 	struct inet_request_sock *ireq = inet_rsk(req);
3894 	const struct tcp_sock *tp = tcp_sk(sk);
3895 	struct tcp_out_options opts;
3896 	struct tcp_key key = {};
3897 	struct sk_buff *skb;
3898 	int tcp_header_size;
3899 	struct tcphdr *th;
3900 	int mss;
3901 	u64 now;
3902 
3903 	skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3904 	if (unlikely(!skb)) {
3905 		dst_release(dst);
3906 		return NULL;
3907 	}
3908 	/* Reserve space for headers. */
3909 	skb_reserve(skb, MAX_TCP_HEADER);
3910 
3911 	switch (synack_type) {
3912 	case TCP_SYNACK_NORMAL:
3913 		skb_set_owner_edemux(skb, req_to_sk(req));
3914 		break;
3915 	case TCP_SYNACK_COOKIE:
3916 		/* Under synflood, we do not attach skb to a socket,
3917 		 * to avoid false sharing.
3918 		 */
3919 		break;
3920 	case TCP_SYNACK_FASTOPEN:
3921 		/* sk is a const pointer, because we want to express multiple
3922 		 * cpu might call us concurrently.
3923 		 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3924 		 */
3925 		skb_set_owner_w(skb, (struct sock *)sk);
3926 		break;
3927 	}
3928 	skb_dst_set(skb, dst);
3929 
3930 	mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3931 
3932 	memset(&opts, 0, sizeof(opts));
3933 	now = tcp_clock_ns();
3934 #ifdef CONFIG_SYN_COOKIES
3935 	if (unlikely(synack_type == TCP_SYNACK_COOKIE && ireq->tstamp_ok))
3936 		skb_set_delivery_time(skb, cookie_init_timestamp(req, now),
3937 				      SKB_CLOCK_MONOTONIC);
3938 	else
3939 #endif
3940 	{
3941 		skb_set_delivery_time(skb, now, SKB_CLOCK_MONOTONIC);
3942 		if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */
3943 			tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb);
3944 	}
3945 
3946 #if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO)
3947 	rcu_read_lock();
3948 #endif
3949 	if (tcp_rsk_used_ao(req)) {
3950 #ifdef CONFIG_TCP_AO
3951 		struct tcp_ao_key *ao_key = NULL;
3952 		u8 keyid = tcp_rsk(req)->ao_keyid;
3953 		u8 rnext = tcp_rsk(req)->ao_rcv_next;
3954 
3955 		ao_key = tcp_sk(sk)->af_specific->ao_lookup(sk, req_to_sk(req),
3956 							    keyid, -1);
3957 		/* If there is no matching key - avoid sending anything,
3958 		 * especially usigned segments. It could try harder and lookup
3959 		 * for another peer-matching key, but the peer has requested
3960 		 * ao_keyid (RFC5925 RNextKeyID), so let's keep it simple here.
