]> git.saurik.com Git - apple/xnu.git/blob - bsd/netinet/tcp_timer.c
projects / apple / xnu.git / blob
? search:


b1ac3138b186b3b102ab02419ae4442d0c95eff5
[apple/xnu.git] / bsd / netinet / tcp_timer.c
1 /*
2 * Copyright (c) 2000-2014 Apple Inc. All rights reserved.
3 *
4 * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
5 *
6 * This file contains Original Code and/or Modifications of Original Code
7 * as defined in and that are subject to the Apple Public Source License
8 * Version 2.0 (the 'License'). You may not use this file except in
9 * compliance with the License. The rights granted to you under the License
10 * may not be used to create, or enable the creation or redistribution of,
11 * unlawful or unlicensed copies of an Apple operating system, or to
12 * circumvent, violate, or enable the circumvention or violation of, any
13 * terms of an Apple operating system software license agreement.
14 *
15 * Please obtain a copy of the License at
16 * http://www.opensource.apple.com/apsl/ and read it before using this file.
17 *
18 * The Original Code and all software distributed under the License are
19 * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
20 * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
21 * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
22 * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
23 * Please see the License for the specific language governing rights and
24 * limitations under the License.
25 *
26 * @APPLE_OSREFERENCE_LICENSE_HEADER_END@
27 */
28 /*
29 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
30 * The Regents of the University of California. All rights reserved.
31 *
32 * Redistribution and use in source and binary forms, with or without
33 * modification, are permitted provided that the following conditions
34 * are met:
35 * 1. Redistributions of source code must retain the above copyright
36 * notice, this list of conditions and the following disclaimer.
37 * 2. Redistributions in binary form must reproduce the above copyright
38 * notice, this list of conditions and the following disclaimer in the
39 * documentation and/or other materials provided with the distribution.
40 * 3. All advertising materials mentioning features or use of this software
41 * must display the following acknowledgement:
42 * This product includes software developed by the University of
43 * California, Berkeley and its contributors.
44 * 4. Neither the name of the University nor the names of its contributors
45 * may be used to endorse or promote products derived from this software
46 * without specific prior written permission.
47 *
48 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
49 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
50 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
51 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
52 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
53 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
54 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
55 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
56 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
57 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
58 * SUCH DAMAGE.
59 *
60 * @(#)tcp_timer.c 8.2 (Berkeley) 5/24/95
61 * $FreeBSD: src/sys/netinet/tcp_timer.c,v 1.34.2.11 2001/08/22 00:59:12 silby Exp $
62 */
63
64
65 #include <sys/param.h>
66 #include <sys/systm.h>
67 #include <sys/kernel.h>
68 #include <sys/mbuf.h>
69 #include <sys/sysctl.h>
70 #include <sys/socket.h>
71 #include <sys/socketvar.h>
72 #include <sys/protosw.h>
73 #include <sys/domain.h>
74 #include <sys/mcache.h>
75 #include <sys/queue.h>
76 #include <kern/locks.h>
77 #include <kern/cpu_number.h> /* before tcp_seq.h, for tcp_random18() */
78 #include <mach/boolean.h>
79
80 #include <net/route.h>
81 #include <net/if_var.h>
82
83 #include <netinet/in.h>
84 #include <netinet/in_systm.h>
85 #include <netinet/in_pcb.h>
86 #if INET6
87 #include <netinet6/in6_pcb.h>
88 #endif
89 #include <netinet/ip_var.h>
90 #include <netinet/tcp.h>
91 #include <netinet/tcp_fsm.h>
92 #include <netinet/tcp_seq.h>
93 #include <netinet/tcp_timer.h>
94 #include <netinet/tcp_var.h>
95 #include <netinet/tcp_cc.h>
96 #if INET6
97 #include <netinet6/tcp6_var.h>
98 #endif
99 #include <netinet/tcpip.h>
100 #if TCPDEBUG
101 #include <netinet/tcp_debug.h>
102 #endif
103 #include <sys/kdebug.h>
104 #include <mach/sdt.h>
105 #include <netinet/mptcp_var.h>
106
107 extern void postevent(struct socket *, struct sockbuf *,
108 int);
109 #define DBG_FNC_TCP_FAST NETDBG_CODE(DBG_NETTCP, (5 << 8))
110 #define DBG_FNC_TCP_SLOW NETDBG_CODE(DBG_NETTCP, (5 << 8) | 1)
111
112 #define TIMERENTRY_TO_TP(te) ((struct tcpcb *)((uintptr_t)te - offsetof(struct tcpcb, tentry.le.le_next)))
113
114 #define VERIFY_NEXT_LINK(elm,field) do { \
115 if (LIST_NEXT((elm),field) != NULL && \
116 LIST_NEXT((elm),field)->field.le_prev != \
117 &((elm)->field.le_next)) \
118 panic("Bad link elm %p next->prev != elm", (elm)); \
119 } while(0)
120
121 #define VERIFY_PREV_LINK(elm,field) do { \
122 if (*(elm)->field.le_prev != (elm)) \
123 panic("Bad link elm %p prev->next != elm", (elm)); \
124 } while(0)
125
126 /* tcp timer list */
127 struct tcptimerlist tcp_timer_list;
128
129 /* List of pcbs in timewait state, protected by tcbinfo's ipi_lock */
130 struct tcptailq tcp_tw_tailq;
131
132 static int background_io_trigger = 5;
133 SYSCTL_INT(_net_inet_tcp, OID_AUTO, background_io_trigger, CTLFLAG_RW | CTLFLAG_LOCKED,
134 &background_io_trigger, 0, "Background IO Trigger Setting");
135
136 static int
137 sysctl_msec_to_ticks SYSCTL_HANDLER_ARGS
138 {
139 #pragma unused(arg1, arg2)
140 int error, s, tt;
141
142 tt = *(int *)oidp->oid_arg1;
143 s = tt * 1000 / TCP_RETRANSHZ;;
144
145 error = sysctl_handle_int(oidp, &s, 0, req);
146 if (error || !req->newptr)
147 return (error);
148
149 tt = s * TCP_RETRANSHZ / 1000;
150 if (tt < 1)
151 return (EINVAL);
152
153 *(int *)oidp->oid_arg1 = tt;
154 return (0);
155 }
156
157 int tcp_keepinit;
158 SYSCTL_PROC(_net_inet_tcp, TCPCTL_KEEPINIT, keepinit, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_LOCKED,
159 &tcp_keepinit, 0, sysctl_msec_to_ticks, "I", "");
160
161 int tcp_keepidle;
162 SYSCTL_PROC(_net_inet_tcp, TCPCTL_KEEPIDLE, keepidle, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_LOCKED,
163 &tcp_keepidle, 0, sysctl_msec_to_ticks, "I", "");
164
165 int tcp_keepintvl;
166 SYSCTL_PROC(_net_inet_tcp, TCPCTL_KEEPINTVL, keepintvl, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_LOCKED,
167 &tcp_keepintvl, 0, sysctl_msec_to_ticks, "I", "");
168
169 int tcp_keepcnt;
170 SYSCTL_INT(_net_inet_tcp, OID_AUTO, keepcnt, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_LOCKED,
171 &tcp_keepcnt, 0, "number of times to repeat keepalive");
172
173 int tcp_msl;
174 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, msl, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_LOCKED,
175 &tcp_msl, 0, sysctl_msec_to_ticks, "I", "Maximum segment lifetime");
176
177 /*
178 * Avoid DoS via TCP Robustness in Persist Condition (see http://www.ietf.org/id/draft-ananth-tcpm-persist-02.txt)
179 * by allowing a system wide maximum persistence timeout value when in Zero Window Probe mode.
180 * Expressed in milliseconds to be consistent without timeout related values, the TCP socket option is in seconds.