3961 		 */
3962 		if (unlikely(!ao_key)) {
3963 			trace_tcp_ao_synack_no_key(sk, keyid, rnext);
3964 			rcu_read_unlock();
3965 			kfree_skb(skb);
3966 			net_warn_ratelimited("TCP-AO: the keyid %u from SYN packet is not present - not sending SYNACK\n",
3967 					     keyid);
3968 			return NULL;
3969 		}
3970 		key.ao_key = ao_key;
3971 		key.type = TCP_KEY_AO;
3972 #endif
3973 	} else {
3974 #ifdef CONFIG_TCP_MD5SIG
3975 		key.md5_key = tcp_rsk(req)->af_specific->req_md5_lookup(sk,
3976 					req_to_sk(req));
3977 		if (key.md5_key)
3978 			key.type = TCP_KEY_MD5;
3979 #endif
3980 	}
3981 	skb_set_hash(skb, READ_ONCE(tcp_rsk(req)->txhash), PKT_HASH_TYPE_L4);
3982 	/* bpf program will be interested in the tcp_flags */
3983 	TCP_SKB_CB(skb)->tcp_flags = TCPHDR_SYN | TCPHDR_ACK;
3984 	tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts,
3985 					     &key, foc, synack_type, syn_skb)
3986 					+ sizeof(*th);
3987 
3988 	skb_push(skb, tcp_header_size);
3989 	skb_reset_transport_header(skb);
3990 
3991 	th = (struct tcphdr *)skb->data;
3992 	memset(th, 0, sizeof(struct tcphdr));
3993 	th->syn = 1;
3994 	th->ack = 1;
3995 	tcp_ecn_make_synack(req, th);
3996 	th->source = htons(ireq->ir_num);
3997 	th->dest = ireq->ir_rmt_port;
3998 	skb->mark = ireq->ir_mark;
3999 	skb->ip_summed = CHECKSUM_PARTIAL;
4000 	th->seq = htonl(tcp_rsk(req)->snt_isn);
4001 	/* XXX data is queued and acked as is. No buffer/window check */
4002 	th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
4003 
4004 	/* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
4005 	th->window = htons(min(req->rsk_rcv_wnd, 65535U));
4006 	tcp_options_write(th, NULL, tcp_rsk(req), &opts, &key);
4007 	th->doff = (tcp_header_size >> 2);
4008 	TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
4009 
4010 	/* Okay, we have all we need - do the md5 hash if needed */
4011 	if (tcp_key_is_md5(&key)) {
4012 #ifdef CONFIG_TCP_MD5SIG
4013 		tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
4014 					key.md5_key, req_to_sk(req), skb);
4015 #endif
4016 	} else if (tcp_key_is_ao(&key)) {
4017 #ifdef CONFIG_TCP_AO
4018 		tcp_rsk(req)->af_specific->ao_synack_hash(opts.hash_location,
4019 					key.ao_key, req, skb,
4020 					opts.hash_location - (u8 *)th, 0);
4021 #endif
4022 	}
4023 #if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO)
4024 	rcu_read_unlock();
4025 #endif
4026 
4027 	bpf_skops_write_hdr_opt((struct sock *)sk, skb, req, syn_skb,
4028 				synack_type, &opts);
4029 
4030 	skb_set_delivery_time(skb, now, SKB_CLOCK_MONOTONIC);
4031 	tcp_add_tx_delay(skb, tp);
4032 
4033 	return skb;
4034 }
4035 EXPORT_IPV6_MOD(tcp_make_synack);
4036 
tcp_ca_dst_init(struct sock * sk,const struct dst_entry * dst)4037 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
4038 {
4039 	struct inet_connection_sock *icsk = inet_csk(sk);
4040 	const struct tcp_congestion_ops *ca;
4041 	u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
4042 
4043 	if (ca_key == TCP_CA_UNSPEC)
4044 		return;
4045 
4046 	rcu_read_lock();
4047 	ca = tcp_ca_find_key(ca_key);
4048 	if (likely(ca && bpf_try_module_get(ca, ca->owner))) {
4049 		bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner);
4050 		icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
4051 		icsk->icsk_ca_ops = ca;
4052 	}
4053 	rcu_read_unlock();
4054 }
4055 
4056 /* Do all connect socket setups that can be done AF independent. */
tcp_connect_init(struct sock * sk)4057 static void tcp_connect_init(struct sock *sk)
4058 {
4059 	const struct dst_entry *dst = __sk_dst_get(sk);
4060 	struct tcp_sock *tp = tcp_sk(sk);
4061 	__u8 rcv_wscale;
4062 	u16 user_mss;
4063 	u32 rcv_wnd;
4064 
4065 	/* We'll fix this up when we get a response from the other end.