181 */
182 u_int32_t tcp_max_persist_timeout = 0;
183 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, max_persist_timeout, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_LOCKED,
184 &tcp_max_persist_timeout, 0, sysctl_msec_to_ticks, "I", "Maximum persistence timout for ZWP");
185
186 static int always_keepalive = 0;
187 SYSCTL_INT(_net_inet_tcp, OID_AUTO, always_keepalive, CTLFLAG_RW | CTLFLAG_LOCKED,
188 &always_keepalive , 0, "Assume SO_KEEPALIVE on all TCP connections");
189
190 /* This parameter determines how long the timer list will stay in fast mode even
191 * though all connections are idle. In fast mode, the timer will fire more frequently
192 * anticipating new data.
193 */
194 int timer_fastmode_idlemax = TCP_FASTMODE_IDLEGEN_MAX;
195 SYSCTL_INT(_net_inet_tcp, OID_AUTO, timer_fastmode_idlemax, CTLFLAG_RW | CTLFLAG_LOCKED,
196 &timer_fastmode_idlemax, 0, "Maximum idle generations in fast mode");
197
198 /*
199 * See tcp_syn_backoff[] for interval values between SYN retransmits;
200 * the value set below defines the number of retransmits, before we
201 * disable the timestamp and window scaling options during subsequent
202 * SYN retransmits. Setting it to 0 disables the dropping off of those
203 * two options.
204 */
205 static int tcp_broken_peer_syn_rxmit_thres = 7;
206 SYSCTL_INT(_net_inet_tcp, OID_AUTO, broken_peer_syn_rxmit_thres, CTLFLAG_RW | CTLFLAG_LOCKED,
207 &tcp_broken_peer_syn_rxmit_thres, 0, "Number of retransmitted SYNs before "
208 "TCP disables rfc1323 and rfc1644 during the rest of attempts");
209
210 /* A higher threshold on local connections for disabling RFC 1323 options */
211 static int tcp_broken_peer_syn_rxmit_thres_local = 10;
212 SYSCTL_INT(_net_inet_tcp, OID_AUTO, broken_peer_syn_rexmit_thres_local,
213 CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_broken_peer_syn_rxmit_thres_local, 0,
214 "Number of retransmitted SYNs before disabling RFC 1323 options on local connections");
215
216 static int tcp_timer_advanced = 0;
217 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcp_timer_advanced, CTLFLAG_RD | CTLFLAG_LOCKED,
218 &tcp_timer_advanced, 0, "Number of times one of the timers was advanced");
219
220 static int tcp_resched_timerlist = 0;
221 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcp_resched_timerlist, CTLFLAG_RD | CTLFLAG_LOCKED,
222 &tcp_resched_timerlist, 0,
223 "Number of times timer list was rescheduled as part of processing a packet");
224
225 int tcp_pmtud_black_hole_detect = 1 ;
226 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pmtud_blackhole_detection, CTLFLAG_RW | CTLFLAG_LOCKED,
227 &tcp_pmtud_black_hole_detect, 0, "Path MTU Discovery Black Hole Detection");
228
229 int tcp_pmtud_black_hole_mss = 1200 ;
230 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pmtud_blackhole_mss, CTLFLAG_RW | CTLFLAG_LOCKED,
231 &tcp_pmtud_black_hole_mss, 0, "Path MTU Discovery Black Hole Detection lowered MSS");
232
233 /* performed garbage collection of "used" sockets */
234 static boolean_t tcp_gc_done = FALSE;
235
236 /* max idle probes */
237 int tcp_maxpersistidle;
238
239 /* TCP delack timer is set to 100 ms. Since the processing of timer list in fast
240 * mode will happen no faster than 100 ms, the delayed ack timer will fire some where
241 * between 100 and 200 ms.
242 */
243 int tcp_delack = TCP_RETRANSHZ / 10;
244
245 #if MPTCP
246 /*
247 * MP_JOIN retransmission of 3rd ACK will be every 500 msecs without backoff
248 */
249 int tcp_jack_rxmt = TCP_RETRANSHZ / 2;
250 #endif /* MPTCP */
251
252 /* The frequency of running through the TCP timer list in
253 * fast and slow mode can be configured.
254 */
255 SYSCTL_UINT(_net_inet_tcp, OID_AUTO, timer_fastquantum, CTLFLAG_RW | CTLFLAG_LOCKED,
256 &tcp_timer_list.fast_quantum, TCP_FASTTIMER_QUANTUM,
257 "Frequency of running timer list in fast mode");
258
259 SYSCTL_UINT(_net_inet_tcp, OID_AUTO, timer_slowquantum, CTLFLAG_RW | CTLFLAG_LOCKED,
260 &tcp_timer_list.slow_quantum, TCP_SLOWTIMER_QUANTUM,
261 "Frequency of running timer list in slow mode");
262
263 static void tcp_remove_timer(struct tcpcb *tp);
264 static void tcp_sched_timerlist(uint32_t offset);
265 static uint32_t tcp_run_conn_timer(struct tcpcb *tp, uint16_t *next_index);
266 static void tcp_sched_timers(struct tcpcb *tp);
267 static inline void tcp_set_lotimer_index(struct tcpcb *);
268 static void tcp_rexmt_save_state(struct tcpcb *tp);
269 void tcp_remove_from_time_wait(struct inpcb *inp);
270
271 /* Macro to compare two timers. If there is a reset of the sign bit, it is
272 * safe to assume that the timer has wrapped around. By doing signed comparision,
273 * we take care of wrap around such that the value with the sign bit reset is
274 * actually ahead of the other.
275 */
276
277 static inline int32_t
278 timer_diff(uint32_t t1, uint32_t toff1, uint32_t t2, uint32_t toff2) {
279 return (int32_t)((t1 + toff1) - (t2 + toff2));
280 };
281
282 /* Returns true if the timer is on the timer list */
283 #define TIMER_IS_ON_LIST(tp) ((tp)->t_flags & TF_TIMER_ONLIST)
284
285 /* Run the TCP timerlist atleast once every hour */
286 #define TCP_TIMERLIST_MAX_OFFSET (60 * 60 * TCP_RETRANSHZ)
287
288 static void add_to_time_wait_locked(struct tcpcb *tp, uint32_t delay);
289 void add_to_time_wait(struct tcpcb *tp, uint32_t delay) ;
290
291 static boolean_t tcp_garbage_collect(struct inpcb *, int);
292
293 /*
294 * Add to tcp timewait list, delay is given in milliseconds.
295 */
296 static void
297 add_to_time_wait_locked(struct tcpcb *tp, uint32_t delay)
298 {
299 struct inpcbinfo *pcbinfo = &tcbinfo;
300 struct inpcb *inp = tp->t_inpcb;
301 uint32_t timer;
302
303 /* pcb list should be locked when we get here */
304 lck_rw_assert(pcbinfo->ipi_lock, LCK_RW_ASSERT_EXCLUSIVE);
305
306 /* We may get here multiple times, so check */
307 if (!(inp->inp_flags2 & INP2_TIMEWAIT)) {
308 pcbinfo->ipi_twcount++;
309 inp->inp_flags2 |= INP2_TIMEWAIT;
310
311 /* Remove from global inp list */
312 LIST_REMOVE(inp, inp_list);
313 } else {
314 TAILQ_REMOVE(&tcp_tw_tailq, tp, t_twentry);
315 }
316
317 /* Compute the time at which this socket can be closed */
318 timer = tcp_now + delay;
319
320 /* We will use the TCPT_2MSL timer for tracking this delay */
321
322 if (TIMER_IS_ON_LIST(tp))
323 tcp_remove_timer(tp);
324 tp->t_timer[TCPT_2MSL] = timer;
325
326 TAILQ_INSERT_TAIL(&tcp_tw_tailq, tp, t_twentry);
327 }
328
329 void
330 add_to_time_wait(struct tcpcb *tp, uint32_t delay)
331 {
332 struct inpcbinfo *pcbinfo = &tcbinfo;
333
334 if (!lck_rw_try_lock_exclusive(pcbinfo->ipi_lock)) {
335 tcp_unlock(tp->t_inpcb->inp_socket, 0, 0);
336 lck_rw_lock_exclusive(pcbinfo->ipi_lock);
337 tcp_lock(tp->t_inpcb->inp_socket, 0, 0);
338 }
339 add_to_time_wait_locked(tp, delay);
340 lck_rw_done(pcbinfo->ipi_lock);
341
342 inpcb_gc_sched(pcbinfo, INPCB_TIMER_LAZY);
343 }
344
345 /* If this is on time wait queue, remove it. */
346 void
347 tcp_remove_from_time_wait(struct inpcb *inp)
348 {
349 struct tcpcb *tp = intotcpcb(inp);
350 if (inp->inp_flags2 & INP2_TIMEWAIT)
351 TAILQ_REMOVE(&tcp_tw_tailq, tp, t_twentry);
352 }
353
354 static boolean_t
355 tcp_garbage_collect(struct inpcb *inp, int istimewait)
356 {
357 boolean_t active = FALSE;
358 struct socket *so;
359 struct tcpcb *tp;
360
361 so = inp->inp_socket;
362 tp = intotcpcb(inp);
363
364 /*
365 * Skip if still in use or busy; it would have been more efficient
366 * if we were to test so_usecount against 0, but this isn't possible
367 * due to the current implementation of tcp_dropdropablreq() where
368 * overflow sockets that are eligible for garbage collection have
369 * their usecounts set to 1.