4066 	 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
4067 	 */
4068 	tp->tcp_header_len = sizeof(struct tcphdr);
4069 	if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps))
4070 		tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
4071 
4072 	tcp_ao_connect_init(sk);
4073 
4074 	/* If user gave his TCP_MAXSEG, record it to clamp */
4075 	user_mss = READ_ONCE(tp->rx_opt.user_mss);
4076 	if (user_mss)
4077 		tp->rx_opt.mss_clamp = user_mss;
4078 	tp->max_window = 0;
4079 	tcp_mtup_init(sk);
4080 	tcp_sync_mss(sk, dst_mtu(dst));
4081 
4082 	tcp_ca_dst_init(sk, dst);
4083 
4084 	if (!tp->window_clamp)
4085 		WRITE_ONCE(tp->window_clamp, dst_metric(dst, RTAX_WINDOW));
4086 	tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
4087 
4088 	tcp_initialize_rcv_mss(sk);
4089 
4090 	/* limit the window selection if the user enforce a smaller rx buffer */
4091 	if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
4092 	    (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
4093 		WRITE_ONCE(tp->window_clamp, tcp_full_space(sk));
4094 
4095 	rcv_wnd = tcp_rwnd_init_bpf(sk);
4096 	if (rcv_wnd == 0)
4097 		rcv_wnd = dst_metric(dst, RTAX_INITRWND);
4098 
4099 	tcp_select_initial_window(sk, tcp_full_space(sk),
4100 				  tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
4101 				  &tp->rcv_wnd,
4102 				  &tp->window_clamp,
4103 				  READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling),
4104 				  &rcv_wscale,
4105 				  rcv_wnd);
4106 
4107 	tp->rx_opt.rcv_wscale = rcv_wscale;
4108 	tp->rcv_ssthresh = tp->rcv_wnd;
4109 
4110 	WRITE_ONCE(sk->sk_err, 0);
4111 	sock_reset_flag(sk, SOCK_DONE);
4112 	tp->snd_wnd = 0;
4113 	tcp_init_wl(tp, 0);
4114 	tcp_write_queue_purge(sk);
4115 	tp->snd_una = tp->write_seq;
4116 	tp->snd_sml = tp->write_seq;
4117 	tp->snd_up = tp->write_seq;
4118 	WRITE_ONCE(tp->snd_nxt, tp->write_seq);
4119 
4120 	if (likely(!tp->repair))
4121 		tp->rcv_nxt = 0;
4122 	else
4123 		tp->rcv_tstamp = tcp_jiffies32;
4124 	tp->rcv_wup = tp->rcv_nxt;
4125 	WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
4126 
4127 	inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
4128 	WRITE_ONCE(inet_csk(sk)->icsk_retransmits, 0);
4129 	tcp_clear_retrans(tp);
4130 }
4131 
tcp_connect_queue_skb(struct sock * sk,struct sk_buff * skb)4132 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
4133 {
4134 	struct tcp_sock *tp = tcp_sk(sk);
4135 	struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
4136 
4137 	tcb->end_seq += skb->len;
4138 	__skb_header_release(skb);
4139 	sk_wmem_queued_add(sk, skb->truesize);
4140 	sk_mem_charge(sk, skb->truesize);
4141 	WRITE_ONCE(tp->write_seq, tcb->end_seq);
4142 	tp->packets_out += tcp_skb_pcount(skb);
4143 }
4144 
4145 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
4146  * queue a data-only packet after the regular SYN, such that regular SYNs
4147  * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
4148  * only the SYN sequence, the data are retransmitted in the first ACK.
4149  * If cookie is not cached or other error occurs, falls back to send a
4150  * regular SYN with Fast Open cookie request option.