370 */
371 if (!lck_mtx_try_lock_spin(&inp->inpcb_mtx))
372 return (TRUE);
373
374 /* Check again under the lock */
375 if (so->so_usecount > 1) {
376 if (inp->inp_wantcnt == WNT_STOPUSING)
377 active = TRUE;
378 lck_mtx_unlock(&inp->inpcb_mtx);
379 return (active);
380 }
381
382 if (istimewait &&
383 TSTMP_GEQ(tcp_now, tp->t_timer[TCPT_2MSL]) &&
384 tp->t_state != TCPS_CLOSED) {
385 /* Become a regular mutex */
386 lck_mtx_convert_spin(&inp->inpcb_mtx);
387 tcp_close(tp);
388 }
389
390 /*
391 * Overflowed socket dropped from the listening queue? Do this
392 * only if we are called to clean up the time wait slots, since
393 * tcp_dropdropablreq() considers a socket to have been fully
394 * dropped after add_to_time_wait() is finished.
395 * Also handle the case of connections getting closed by the peer
396 * while in the queue as seen with rdar://6422317
397 *
398 */
399 if (so->so_usecount == 1 &&
400 ((istimewait && (so->so_flags & SOF_OVERFLOW)) ||
401 ((tp != NULL) && (tp->t_state == TCPS_CLOSED) &&
402 (so->so_head != NULL) &&
403 ((so->so_state & (SS_INCOMP|SS_CANTSENDMORE|SS_CANTRCVMORE)) ==
404 (SS_INCOMP|SS_CANTSENDMORE|SS_CANTRCVMORE))))) {
405
406 if (inp->inp_state != INPCB_STATE_DEAD) {
407 /* Become a regular mutex */
408 lck_mtx_convert_spin(&inp->inpcb_mtx);
409 #if INET6
410 if (SOCK_CHECK_DOM(so, PF_INET6))
411 in6_pcbdetach(inp);
412 else
413 #endif /* INET6 */
414 in_pcbdetach(inp);
415 }
416 so->so_usecount--;
417 if (inp->inp_wantcnt == WNT_STOPUSING)
418 active = TRUE;
419 lck_mtx_unlock(&inp->inpcb_mtx);
420 return (active);
421 } else if (inp->inp_wantcnt != WNT_STOPUSING) {
422 lck_mtx_unlock(&inp->inpcb_mtx);
423 return (FALSE);
424 }
425
426 /*
427 * We get here because the PCB is no longer searchable
428 * (WNT_STOPUSING); detach (if needed) and dispose if it is dead
429 * (usecount is 0). This covers all cases, including overflow
430 * sockets and those that are considered as "embryonic",
431 * i.e. created by sonewconn() in TCP input path, and have
432 * not yet been committed. For the former, we reduce the usecount
433 * to 0 as done by the code above. For the latter, the usecount
434 * would have reduced to 0 as part calling soabort() when the
435 * socket is dropped at the end of tcp_input().
436 */
437 if (so->so_usecount == 0) {
438 DTRACE_TCP4(state__change, void, NULL, struct inpcb *, inp,
439 struct tcpcb *, tp, int32_t, TCPS_CLOSED);
440 /* Become a regular mutex */
441 lck_mtx_convert_spin(&inp->inpcb_mtx);
442
443 /*
444 * If this tp still happens to be on the timer list,
445 * take it out
446 */
447 if (TIMER_IS_ON_LIST(tp)) {
448 tcp_remove_timer(tp);
449 }
450
451 if (inp->inp_state != INPCB_STATE_DEAD) {
452 #if INET6
453 if (SOCK_CHECK_DOM(so, PF_INET6))
454 in6_pcbdetach(inp);
455 else
456 #endif /* INET6 */
457 in_pcbdetach(inp);
458 }
459 in_pcbdispose(inp);
460 return (FALSE);
461 }
462
463 lck_mtx_unlock(&inp->inpcb_mtx);
464 return (TRUE);
465 }
466
467 /*
468 * TCP garbage collector callback (inpcb_timer_func_t).
469 *
470 * Returns the number of pcbs that will need to be gc-ed soon,
471 * returnining > 0 will keep timer active.
472 */
473 void
474 tcp_gc(struct inpcbinfo *ipi)
475 {
476 struct inpcb *inp, *nxt;
477 struct tcpcb *tw_tp, *tw_ntp;
478 #if TCPDEBUG
479 int ostate;
480 #endif
481 #if KDEBUG
482 static int tws_checked = 0;
483 #endif
484
485 KERNEL_DEBUG(DBG_FNC_TCP_SLOW | DBG_FUNC_START, 0, 0, 0, 0, 0);
486
487 /*
488 * Update tcp_now here as it may get used while
489 * processing the slow timer.
490 */
491 calculate_tcp_clock();
492
493 /*
494 * Garbage collect socket/tcpcb: We need to acquire the list lock
495 * exclusively to do this
496 */
497
498 if (lck_rw_try_lock_exclusive(ipi->ipi_lock) == FALSE) {
499 /* don't sweat it this time; cleanup was done last time */
500 if (tcp_gc_done == TRUE) {
501 tcp_gc_done = FALSE;
502 KERNEL_DEBUG(DBG_FNC_TCP_SLOW | DBG_FUNC_END,
503 tws_checked, cur_tw_slot, 0, 0, 0);
504 /* Lock upgrade failed, give up this round */
505 atomic_add_32(&ipi->ipi_gc_req.intimer_fast, 1);
506 return;
507 }
508 /* Upgrade failed, lost lock now take it again exclusive */
509 lck_rw_lock_exclusive(ipi->ipi_lock);
510 }
511 tcp_gc_done = TRUE;
512
513 LIST_FOREACH_SAFE(inp, &tcb, inp_list, nxt) {
514 if (tcp_garbage_collect(inp, 0))
515 atomic_add_32(&ipi->ipi_gc_req.intimer_fast, 1);
516 }
517
518 /* Now cleanup the time wait ones */
519 TAILQ_FOREACH_SAFE(tw_tp, &tcp_tw_tailq, t_twentry, tw_ntp) {
520 /*
521 * We check the timestamp here without holding the
522 * socket lock for better performance. If there are
523 * any pcbs in time-wait, the timer will get rescheduled.
524 * Hence some error in this check can be tolerated.
525 *
526 * Sometimes a socket on time-wait queue can be closed if
527 * 2MSL timer expired but the application still has a
528 * usecount on it.