4151  */
tcp_send_syn_data(struct sock * sk,struct sk_buff * syn)4152 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
4153 {
4154 	struct inet_connection_sock *icsk = inet_csk(sk);
4155 	struct tcp_sock *tp = tcp_sk(sk);
4156 	struct tcp_fastopen_request *fo = tp->fastopen_req;
4157 	struct page_frag *pfrag = sk_page_frag(sk);
4158 	struct sk_buff *syn_data;
4159 	int space, err = 0;
4160 
4161 	tp->rx_opt.mss_clamp = tp->advmss;  /* If MSS is not cached */
4162 	if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
4163 		goto fallback;
4164 
4165 	/* MSS for SYN-data is based on cached MSS and bounded by PMTU and
4166 	 * user-MSS. Reserve maximum option space for middleboxes that add
4167 	 * private TCP options. The cost is reduced data space in SYN :(
4168 	 */
4169 	tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
4170 	/* Sync mss_cache after updating the mss_clamp */
4171 	tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4172 
4173 	space = __tcp_mtu_to_mss(sk, icsk->icsk_pmtu_cookie) -
4174 		MAX_TCP_OPTION_SPACE;
4175 
4176 	space = min_t(size_t, space, fo->size);
4177 
4178 	if (space &&
4179 	    !skb_page_frag_refill(min_t(size_t, space, PAGE_SIZE),
4180 				  pfrag, sk->sk_allocation))
4181 		goto fallback;
4182 	syn_data = tcp_stream_alloc_skb(sk, sk->sk_allocation, false);
4183 	if (!syn_data)
4184 		goto fallback;
4185 	memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
4186 	if (space) {
4187 		space = min_t(size_t, space, pfrag->size - pfrag->offset);
4188 		space = tcp_wmem_schedule(sk, space);
4189 	}
4190 	if (space) {
4191 		space = copy_page_from_iter(pfrag->page, pfrag->offset,
4192 					    space, &fo->data->msg_iter);
4193 		if (unlikely(!space)) {
4194 			tcp_skb_tsorted_anchor_cleanup(syn_data);
4195 			kfree_skb(syn_data);
4196 			goto fallback;
4197 		}
4198 		skb_fill_page_desc(syn_data, 0, pfrag->page,
4199 				   pfrag->offset, space);
4200 		page_ref_inc(pfrag->page);
4201 		pfrag->offset += space;
4202 		skb_len_add(syn_data, space);
4203 		skb_zcopy_set(syn_data, fo->uarg, NULL);
4204 	}
4205 	/* No more data pending in inet_wait_for_connect() */
4206 	if (space == fo->size)
4207 		fo->data = NULL;
4208 	fo->copied = space;
4209 
4210 	tcp_connect_queue_skb(sk, syn_data);
4211 	if (syn_data->len)
4212 		tcp_chrono_start(sk, TCP_CHRONO_BUSY);
4213 
4214 	err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
4215 
4216 	skb_set_delivery_time(syn, syn_data->skb_mstamp_ns, SKB_CLOCK_MONOTONIC);
4217 
4218 	/* Now full SYN+DATA was cloned and sent (or not),
4219 	 * remove the SYN from the original skb (syn_data)
4220 	 * we keep in write queue in case of a retransmit, as we
4221 	 * also have the SYN packet (with no data) in the same queue.
4222 	 */
4223 	TCP_SKB_CB(syn_data)->seq++;
4224 	TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
4225 	if (!err) {
4226 		tp->syn_data = (fo->copied > 0);
4227 		tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
4228 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
4229 		goto done;
4230 	}
4231 
4232 	/* data was not sent, put it in write_queue */
4233 	__skb_queue_tail(&sk->sk_write_queue, syn_data);
4234 	tp->packets_out -= tcp_skb_pcount(syn_data);
4235 
4236 fallback:
4237 	/* Send a regular SYN with Fast Open cookie request option */
4238 	if (fo->cookie.len > 0)
4239 		fo->cookie.len = 0;
4240 	err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
4241 	if (err)
4242 		tp->syn_fastopen = 0;
4243 done:
4244 	fo->cookie.len = -1;  /* Exclude Fast Open option for SYN retries */
4245 	return err;
4246 }
4247 
4248 /* Build a SYN and send it off. */
tcp_connect(struct sock * sk)4249 int tcp_connect(struct sock *sk)
4250 {
4251 	struct tcp_sock *tp = tcp_sk(sk);
4252 	struct sk_buff *buff;
4253 	int err;
4254 
4255 	tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
4256 
4257 #if defined(CONFIG_TCP_MD5SIG) && defined(CONFIG_TCP_AO)
4258 	/* Has to be checked late, after setting daddr/saddr/ops.
4259 	 * Return error if the peer has both a md5 and a tcp-ao key
4260 	 * configured as this is ambiguous.
4261 	 */
4262 	if (unlikely(rcu_dereference_protected(tp->md5sig_info,
4263 					       lockdep_sock_is_held(sk)))) {
4264 		bool needs_ao = !!tp->af_specific->ao_lookup(sk, sk, -1, -1);
4265 		bool needs_md5 = !!tp->af_specific->md5_lookup(sk, sk);
4266 		struct tcp_ao_info *ao_info;
4267 
4268 		ao_info = rcu_dereference_check(tp->ao_info,
4269 						lockdep_sock_is_held(sk));
4270 		if (ao_info) {
4271 			/* This is an extra check: tcp_ao_required() in
4272 			 * tcp_v{4,6}_parse_md5_keys() should prevent adding
4273 			 * md5 keys on ao_required socket.
4274 			 */
4275 			needs_ao |= ao_info->ao_required;
4276 			WARN_ON_ONCE(ao_info->ao_required && needs_md5);
4277 		}
4278 		if (needs_md5 && needs_ao)
4279 			return -EKEYREJECTED;
4280 
4281 		/* If we have a matching md5 key and no matching tcp-ao key
4282 		 * then free up ao_info if allocated.
4283 		 */
4284 		if (needs_md5) {
4285 			tcp_ao_destroy_sock(sk, false);
4286 		} else if (needs_ao) {
4287 			tcp_clear_md5_list(sk);
4288 			kfree(rcu_replace_pointer(tp->md5sig_info, NULL,
4289 						  lockdep_sock_is_held(sk)));
4290 		}
4291 	}
4292 #endif
4293 #ifdef CONFIG_TCP_AO
4294 	if (unlikely(rcu_dereference_protected(tp->ao_info,
4295 					       lockdep_sock_is_held(sk)))) {
4296 		/* Don't allow connecting if ao is configured but no
4297 		 * matching key is found.
4298 		 */
4299 		if (!tp->af_specific->ao_lookup(sk, sk, -1, -1))
4300 			return -EKEYREJECTED;
4301 	}
4302 #endif
4303 
4304 	if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
4305 		return -EHOSTUNREACH; /* Routing failure or similar. */
4306 
4307 	tcp_connect_init(sk);
4308 
4309 	if (unlikely(tp->repair)) {
4310 		tcp_finish_connect(sk, NULL);
4311 		return 0;
4312 	}
4313 
4314 	buff = tcp_stream_alloc_skb(sk, sk->sk_allocation, true);
4315 	if (unlikely(!buff))
4316 		return -ENOBUFS;
4317 
4318 	/* SYN eats a sequence byte, write_seq updated by
4319 	 * tcp_connect_queue_skb().
4320 	 */
4321 	tcp_init_nondata_skb(buff, sk, tp->write_seq, TCPHDR_SYN);
4322 	tcp_mstamp_refresh(tp);
4323 	tp->retrans_stamp = tcp_time_stamp_ts(tp);
4324 	tcp_connect_queue_skb(sk, buff);
4325 	tcp_ecn_send_syn(sk, buff);
4326 	tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
4327 
4328 	/* Send off SYN; include data in Fast Open. */
4329 	err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
4330 	      tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
4331 	if (err == -ECONNREFUSED)
4332 		return err;
4333 
4334 	/* We change tp->snd_nxt after the tcp_transmit_skb() call
4335 	 * in order to make this packet get counted in tcpOutSegs.