529 */
530 if (tw_tp->t_state == TCPS_CLOSED ||
531 TSTMP_GEQ(tcp_now, tw_tp->t_timer[TCPT_2MSL])) {
532 if (tcp_garbage_collect(tw_tp->t_inpcb, 1))
533 atomic_add_32(&ipi->ipi_gc_req.intimer_lazy, 1);
534 }
535 }
536
537 /* take into account pcbs that are still in time_wait_slots */
538 atomic_add_32(&ipi->ipi_gc_req.intimer_lazy, ipi->ipi_twcount);
539
540 lck_rw_done(ipi->ipi_lock);
541
542 /* Clean up the socache while we are here */
543 if (so_cache_timer())
544 atomic_add_32(&ipi->ipi_gc_req.intimer_lazy, 1);
545
546 KERNEL_DEBUG(DBG_FNC_TCP_SLOW | DBG_FUNC_END, tws_checked,
547 cur_tw_slot, 0, 0, 0);
548
549 return;
550 }
551
552 /*
553 * Cancel all timers for TCP tp.
554 */
555 void
556 tcp_canceltimers(tp)
557 struct tcpcb *tp;
558 {
559 register int i;
560
561 tcp_remove_timer(tp);
562 for (i = 0; i < TCPT_NTIMERS; i++)
563 tp->t_timer[i] = 0;
564 tp->tentry.timer_start = tcp_now;
565 tp->tentry.index = TCPT_NONE;
566 }
567
568 int tcp_syn_backoff[TCP_MAXRXTSHIFT + 1] =
569 { 1, 1, 1, 1, 1, 2, 4, 8, 16, 32, 64, 64, 64 };
570
571 int tcp_backoff[TCP_MAXRXTSHIFT + 1] =
572 { 1, 2, 4, 8, 16, 32, 64, 64, 64, 64, 64, 64, 64 };
573
574 static int tcp_totbackoff = 511; /* sum of tcp_backoff[] */
575
576 static void tcp_rexmt_save_state(struct tcpcb *tp)
577 {
578 u_int32_t fsize;
579 if (TSTMP_SUPPORTED(tp)) {
580 /*
581 * Since timestamps are supported on the connection,
582 * we can do recovery as described in rfc 4015.
583 */
584 fsize = tp->snd_max - tp->snd_una;
585 tp->snd_ssthresh_prev = max(fsize, tp->snd_ssthresh);
586 tp->snd_recover_prev = tp->snd_recover;
587 } else {
588 /*
589 * Timestamp option is not supported on this connection.
590 * Record ssthresh and cwnd so they can
591 * be recovered if this turns out to be a "bad" retransmit.
592 * A retransmit is considered "bad" if an ACK for this
593 * segment is received within RTT/2 interval; the assumption
594 * here is that the ACK was already in flight. See
595 * "On Estimating End-to-End Network Path Properties" by
596 * Allman and Paxson for more details.
597 */
598 tp->snd_cwnd_prev = tp->snd_cwnd;
599 tp->snd_ssthresh_prev = tp->snd_ssthresh;
600 tp->snd_recover_prev = tp->snd_recover;
601 if (IN_FASTRECOVERY(tp))
602 tp->t_flags |= TF_WASFRECOVERY;
603 else
604 tp->t_flags &= ~TF_WASFRECOVERY;
605 }
606 tp->t_srtt_prev = (tp->t_srtt >> TCP_RTT_SHIFT) + 2;
607 tp->t_rttvar_prev = (tp->t_rttvar >> TCP_RTTVAR_SHIFT);
608 tp->t_flagsext &= ~(TF_RECOMPUTE_RTT);
609 }
610
611 /*
612 * TCP timer processing.
613 */
614 struct tcpcb *
615 tcp_timers(tp, timer)
616 register struct tcpcb *tp;
617 int timer;
618 {
619 register int rexmt;
620 struct socket *so;
621 struct tcptemp *t_template;
622 int optlen = 0;
623 int idle_time = 0;
624
625 #if TCPDEBUG
626 int ostate;
627 #endif
628
629 #if INET6
630 int isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV4) == 0;
631 #endif /* INET6 */
632
633 so = tp->t_inpcb->inp_socket;
634 idle_time = tcp_now - tp->t_rcvtime;
635
636 switch (timer) {
637
638 /*
639 * 2 MSL timeout in shutdown went off. If we're closed but
640 * still waiting for peer to close and connection has been idle
641 * too long, or if 2MSL time is up from TIME_WAIT or FIN_WAIT_2,
642 * delete connection control block.
643 * Otherwise, (this case shouldn't happen) check again in a bit
644 * we keep the socket in the main list in that case.
645 */
646 case TCPT_2MSL:
647 tcp_free_sackholes(tp);
648 if (tp->t_state != TCPS_TIME_WAIT &&
649 tp->t_state != TCPS_FIN_WAIT_2 &&
650 ((idle_time > 0) && (idle_time < TCP_CONN_MAXIDLE(tp)))) {
651 tp->t_timer[TCPT_2MSL] = OFFSET_FROM_START(tp,
652 (u_int32_t)TCP_CONN_KEEPINTVL(tp));
653 } else {
654 tp = tcp_close(tp);
655 return(tp);
656 }
657 break;
658
659 /*
660 * Retransmission timer went off. Message has not
661 * been acked within retransmit interval. Back off
662 * to a longer retransmit interval and retransmit one segment.
663 */
664 case TCPT_REXMT:
665 /* Drop a connection in the retransmit timer
666 * 1. If we have retransmitted more than TCP_MAXRXTSHIFT times
667 * 2. If the time spent in this retransmission episode is more than
668 * the time limit set with TCP_RXT_CONNDROPTIME socket option
669 * 3. If TCP_RXT_FINDROP socket option was set and we have already
670 * retransmitted the FIN 3 times without receiving an ack
671 */
672 if (++tp->t_rxtshift > TCP_MAXRXTSHIFT ||
673 (tp->t_rxt_conndroptime > 0 && tp->t_rxtstart > 0 &&
674 (tcp_now - tp->t_rxtstart) >= tp->t_rxt_conndroptime) ||
675 ((tp->t_flagsext & TF_RXTFINDROP) != 0 &&
676 (tp->t_flags & TF_SENTFIN) != 0 &&
677 tp->t_rxtshift >= 4)) {
678
679 if ((tp->t_flagsext & TF_RXTFINDROP) != 0) {
680 tcpstat.tcps_rxtfindrop++;
681 } else {
682 tcpstat.tcps_timeoutdrop++;
683 }
684 tp->t_rxtshift = TCP_MAXRXTSHIFT;
685 postevent(so, 0, EV_TIMEOUT);
686 soevent(so,
687 (SO_FILT_HINT_LOCKED|SO_FILT_HINT_TIMEOUT));
688 tp = tcp_drop(tp, tp->t_softerror ?
689 tp->t_softerror : ETIMEDOUT);
690
691 break;
692 }
693
694 tcpstat.tcps_rexmttimeo++;
695
696 if (tp->t_rxtshift == 1 &&
697 tp->t_state == TCPS_ESTABLISHED) {
698 /* Set the time at which retransmission started. */
699 tp->t_rxtstart = tcp_now;
700
701 /*
702 * if this is the first retransmit timeout, save
703 * the state so that we can recover if the timeout
704 * is spurious.