4336 	 */
4337 	WRITE_ONCE(tp->snd_nxt, tp->write_seq);
4338 	tp->pushed_seq = tp->write_seq;
4339 	buff = tcp_send_head(sk);
4340 	if (unlikely(buff)) {
4341 		WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq);
4342 		tp->pushed_seq	= TCP_SKB_CB(buff)->seq;
4343 	}
4344 	TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
4345 
4346 	/* Timer for repeating the SYN until an answer. */
4347 	tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
4348 			     inet_csk(sk)->icsk_rto, false);
4349 	return 0;
4350 }
4351 EXPORT_SYMBOL(tcp_connect);
4352 
tcp_delack_max(const struct sock * sk)4353 u32 tcp_delack_max(const struct sock *sk)
4354 {
4355 	u32 delack_from_rto_min = max(tcp_rto_min(sk), 2) - 1;
4356 
4357 	return min(READ_ONCE(inet_csk(sk)->icsk_delack_max), delack_from_rto_min);
4358 }
4359 
4360 /* Send out a delayed ack, the caller does the policy checking
4361  * to see if we should even be here.  See tcp_input.c:tcp_ack_snd_check()
4362  * for details.
4363  */
tcp_send_delayed_ack(struct sock * sk)4364 void tcp_send_delayed_ack(struct sock *sk)
4365 {
4366 	struct inet_connection_sock *icsk = inet_csk(sk);
4367 	int ato = icsk->icsk_ack.ato;
4368 	unsigned long timeout;
4369 
4370 	if (ato > TCP_DELACK_MIN) {
4371 		const struct tcp_sock *tp = tcp_sk(sk);
4372 		int max_ato = HZ / 2;
4373 
4374 		if (inet_csk_in_pingpong_mode(sk) ||
4375 		    (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
4376 			max_ato = TCP_DELACK_MAX;
4377 
4378 		/* Slow path, intersegment interval is "high". */
4379 
4380 		/* If some rtt estimate is known, use it to bound delayed ack.
4381 		 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
4382 		 * directly.
4383 		 */
4384 		if (tp->srtt_us) {
4385 			int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
4386 					TCP_DELACK_MIN);
4387 
4388 			if (rtt < max_ato)
4389 				max_ato = rtt;
4390 		}
4391 
4392 		ato = min(ato, max_ato);
4393 	}
4394 
4395 	ato = min_t(u32, ato, tcp_delack_max(sk));
4396 
4397 	/* Stay within the limit we were given */
4398 	timeout = jiffies + ato;
4399 
4400 	/* Use new timeout only if there wasn't a older one earlier. */
4401 	if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
4402 		/* If delack timer is about to expire, send ACK now. */
4403 		if (time_before_eq(icsk_delack_timeout(icsk), jiffies + (ato >> 2))) {
4404 			tcp_send_ack(sk);
4405 			return;
4406 		}
4407 
4408 		if (!time_before(timeout, icsk_delack_timeout(icsk)))
4409 			timeout = icsk_delack_timeout(icsk);
4410 	}
4411 	smp_store_release(&icsk->icsk_ack.pending,
4412 			  icsk->icsk_ack.pending | ICSK_ACK_SCHED | ICSK_ACK_TIMER);
4413 	sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
4414 }
4415 
4416 /* This routine sends an ack and also updates the window. */
__tcp_send_ack(struct sock * sk,u32 rcv_nxt,u16 flags)4417 void __tcp_send_ack(struct sock *sk, u32 rcv_nxt, u16 flags)
4418 {
4419 	struct sk_buff *buff;
4420 
4421 	/* If we have been reset, we may not send again. */
4422 	if (sk->sk_state == TCP_CLOSE)
4423 		return;
4424 
4425 	/* We are not putting this on the write queue, so
4426 	 * tcp_transmit_skb() will set the ownership to this
4427 	 * sock.