705 */
706 tcp_rexmt_save_state(tp);
707 }
708 #if MPTCP
709 if ((tp->t_rxtshift == mptcp_fail_thresh) &&
710 (tp->t_state == TCPS_ESTABLISHED) &&
711 (tp->t_mpflags & TMPF_MPTCP_TRUE)) {
712 mptcp_act_on_txfail(so);
713
714 }
715 #endif /* MPTCP */
716
717 if (tp->t_adaptive_wtimo > 0 &&
718 tp->t_rxtshift > tp->t_adaptive_wtimo &&
719 TCPS_HAVEESTABLISHED(tp->t_state)) {
720 /* Send an event to the application */
721 soevent(so,
722 (SO_FILT_HINT_LOCKED|
723 SO_FILT_HINT_ADAPTIVE_WTIMO));
724 }
725
726 if (tp->t_state == TCPS_SYN_SENT) {
727 rexmt = TCP_REXMTVAL(tp) * tcp_syn_backoff[tp->t_rxtshift];
728 tp->t_stat.synrxtshift = tp->t_rxtshift;
729 }
730 else
731 rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift];
732 TCPT_RANGESET(tp->t_rxtcur, rexmt,
733 tp->t_rttmin, TCPTV_REXMTMAX,
734 TCP_ADD_REXMTSLOP(tp));
735 tp->t_timer[TCPT_REXMT] = OFFSET_FROM_START(tp, tp->t_rxtcur);
736
737 if (INP_WAIT_FOR_IF_FEEDBACK(tp->t_inpcb))
738 goto fc_output;
739
740 tcp_free_sackholes(tp);
741 /*
742 * Check for potential Path MTU Discovery Black Hole
743 */
744
745 if (tcp_pmtud_black_hole_detect && (tp->t_state == TCPS_ESTABLISHED)) {
746 if (((tp->t_flags & (TF_PMTUD|TF_MAXSEGSNT)) == (TF_PMTUD|TF_MAXSEGSNT)) &&
747 (tp->t_rxtshift == 2)) {
748 /*
749 * Enter Path MTU Black-hole Detection mechanism:
750 * - Disable Path MTU Discovery (IP "DF" bit).
751 * - Reduce MTU to lower value than what we negociated with peer.
752 */
753 /* Disable Path MTU Discovery for now */
754 tp->t_flags &= ~TF_PMTUD;
755 /* Record that we may have found a black hole */
756 tp->t_flags |= TF_BLACKHOLE;
757 optlen = tp->t_maxopd - tp->t_maxseg;
758 /* Keep track of previous MSS */
759 tp->t_pmtud_saved_maxopd = tp->t_maxopd;
760 /* Reduce the MSS to intermediary value */
761 if (tp->t_maxopd > tcp_pmtud_black_hole_mss) {
762 tp->t_maxopd = tcp_pmtud_black_hole_mss;
763 } else {
764 tp->t_maxopd = /* use the default MSS */
765 #if INET6
766 isipv6 ? tcp_v6mssdflt :
767 #endif /* INET6 */
768 tcp_mssdflt;
769 }
770 tp->t_maxseg = tp->t_maxopd - optlen;
771
772 /*
773 * Reset the slow-start flight size
774 * as it may depend on the new MSS
775 */
776 if (CC_ALGO(tp)->cwnd_init != NULL)
777 CC_ALGO(tp)->cwnd_init(tp);
778 }
779 /*
780 * If further retransmissions are still unsuccessful with a lowered MTU,
781 * maybe this isn't a Black Hole and we restore the previous MSS and
782 * blackhole detection flags.
783 */
784 else {
785
786 if ((tp->t_flags & TF_BLACKHOLE) && (tp->t_rxtshift > 4)) {
787 tp->t_flags |= TF_PMTUD;
788 tp->t_flags &= ~TF_BLACKHOLE;
789 optlen = tp->t_maxopd - tp->t_maxseg;
790 tp->t_maxopd = tp->t_pmtud_saved_maxopd;
791 tp->t_maxseg = tp->t_maxopd - optlen;
792 /*
793 * Reset the slow-start flight size as it
794 * may depend on the new MSS
795 */
796 if (CC_ALGO(tp)->cwnd_init != NULL)
797 CC_ALGO(tp)->cwnd_init(tp);
798 }
799 }
800 }
801
802
803 /*
804 * Disable rfc1323 and rfc1644 if we haven't got any response to
805 * our SYN (after we reach the threshold) to work-around some
806 * broken terminal servers (most of which have hopefully been
807 * retired) that have bad VJ header compression code which
808 * trashes TCP segments containing unknown-to-them TCP options.
809 * Do this only on non-local connections.
810 */
811 if (tp->t_state == TCPS_SYN_SENT &&
812 ((!(tp->t_flags & TF_LOCAL) &&
813 tp->t_rxtshift == tcp_broken_peer_syn_rxmit_thres) ||
814 ((tp->t_flags & TF_LOCAL) &&
815 tp->t_rxtshift == tcp_broken_peer_syn_rxmit_thres_local)))
816 tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_REQ_CC);
817
818 /*
819 * If losing, let the lower level know and try for
820 * a better route. Also, if we backed off this far,
821 * our srtt estimate is probably bogus. Clobber it
822 * so we'll take the next rtt measurement as our srtt;
823 * move the current srtt into rttvar to keep the current
824 * retransmit times until then.
825 */
826 if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
827 #if INET6
828 if (isipv6)
829 in6_losing(tp->t_inpcb);
830 else
831 #endif /* INET6 */
832 in_losing(tp->t_inpcb);
833 tp->t_rttvar += (tp->t_srtt >> TCP_RTT_SHIFT);
834 tp->t_srtt = 0;
835 }
836 tp->snd_nxt = tp->snd_una;
837 /*
838 * Note: We overload snd_recover to function also as the
839 * snd_last variable described in RFC 2582
840 */
841 tp->snd_recover = tp->snd_max;
842 /*
843 * Force a segment to be sent.
844 */
845 tp->t_flags |= TF_ACKNOW;
846 /*
847 * If timing a segment in this window, stop the timer.
848 */
849 tp->t_rtttime = 0;
850
851 EXIT_FASTRECOVERY(tp);
852
853 /* RFC 5681 says: when a TCP sender detects segment loss
854 * using retransmit timer and the given segment has already
855 * been retransmitted by way of the retransmission timer at
856 * least once, the value of ssthresh is held constant
857 */
858 if (tp->t_rxtshift == 1 &&
859 CC_ALGO(tp)->after_timeout != NULL)
860 CC_ALGO(tp)->after_timeout(tp);
861
862
863 /* CWR notifications are to be sent on new data right after
864 * RTOs, Fast Retransmits and ECE notification receipts.
865 */
866 if ((tp->ecn_flags & TE_ECN_ON) == TE_ECN_ON) {
867 tp->ecn_flags |= TE_SENDCWR;
868 }
869 fc_output:
870 DTRACE_TCP5(cc, void, NULL, struct inpcb *, tp->t_inpcb,
871 struct tcpcb *, tp, struct tcphdr *, NULL,
872 int32_t, TCP_CC_REXMT_TIMEOUT);
873
874 (void) tcp_output(tp);
875 break;
876
877 /*
878 * Persistance timer into zero window.
879 * Force a byte to be output, if possible.
880 */
881 case TCPT_PERSIST:
882 tcpstat.tcps_persisttimeo++;
883 /*
884 * Hack: if the peer is dead/unreachable, we do not
885 * time out if the window is closed. After a full
886 * backoff, drop the connection if the idle time
887 * (no responses to probes) reaches the maximum
888 * backoff that we would use if retransmitting.
889 *
890 * Drop the connection if we reached the maximum allowed time for
891 * Zero Window Probes without a non-zero update from the peer.
892 * See rdar://5805356
893 */
894 if ((tp->t_rxtshift == TCP_MAXRXTSHIFT &&
895 (idle_time >= tcp_maxpersistidle ||
896 idle_time >= TCP_REXMTVAL(tp) * tcp_totbackoff)) ||
897 ((tp->t_persist_stop != 0) &&
898 TSTMP_LEQ(tp->t_persist_stop, tcp_now))) {
899 tcpstat.tcps_persistdrop++;
900 postevent(so, 0, EV_TIMEOUT);
901 soevent(so,
902 (SO_FILT_HINT_LOCKED|SO_FILT_HINT_TIMEOUT));
903 tp = tcp_drop(tp, ETIMEDOUT);
904 break;
905 }
906 tcp_setpersist(tp);
907 tp->t_force = 1;
908 (void) tcp_output(tp);
909 tp->t_force = 0;
910 break;
911
912 /*
913 * Keep-alive timer went off; send something
914 * or drop connection if idle for too long.
915 */
916 case TCPT_KEEP:
917 tcpstat.tcps_keeptimeo++;
918 #if MPTCP
919 /*
920 * Regular TCP connections do not send keepalives after closing
921 * MPTCP must not also, after sending Data FINs.
922 */
923 struct mptcb *mp_tp = tp->t_mptcb;
924 if ((tp->t_mpflags & TMPF_MPTCP_TRUE) &&
925 (mp_tp == NULL)) {
926 goto dropit;
927 } else if (mp_tp != NULL) {
928 if ((mptcp_ok_to_keepalive(mp_tp) == 0))
929 goto dropit;
930 }
931 #endif /* MPTCP */
932 if (tp->t_state < TCPS_ESTABLISHED)
933 goto dropit;
934 if ((always_keepalive ||
935 (tp->t_inpcb->inp_socket->so_options & SO_KEEPALIVE) ||
936 (tp->t_flagsext & TF_DETECT_READSTALL)) &&
937 (tp->t_state <= TCPS_CLOSING || tp->t_state == TCPS_FIN_WAIT_2)) {
938 if (idle_time >= TCP_CONN_KEEPIDLE(tp) + TCP_CONN_MAXIDLE(tp))
939 goto dropit;
940 /*
941 * Send a packet designed to force a response
942 * if the peer is up and reachable:
943 * either an ACK if the connection is still alive,
944 * or an RST if the peer has closed the connection
945 * due to timeout or reboot.
946 * Using sequence number tp->snd_una-1
947 * causes the transmitted zero-length segment
948 * to lie outside the receive window;
949 * by the protocol spec, this requires the
950 * correspondent TCP to respond.
951 */
952 tcpstat.tcps_keepprobe++;
953 t_template = tcp_maketemplate(tp);
954 if (t_template) {
955 unsigned int ifscope, nocell = 0;
956
957 if (tp->t_inpcb->inp_flags & INP_BOUND_IF)
958 ifscope = tp->t_inpcb->inp_boundifp->if_index;
959 else
960 ifscope = IFSCOPE_NONE;
961
962 /*
963 * If the socket isn't allowed to use the
964 * cellular interface, indicate it as such.
965 */
966 if (tp->t_inpcb->inp_flags & INP_NO_IFT_CELLULAR)
967 nocell = 1;
968
969 tcp_respond(tp, t_template->tt_ipgen,
970 &t_template->tt_t, (struct mbuf *)NULL,
971 tp->rcv_nxt, tp->snd_una - 1, 0, ifscope,
972 nocell);
973 (void) m_free(dtom(t_template));
974 if (tp->t_flagsext & TF_DETECT_READSTALL)
975 tp->t_rtimo_probes++;
976 }
977 tp->t_timer[TCPT_KEEP] = OFFSET_FROM_START(tp,
978 TCP_CONN_KEEPINTVL(tp));
979 } else {
980 tp->t_timer[TCPT_KEEP] = OFFSET_FROM_START(tp,
981 TCP_CONN_KEEPIDLE(tp));
982 }
983 if (tp->t_flagsext & TF_DETECT_READSTALL) {
984 /*
985 * The keep alive packets sent to detect a read
986 * stall did not get a response from the
987 * peer. Generate more keep-alives to confirm this.
988 * If the number of probes sent reaches the limit,
989 * generate an event.
990 */
991 if (tp->t_rtimo_probes > tp->t_adaptive_rtimo) {
992 /* Generate an event */
993 soevent(so,
994 (SO_FILT_HINT_LOCKED|
995 SO_FILT_HINT_ADAPTIVE_RTIMO));
996 tcp_keepalive_reset(tp);
997 } else {
998 tp->t_timer[TCPT_KEEP] = OFFSET_FROM_START(
999 tp, TCP_REXMTVAL(tp));
1000 }
1001 }
1002 break;
1003 case TCPT_DELACK:
1004 if (tcp_delack_enabled && (tp->t_flags & TF_DELACK)) {
1005 tp->t_flags &= ~TF_DELACK;
1006 tp->t_timer[TCPT_DELACK] = 0;
1007 tp->t_flags |= TF_ACKNOW;
1008
1009 /* If delayed ack timer fired while stretching acks
1010 * go back to acking every other packet
1011 */
1012 if ((tp->t_flags & TF_STRETCHACK) != 0)
1013 tcp_reset_stretch_ack(tp);
1014
1015 /* If we are measuring inter packet arrival jitter for
1016 * throttling a connection, this delayed ack might be
1017 * the reason for accumulating some jitter. So let's
1018 * restart the measurement.
1019 */
1020 CLEAR_IAJ_STATE(tp);
1021
1022 tcpstat.tcps_delack++;
1023 (void) tcp_output(tp);
1024 }
1025 break;
1026
1027 #if MPTCP
1028 case TCPT_JACK_RXMT:
1029 if ((tp->t_state == TCPS_ESTABLISHED) &&
1030 (tp->t_mpflags & TMPF_PREESTABLISHED) &&
1031 (tp->t_mpflags & TMPF_JOINED_FLOW)) {
1032 if (++tp->t_mprxtshift > TCP_MAXRXTSHIFT) {
1033 tcpstat.tcps_timeoutdrop++;
1034 postevent(so, 0, EV_TIMEOUT);
1035 soevent(so,
1036 (SO_FILT_HINT_LOCKED|
1037 SO_FILT_HINT_TIMEOUT));
1038 tp = tcp_drop(tp, tp->t_softerror ?
1039 tp->t_softerror : ETIMEDOUT);
1040 break;
1041 }
1042 tcpstat.tcps_join_rxmts++;
1043 tp->t_flags |= TF_ACKNOW;
1044
1045 /*
1046 * No backoff is implemented for simplicity for this
1047 * corner case.
1048 */
1049 (void) tcp_output(tp);
1050 }
1051 break;
1052 #endif /* MPTCP */
1053
1054 #if TCPDEBUG
1055 if (tp->t_inpcb->inp_socket->so_options & SO_DEBUG)
1056 tcp_trace(TA_USER, ostate, tp, (void *)0, (struct tcphdr *)0,
1057 PRU_SLOWTIMO);
1058 #endif
1059 dropit:
1060 tcpstat.tcps_keepdrops++;
1061 postevent(so, 0, EV_TIMEOUT);
1062 soevent(so,
1063 (SO_FILT_HINT_LOCKED|SO_FILT_HINT_TIMEOUT));
1064 tp = tcp_drop(tp, ETIMEDOUT);
1065 break;
1066 }
1067 return (tp);
1068 }
1069
1070 /* Remove a timer entry from timer list */
1071 void
1072 tcp_remove_timer(struct tcpcb *tp)
1073 {
1074 struct tcptimerlist *listp = &tcp_timer_list;
1075
1076 lck_mtx_assert(&tp->t_inpcb->inpcb_mtx, LCK_MTX_ASSERT_OWNED);
1077 if (!(TIMER_IS_ON_LIST(tp))) {
1078 return;
1079 }
1080 lck_mtx_lock(listp->mtx);
1081
1082 /* Check if pcb is on timer list again after acquiring the lock */
1083 if (!(TIMER_IS_ON_LIST(tp))) {
1084 lck_mtx_unlock(listp->mtx);
1085 return;
1086 }
1087
1088 if (listp->next_te != NULL && listp->next_te == &tp->tentry)
1089 listp->next_te = LIST_NEXT(&tp->tentry, le);
1090
1091 LIST_REMOVE(&tp->tentry, le);
1092 tp->t_flags &= ~(TF_TIMER_ONLIST);
1093
1094 listp->entries--;
1095
1096 tp->tentry.le.le_next = NULL;
1097 tp->tentry.le.le_prev = NULL;
1098 lck_mtx_unlock(listp->mtx);
1099 }
1100
1101 /* Function to check if the timerlist needs to be rescheduled to run
1102 * the timer entry correctly. Basically, this is to check if we can avoid
1103 * taking the list lock.