4428 	 */
4429 	buff = alloc_skb(MAX_TCP_HEADER,
4430 			 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
4431 	if (unlikely(!buff)) {
4432 		struct inet_connection_sock *icsk = inet_csk(sk);
4433 		unsigned long delay;
4434 
4435 		delay = TCP_DELACK_MAX << icsk->icsk_ack.retry;
4436 		if (delay < tcp_rto_max(sk))
4437 			icsk->icsk_ack.retry++;
4438 		inet_csk_schedule_ack(sk);
4439 		icsk->icsk_ack.ato = TCP_ATO_MIN;
4440 		tcp_reset_xmit_timer(sk, ICSK_TIME_DACK, delay, false);
4441 		return;
4442 	}
4443 
4444 	/* Reserve space for headers and prepare control bits. */
4445 	skb_reserve(buff, MAX_TCP_HEADER);
4446 	tcp_init_nondata_skb(buff, sk,
4447 			     tcp_acceptable_seq(sk), TCPHDR_ACK | flags);
4448 
4449 	/* We do not want pure acks influencing TCP Small Queues or fq/pacing
4450 	 * too much.
4451 	 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
4452 	 */
4453 	skb_set_tcp_pure_ack(buff);
4454 
4455 	/* Send it off, this clears delayed acks for us. */
4456 	__tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
4457 }
4458 EXPORT_SYMBOL_GPL(__tcp_send_ack);
4459 
tcp_send_ack(struct sock * sk)4460 void tcp_send_ack(struct sock *sk)
4461 {
4462 	__tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt, 0);
4463 }
4464 
4465 /* This routine sends a packet with an out of date sequence
4466  * number. It assumes the other end will try to ack it.
4467  *
4468  * Question: what should we make while urgent mode?
4469  * 4.4BSD forces sending single byte of data. We cannot send
4470  * out of window data, because we have SND.NXT==SND.MAX...
4471  *
4472  * Current solution: to send TWO zero-length segments in urgent mode:
4473  * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
4474  * out-of-date with SND.UNA-1 to probe window.
4475  */
tcp_xmit_probe_skb(struct sock * sk,int urgent,int mib)4476 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
4477 {
4478 	struct tcp_sock *tp = tcp_sk(sk);
4479 	struct sk_buff *skb;
4480 
4481 	/* We don't queue it, tcp_transmit_skb() sets ownership. */
4482 	skb = alloc_skb(MAX_TCP_HEADER,
4483 			sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
4484 	if (!skb)
4485 		return -1;
4486 
4487 	/* Reserve space for headers and set control bits. */
4488 	skb_reserve(skb, MAX_TCP_HEADER);
4489 	/* Use a previous sequence.  This should cause the other
4490 	 * end to send an ack.  Don't queue or clone SKB, just
4491 	 * send it.
4492 	 */
4493 	tcp_init_nondata_skb(skb, sk, tp->snd_una - !urgent, TCPHDR_ACK);
4494 	NET_INC_STATS(sock_net(sk), mib);
4495 	return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
4496 }
4497 
4498 /* Called from setsockopt( ... TCP_REPAIR ) */
tcp_send_window_probe(struct sock * sk)4499 void tcp_send_window_probe(struct sock *sk)
4500 {
4501 	if (sk->sk_state == TCP_ESTABLISHED) {
4502 		tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
4503 		tcp_mstamp_refresh(tcp_sk(sk));
4504 		tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
4505 	}
4506 }
4507 
4508 /* Initiate keepalive or window probe from timer. */
tcp_write_wakeup(struct sock * sk,int mib)4509 int tcp_write_wakeup(struct sock *sk, int mib)
4510 {
4511 	struct tcp_sock *tp = tcp_sk(sk);
4512 	struct sk_buff *skb;
4513 
4514 	if (sk->sk_state == TCP_CLOSE)
4515 		return -1;
4516 
4517 	skb = tcp_send_head(sk);
4518 	if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
4519 		int err;
4520 		unsigned int mss = tcp_current_mss(sk);
4521 		unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
4522 
4523 		if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
4524 			tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
4525 
4526 		/* We are probing the opening of a window
4527 		 * but the window size is != 0
4528 		 * must have been a result SWS avoidance ( sender )
4529 		 */
4530 		if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
4531 		    skb->len > mss) {
4532 			seg_size = min(seg_size, mss);
4533 			TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
4534 			if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
4535 					 skb, seg_size, mss, GFP_ATOMIC))
4536 				return -1;
4537 		} else if (!tcp_skb_pcount(skb))
4538 			tcp_set_skb_tso_segs(skb, mss);
4539 
4540 		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
4541 		err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
4542 		if (!err)
4543 			tcp_event_new_data_sent(sk, skb);
4544 		return err;
4545 	} else {
4546 		if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
4547 			tcp_xmit_probe_skb(sk, 1, mib);
4548 		return tcp_xmit_probe_skb(sk, 0, mib);
4549 	}
4550 }
4551 
4552 /* A window probe timeout has occurred.  If window is not closed send
4553  * a partial packet else a zero probe.