1104 */
1105
1106 static boolean_t
1107 need_to_resched_timerlist(uint32_t runtime, uint16_t index) {
1108 struct tcptimerlist *listp = &tcp_timer_list;
1109 int32_t diff;
1110 boolean_t is_fast;
1111
1112 if (index == TCPT_NONE)
1113 return FALSE;
1114 is_fast = !(IS_TIMER_SLOW(index));
1115
1116 /* If the list is being processed then the state of the list is in flux.
1117 * In this case always acquire the lock and set the state correctly.
1118 */
1119 if (listp->running)
1120 return TRUE;
1121
1122 if (!listp->scheduled)
1123 return (TRUE);
1124
1125 diff = timer_diff(listp->runtime, 0, runtime, 0);
1126 if (diff <= 0) {
1127 /* The list is going to run before this timer */
1128 return FALSE;
1129 } else {
1130 if (is_fast) {
1131 if (diff <= listp->fast_quantum)
1132 return FALSE;
1133 } else {
1134 if (diff <= listp->slow_quantum)
1135 return FALSE;
1136 }
1137 }
1138 return TRUE;
1139 }
1140
1141 void
1142 tcp_sched_timerlist(uint32_t offset)
1143 {
1144
1145 uint64_t deadline = 0;
1146 struct tcptimerlist *listp = &tcp_timer_list;
1147
1148 lck_mtx_assert(listp->mtx, LCK_MTX_ASSERT_OWNED);
1149
1150 offset = min(offset, TCP_TIMERLIST_MAX_OFFSET);
1151 listp->runtime = tcp_now + offset;
1152 if (listp->runtime == 0)
1153 listp->runtime++;
1154
1155 clock_interval_to_deadline(offset, NSEC_PER_SEC / TCP_RETRANSHZ,
1156 &deadline);
1157
1158 thread_call_enter_delayed(listp->call, deadline);
1159 listp->scheduled = TRUE;
1160 }
1161
1162 /* Function to run the timers for a connection.
1163 *
1164 * Returns the offset of next timer to be run for this connection which
1165 * can be used to reschedule the timerlist.
1166 */
1167 uint32_t
1168 tcp_run_conn_timer(struct tcpcb *tp, uint16_t *next_index) {
1169
1170 struct socket *so;
1171 uint16_t i = 0, index = TCPT_NONE, lo_index = TCPT_NONE;
1172 uint32_t timer_val, offset = 0, lo_timer = 0;
1173 int32_t diff;
1174 boolean_t needtorun[TCPT_NTIMERS];
1175 int count = 0;
1176
1177 VERIFY(tp != NULL);
1178 bzero(needtorun, sizeof(needtorun));
1179
1180 tcp_lock(tp->t_inpcb->inp_socket, 1, 0);
1181
1182 so = tp->t_inpcb->inp_socket;
1183 /* Release the want count on inp */
1184 if (in_pcb_checkstate(tp->t_inpcb, WNT_RELEASE, 1) == WNT_STOPUSING) {
1185 if (TIMER_IS_ON_LIST(tp)) {
1186 tcp_remove_timer(tp);
1187 }
1188
1189 /* Looks like the TCP connection got closed while we
1190 * were waiting for the lock.. Done
1191 */
1192 goto done;
1193 }
1194
1195 /* Since the timer thread needs to wait for tcp lock, it may race
1196 * with another thread that can cancel or reschedule the timer that is
1197 * about to run. Check if we need to run anything.
1198 */
1199 if ((index = tp->tentry.index) == TCPT_NONE)
1200 goto done;
1201 timer_val = tp->t_timer[index];
1202
1203 diff = timer_diff(tp->tentry.runtime, 0, tcp_now, 0);
1204 if (diff > 0) {
1205 if (tp->tentry.index != TCPT_NONE) {
1206 offset = diff;
1207 *(next_index) = tp->tentry.index;
1208 }
1209 goto done;
1210 }
1211
1212 tp->t_timer[index] = 0;
1213 if (timer_val > 0) {
1214 tp = tcp_timers(tp, index);
1215 if (tp == NULL)
1216 goto done;
1217 }
1218
1219 /* Check if there are any other timers that need to be run. While doing it,
1220 * adjust the timer values wrt tcp_now.
1221 */
1222 for (i = 0; i < TCPT_NTIMERS; ++i) {
1223 if (tp->t_timer[i] != 0) {
1224 diff = timer_diff(tp->tentry.timer_start, tp->t_timer[i], tcp_now, 0);
1225 if (diff <= 0) {
1226 tp->t_timer[i] = 0;
1227 needtorun[i] = TRUE;
1228 count++;
1229 } else {
1230 tp->t_timer[i] = diff;
1231 needtorun[i] = FALSE;
1232 if (lo_timer == 0 || diff < lo_timer) {
1233 lo_timer = diff;
1234 lo_index = i;
1235 }
1236 }
1237 }
1238 }
1239
1240 tp->tentry.timer_start = tcp_now;
1241 tp->tentry.index = lo_index;
1242 if (lo_index != TCPT_NONE) {
1243 tp->tentry.runtime = tp->tentry.timer_start + tp->t_timer[lo_index];
1244 if (tp->tentry.runtime == 0)
1245 tp->tentry.runtime++;
1246 }
1247
1248 if (count > 0) {
1249 /* run any other timers that are also outstanding at this time. */
1250 for (i = 0; i < TCPT_NTIMERS; ++i) {
1251 if (needtorun[i]) {
1252 tp->t_timer[i] = 0;
1253 tp = tcp_timers(tp, i);
1254 if (tp == NULL) {
1255 offset = 0;
1256 *(next_index) = TCPT_NONE;
1257 goto done;
1258 }
1259 }
1260 }
1261 tcp_set_lotimer_index(tp);
1262 }
1263
1264 if (tp->tentry.index < TCPT_NONE) {
1265 offset = tp->t_timer[tp->tentry.index];
1266 *(next_index) = tp->tentry.index;
1267 }
1268
1269 done:
1270 if (tp != NULL && tp->tentry.index == TCPT_NONE) {
1271 tcp_remove_timer(tp);
1272 offset = 0;
1273 }
1274 tcp_unlock(so, 1, 0);
1275 return offset;
1276 }
1277
1278 void
1279 tcp_run_timerlist(void * arg1, void * arg2) {
1280
1281 #pragma unused(arg1, arg2)
1282
1283 struct tcptimerentry *te, *next_te;
1284 struct tcptimerlist *listp = &tcp_timer_list;
1285 struct tcpcb *tp;
1286 uint32_t next_timer = 0;
1287 uint16_t index = TCPT_NONE;
1288 boolean_t need_fast = FALSE;
1289 uint32_t active_count = 0;
1290 uint32_t mode = TCP_TIMERLIST_FASTMODE;
1291
1292 calculate_tcp_clock();
1293
1294 lck_mtx_lock(listp->mtx);
1295
1296 listp->running = TRUE;
1297
1298 LIST_FOREACH_SAFE(te, &listp->lhead, le, next_te) {
1299 uint32_t offset = 0;
1300 uint32_t runtime = te->runtime;
1301 if (te->index < TCPT_NONE && TSTMP_GT(runtime, tcp_now)) {
1302 offset = timer_diff(runtime, 0, tcp_now, 0);
1303 if (next_timer == 0 || offset < next_timer) {
1304 next_timer = offset;
1305 }
1306 continue;
1307 }
1308 active_count++;
1309
1310 tp = TIMERENTRY_TO_TP(te);
1311
1312 /* Acquire an inp wantcnt on the inpcb so that the socket won't get
1313 * detached even if tcp_close is called
1314 */
1315 if (in_pcb_checkstate(tp->t_inpcb, WNT_ACQUIRE, 0) == WNT_STOPUSING) {
1316 /* Some how this pcb went into dead state while on the timer list,
1317 * just take it off the list. Since the timer list entry pointers
1318 * are protected by the timer list lock, we can do it here
1319 */
1320 if (TIMER_IS_ON_LIST(tp)) {
1321 tp->t_flags &= ~(TF_TIMER_ONLIST);
1322 LIST_REMOVE(&tp->tentry, le);
1323 listp->entries--;
1324
1325 tp->tentry.le.le_next = NULL;
1326 tp->tentry.le.le_prev = NULL;
1327 }
1328 continue;
1329 }
1330
1331 /* Store the next timerentry pointer before releasing the list lock.
1332 * If that entry has to be removed when we release the lock, this
1333 * pointer will be updated to the element after that.
1334 */
1335 listp->next_te = next_te;
1336
1337 VERIFY_NEXT_LINK(&tp->tentry, le);
1338 VERIFY_PREV_LINK(&tp->tentry, le);
1339
1340 lck_mtx_unlock(listp->mtx);
1341
1342 index = TCPT_NONE;
1343 offset = tcp_run_conn_timer(tp, &index);
1344
1345 lck_mtx_lock(listp->mtx);
1346
1347 next_te = listp->next_te;
1348 listp->next_te = NULL;
1349
1350 if (offset > 0) {
1351 if (index < TCPT_NONE) {
1352 /* Check if this is a fast_timer. */
1353 if (!need_fast && !(IS_TIMER_SLOW(index))) {
1354 need_fast = TRUE;
1355 }
1356
1357 if (next_timer == 0 || offset < next_timer) {
1358 next_timer = offset;
1359 }
1360 }
1361 }
1362 }
1363
1364 if (!LIST_EMPTY(&listp->lhead)) {
1365 if (listp->mode == TCP_TIMERLIST_FASTMODE) {
1366 if (need_fast || active_count > 0 ||
1367 listp->pref_mode == TCP_TIMERLIST_FASTMODE) {
1368 listp->idlegen = 0;
1369 } else {
1370 listp->idlegen++;
1371 if (listp->idlegen > timer_fastmode_idlemax) {
1372 mode = TCP_TIMERLIST_SLOWMODE;
1373 listp->idlegen = 0;
1374 }
1375 }
1376 } else {
1377 if (!need_fast) {
1378 mode = TCP_TIMERLIST_SLOWMODE;
1379 }
1380 }
1381
1382 if (mode == TCP_TIMERLIST_FASTMODE ||
1383 listp->pref_mode == TCP_TIMERLIST_FASTMODE) {
1384 next_timer = listp->fast_quantum;
1385 } else {
1386 if (listp->pref_offset != 0 &&
1387 listp->pref_offset < next_timer)
1388 next_timer = listp->pref_offset;
1389 if (next_timer < listp->slow_quantum)
1390 next_timer = listp->slow_quantum;
1391 }
1392
1393 listp->mode = mode;
1394
1395 tcp_sched_timerlist(next_timer);
1396 } else {
1397 /*
1398 * No need to reschedule this timer, but always run
1399 * periodically at a much higher granularity.
1400 */
1401 tcp_sched_timerlist(TCP_TIMERLIST_MAX_OFFSET);
1402 }
1403
1404 listp->running = FALSE;
1405 listp->pref_mode = 0;
1406 listp->pref_offset = 0;
1407
1408 lck_mtx_unlock(listp->mtx);
1409 }
1410
1411 /* Function to verify if a change in timer state is required for a connection */
1412 void
1413 tcp_sched_timers(struct tcpcb *tp)
1414 {
1415 struct tcptimerentry *te = &tp->tentry;
1416 uint16_t index = te->index;
1417 struct tcptimerlist *listp = &tcp_timer_list;
1418 int32_t offset = 0;
1419 boolean_t is_fast;
1420 int list_locked = 0;
1421
1422 if (tp->t_inpcb->inp_state == INPCB_STATE_DEAD) {
1423 /* Just return without adding the dead pcb to the list */
1424 if (TIMER_IS_ON_LIST(tp)) {
1425 tcp_remove_timer(tp);
1426 }
1427 return;
1428 }
1429
1430 if (index == TCPT_NONE) {
1431 tcp_remove_timer(tp);
1432 return;
1433 }
1434
1435 is_fast = !(IS_TIMER_SLOW(index));
1436 offset = timer_diff(te->runtime, 0, tcp_now, 0);
1437 if (offset <= 0) {
1438 offset = 1;
1439 tcp_timer_advanced++;
1440 }
1441 if (is_fast)
1442 offset = listp->fast_quantum;
1443
1444 if (!TIMER_IS_ON_LIST(tp)) {
1445 if (!list_locked) {
1446 lck_mtx_lock(listp->mtx);
1447 list_locked = 1;
1448 }
1449
1450 LIST_INSERT_HEAD(&listp->lhead, te, le);
1451 tp->t_flags |= TF_TIMER_ONLIST;
1452
1453 listp->entries++;
1454 if (listp->entries > listp->maxentries)
1455 listp->maxentries = listp->entries;
1456
1457 /* if the list is not scheduled, just schedule it */
1458 if (!listp->scheduled)
1459 goto schedule;
1460
1461 }
1462
1463
1464 /* timer entry is currently on the list */
1465 if (need_to_resched_timerlist(te->runtime, index)) {
1466 tcp_resched_timerlist++;
1467
1468 if (!list_locked) {
1469 lck_mtx_lock(listp->mtx);
1470 list_locked = 1;
1471 }
1472
1473 VERIFY_NEXT_LINK(te, le);
1474 VERIFY_PREV_LINK(te, le);
1475
1476 if (listp->running) {
1477 if (is_fast) {
1478 listp->pref_mode = TCP_TIMERLIST_FASTMODE;
1479 } else if (listp->pref_offset == 0 ||
1480 offset < listp->pref_offset) {
1481 listp->pref_offset = offset;
1482 }
1483 } else {
1484 /*
1485 * The list could have got scheduled while this
1486 * thread was waiting for the lock
1487 */
1488 if (listp->scheduled) {
1489 int32_t diff;
1490 diff = timer_diff(listp->runtime, 0,
1491 tcp_now, offset);
1492 if (diff <= 0)
1493 goto done;
1494 else
1495 goto schedule;
1496 } else {
1497 goto schedule;
1498 }
1499 }
1500 }
1501 goto done;
1502
1503 schedule:
1504 if (is_fast) {
1505 listp->mode = TCP_TIMERLIST_FASTMODE;
1506 listp->idlegen = 0;
1507 }
1508 tcp_sched_timerlist(offset);
1509
1510 done:
1511 if (list_locked)
1512 lck_mtx_unlock(listp->mtx);
1513
1514 return;
1515 }
1516
1517 void
1518 tcp_set_lotimer_index(struct tcpcb *tp) {
1519 uint16_t i, lo_index = TCPT_NONE;
1520 uint32_t lo_timer = 0;
1521 for (i = 0; i < TCPT_NTIMERS; ++i) {
1522 if (tp->t_timer[i] != 0 &&
1523 (lo_timer == 0 || tp->t_timer[i] < lo_timer)) {
1524 lo_timer = tp->t_timer[i];
1525 lo_index = i;
1526 }
1527 }
1528 tp->tentry.index = lo_index;
1529 if (lo_index != TCPT_NONE) {
1530 tp->tentry.runtime = tp->tentry.timer_start + tp->t_timer[lo_index];
1531 if (tp->tentry.runtime == 0)
1532 tp->tentry.runtime++;
1533 }
1534 }
1535
1536 void
1537 tcp_check_timer_state(struct tcpcb *tp) {
1538
1539 lck_mtx_assert(&tp->t_inpcb->inpcb_mtx, LCK_MTX_ASSERT_OWNED);
1540
1541 if (tp->t_inpcb->inp_flags2 & INP2_TIMEWAIT)
1542 return;
1543
1544 tcp_set_lotimer_index(tp);
1545
1546 tcp_sched_timers(tp);
1547 return;
1548 }
mirror of XNU