4554  */
tcp_send_probe0(struct sock * sk)4555 void tcp_send_probe0(struct sock *sk)
4556 {
4557 	struct inet_connection_sock *icsk = inet_csk(sk);
4558 	struct tcp_sock *tp = tcp_sk(sk);
4559 	struct net *net = sock_net(sk);
4560 	unsigned long timeout;
4561 	int err;
4562 
4563 	err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
4564 
4565 	if (tp->packets_out || tcp_write_queue_empty(sk)) {
4566 		/* Cancel probe timer, if it is not required. */
4567 		WRITE_ONCE(icsk->icsk_probes_out, 0);
4568 		icsk->icsk_backoff = 0;
4569 		icsk->icsk_probes_tstamp = 0;
4570 		return;
4571 	}
4572 
4573 	WRITE_ONCE(icsk->icsk_probes_out, icsk->icsk_probes_out + 1);
4574 	if (err <= 0) {
4575 		if (icsk->icsk_backoff < READ_ONCE(net->ipv4.sysctl_tcp_retries2))
4576 			icsk->icsk_backoff++;
4577 		timeout = tcp_probe0_when(sk, tcp_rto_max(sk));
4578 	} else {
4579 		/* If packet was not sent due to local congestion,
4580 		 * Let senders fight for local resources conservatively.
4581 		 */
4582 		timeout = TCP_RESOURCE_PROBE_INTERVAL;
4583 	}
4584 
4585 	timeout = tcp_clamp_probe0_to_user_timeout(sk, timeout);
4586 	tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, true);
4587 }
4588 
tcp_rtx_synack(const struct sock * sk,struct request_sock * req)4589 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
4590 {
4591 	const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
4592 	struct flowi fl;
4593 	int res;
4594 
4595 	/* Paired with WRITE_ONCE() in sock_setsockopt() */
4596 	if (READ_ONCE(sk->sk_txrehash) == SOCK_TXREHASH_ENABLED)
4597 		WRITE_ONCE(tcp_rsk(req)->txhash, net_tx_rndhash());
4598 	res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL,
4599 				  NULL);
4600 	if (!res) {
4601 		TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
4602 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
4603 		if (unlikely(tcp_passive_fastopen(sk))) {
4604 			/* sk has const attribute because listeners are lockless.
4605 			 * However in this case, we are dealing with a passive fastopen
4606 			 * socket thus we can change total_retrans value.
4607 			 */
4608 			tcp_sk_rw(sk)->total_retrans++;
4609 		}
4610 		trace_tcp_retransmit_synack(sk, req);
4611 		WRITE_ONCE(req->num_retrans, req->num_retrans + 1);
4612 	}
4613 	return res;
4614 }
4615 EXPORT_IPV6_MOD(tcp_rtx_synack);
4616